<%BANNER%>

Identification and Classification of Green Citrus by Spectral Characteristics for Precision Agriculture

Permanent Link: http://ufdc.ufl.edu/UFE0020581/00001

Material Information

Title: Identification and Classification of Green Citrus by Spectral Characteristics for Precision Agriculture
Physical Description: 1 online resource (86 p.)
Language: english
Creator: Kane, Kevin Edward
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2007

Subjects

Subjects / Keywords: agriculture, citrus, image, precision, processing, spectral
Agricultural and Biological Engineering -- Dissertations, Academic -- UF
Genre: Agricultural and Biological Engineering thesis, M.E.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: Citrus production is a multi-million dollar industry in the state of Florida. The crop's economic significance is the driving force behind current developments in precision agriculture technologies assuring citrus growers can minimize their costs, protect the environment, and increase overall yield. A real-time citrus yield map while the citrus fruit are still maturing provides information to growers, giving them time to be proactive at improving the groves growth and planning ahead for the harvest. Recent research has shown cameras a image processing techniques have the ability to identify and count orange citrus fruit in the grove. However, there is a desire by the industry to have citrus yield maps earlier in the growing season, a time when citrus fruit are green. In this case, a traditional visible spectral camera can not accurately identify green citrus fruit against their green tree canopy. The objective of this research was to use spectral information from the near infrared (NIR) reflectance spectrum to identify citrus fruit while they are still green. To begin this work, 540 freshly harvested samples of green citrus fruit and leaves were gathered and measured their diffuse reflectance using a spectrophotometer. The resulting spectral curves from 400 nm to 2500 nm were analyzed using discriminability calculations to find critical wavelengths for separation. Fisher linear discriminant analysis showed the wavelengths of 881 nm and 1383 nm provided perfect green leaf and green fruit separation. This research provided a foundation for the design of an in-field NIR camera system for green citrus fruit identification. A highly sensitive NIR camera outfitted with three optical band pass filters (1064 nm, 1150 nm, and 1572 nm) were used for natural in-field image acquisitions. The 256 bit monochromatic images were studied using spectral indexing and image processing schemes. Using training images, an indexing and image processing algorithm was developed and tested on validation images. A 90.3 % correct pixel classification result was obtained, proving that NIR camera images can successfully be used in the identification of green citrus fruit in the grove. An R2 of 0.746 for fruit pixel counts verse a manually masked fruit pixel counts was achieved. Despite these positive accomplishments, the research has also revealed problems that are prohibitive to this identification method. These problems included the high financial cost of an NIR imaging system and the difficulty of proper illumination and multiple image alignments when working in a normal Florida citrus grove environment.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by Kevin Edward Kane.
Thesis: Thesis (M.E.)--University of Florida, 2007.
Local: Adviser: Lee, Won Suk.

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2007
System ID: UFE0020581:00001

Permanent Link: http://ufdc.ufl.edu/UFE0020581/00001

Material Information

Title: Identification and Classification of Green Citrus by Spectral Characteristics for Precision Agriculture
Physical Description: 1 online resource (86 p.)
Language: english
Creator: Kane, Kevin Edward
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2007

Subjects

Subjects / Keywords: agriculture, citrus, image, precision, processing, spectral
Agricultural and Biological Engineering -- Dissertations, Academic -- UF
Genre: Agricultural and Biological Engineering thesis, M.E.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: Citrus production is a multi-million dollar industry in the state of Florida. The crop's economic significance is the driving force behind current developments in precision agriculture technologies assuring citrus growers can minimize their costs, protect the environment, and increase overall yield. A real-time citrus yield map while the citrus fruit are still maturing provides information to growers, giving them time to be proactive at improving the groves growth and planning ahead for the harvest. Recent research has shown cameras a image processing techniques have the ability to identify and count orange citrus fruit in the grove. However, there is a desire by the industry to have citrus yield maps earlier in the growing season, a time when citrus fruit are green. In this case, a traditional visible spectral camera can not accurately identify green citrus fruit against their green tree canopy. The objective of this research was to use spectral information from the near infrared (NIR) reflectance spectrum to identify citrus fruit while they are still green. To begin this work, 540 freshly harvested samples of green citrus fruit and leaves were gathered and measured their diffuse reflectance using a spectrophotometer. The resulting spectral curves from 400 nm to 2500 nm were analyzed using discriminability calculations to find critical wavelengths for separation. Fisher linear discriminant analysis showed the wavelengths of 881 nm and 1383 nm provided perfect green leaf and green fruit separation. This research provided a foundation for the design of an in-field NIR camera system for green citrus fruit identification. A highly sensitive NIR camera outfitted with three optical band pass filters (1064 nm, 1150 nm, and 1572 nm) were used for natural in-field image acquisitions. The 256 bit monochromatic images were studied using spectral indexing and image processing schemes. Using training images, an indexing and image processing algorithm was developed and tested on validation images. A 90.3 % correct pixel classification result was obtained, proving that NIR camera images can successfully be used in the identification of green citrus fruit in the grove. An R2 of 0.746 for fruit pixel counts verse a manually masked fruit pixel counts was achieved. Despite these positive accomplishments, the research has also revealed problems that are prohibitive to this identification method. These problems included the high financial cost of an NIR imaging system and the difficulty of proper illumination and multiple image alignments when working in a normal Florida citrus grove environment.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by Kevin Edward Kane.
Thesis: Thesis (M.E.)--University of Florida, 2007.
Local: Adviser: Lee, Won Suk.

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2007
System ID: UFE0020581:00001


This item has the following downloads:


Full Text
xml version 1.0 encoding UTF-8
REPORT xmlns http:www.fcla.edudlsmddaitss xmlns:xsi http:www.w3.org2001XMLSchema-instance xsi:schemaLocation http:www.fcla.edudlsmddaitssdaitssReport.xsd
INGEST IEID E20101112_AAAAFM INGEST_TIME 2010-11-13T01:22:58Z PACKAGE UFE0020581_00001
AGREEMENT_INFO ACCOUNT UF PROJECT UFDC
FILES
FILE SIZE 21589 DFID F20101112_AADGKZ ORIGIN DEPOSITOR PATH kane_k_Page_38.QC.jpg GLOBAL false PRESERVATION BIT MESSAGE_DIGEST ALGORITHM MD5
c2f0976c33f809308de4a13830681423
SHA-1
858e4acef614236b6722edfdfcda98ff0044ab72
6276 F20101112_AADGIB kane_k_Page_59thm.jpg
6ebe2cc840f6815ec3e1dfc015865674
c070e39c8dab2b01f3959b1fadd446b1f2a56259
41086 F20101112_AADFWG kane_k_Page_63.pro
4d92cd275f24021497d439a5e3a6d263
0182eed80bd29bad463946f1e4dd137e594263ad
6224 F20101112_AADFRJ kane_k_Page_36thm.jpg
cc30c8c8715e21f7bcd7e21f572b416f
889cfacf1943c28d77ff4aaac967d1b297f35825
53993 F20101112_AADGDE kane_k_Page_30.pro
c1ac61955f333d9fa1af28f5f0adf441
c80e3a5f4c39427f29c69bdb155ae16c1cc8d683
6759 F20101112_AADGIC kane_k_Page_61thm.jpg
06b418d154f87afc75d4ae35935b3968
bc274287b0ea4f81e6bcfc4d2f25fe2d7cf6b0b2
125213 F20101112_AADFWH kane_k_Page_84.jp2
e2e54d65ecbae02cb26511c95d6af4d0
5df32dd5429442a4c0faefc803c99ae1f8c88fb4
25271604 F20101112_AADFRK kane_k_Page_42.tif
a6c5d211822f8d79f353df1da67402c1
abe025bc4f48833a2c1cf0ea29adffdcb7dea5c8
25607 F20101112_AADGID kane_k_Page_75.QC.jpg
7f8b2fd9bdd7e2a17bebbc28540be620
8a7f53e95eece1379df5c85e918e9dd6055eca0d
1053954 F20101112_AADFWI kane_k_Page_60.tif
1db7eeb03f1cea7b6a681e1a0d315bbc
c4cd18ccb5a1fbd8496f17cb71031af624798bca
9804 F20101112_AADFRL kane_k_Page_02.jpg
a0db1ee609ea29e2e60dbba42d13bc9f
b2a299bb77ad2cf3d6b1c1680a0b67b75b43d673
34420 F20101112_AADGDF kane_k_Page_31.pro
f70b4b35d62caccc7979c6d42b8bc3d0
afdaed0bff7e9169128d3d591c68c562fa55e2d6
6415 F20101112_AADGIE kane_k_Page_53thm.jpg
fb71d57d63ed4cbe32ee3bbf4cf62e0f
140af714371c9a62e9a09376cba9c0038074de3e
F20101112_AADFWJ kane_k_Page_24.tif
27425b9f9bfdf0634db9dd41ac28fc2f
ee64e4e7ef9416b54f8268e33b627cb69ed31f23
1345 F20101112_AADFRM kane_k_Page_26.txt
9ecd1901b32f3259b6a3b765e0fdf891
a94e76e1363c6deb10774f09872704ef00f85ca3
26337 F20101112_AADGDG kane_k_Page_32.pro
70ce1372880c1df96b8911d160a84864
3a7fcc82746c4f73c6b79941436cb8689d83f241
26557 F20101112_AADGIF kane_k_Page_72.QC.jpg
0ca9e7c3510ae56fbdcfdce7d7eefbb9
c6db2d0df1de9cc965854d407001cdf1317c5d8a
99630 F20101112_AADFWK UFE0020581_00001.mets FULL
686bb87d6253585ddfc0a8d14dc39afa
173aeb33a4666bee372beaf9372f82db4cf62217
942447 F20101112_AADFRN kane_k_Page_41.jp2
f31024db8dc2d057a8de26336eaf700d
84c57943c6010e8673a49251e0c106fa6f27152a
54729 F20101112_AADGDH kane_k_Page_33.pro
da6a649a57048c617d8ed4f923e59c73
0104ccbd17d30756593b619d7b2a811e8eb8dceb
24330 F20101112_AADGIG kane_k_Page_81.QC.jpg
15276115d4b7a02381ac1349579bf076
5895f6f5992a9a005afc6f23fd1749391a6d0be0
5593 F20101112_AADFRO kane_k_Page_28thm.jpg
6ca220fff18ea1153301387fd0f2fc43
bd74d97c7d72a1dcfa33de2ed262ee4d14c07ee3
21003 F20101112_AADGDI kane_k_Page_34.pro
0d6c996f8bc07880fd080e3abdd5e488
2bc061d4533068e5b6b6458083554e3e4815030c
7202 F20101112_AADGIH kane_k_Page_45thm.jpg
2555e37b8af80ab026848ad278b240a9
9bf12dd46e51239ad65f8c6fc2c1d840d90d7e84
14309 F20101112_AADFRP kane_k_Page_73.QC.jpg
43de5fa3fd93c99c42ca0c76549eac35
886c6e570612d033bf3e7e438437ba46f0f06f14
38112 F20101112_AADGDJ kane_k_Page_35.pro
cde39d756223a2918597bdc0df8e2355
7559f406455bbeb079c2f7e1cb60d09e3df84ed3
6815 F20101112_AADGII kane_k_Page_19thm.jpg
a18a8f0330294b673d97fd4e39625d7a
0328e0fd554fdf6aa580b646396f7ffdfa87a206
24146 F20101112_AADFWN kane_k_Page_01.jpg
5b44a62aad617b926cf232e4a5cc455c
577dbefaf60d19ec2fdd48d0a9a762bded0a3680
F20101112_AADFRQ kane_k_Page_08.tif
7ea423d059a37293b3303919856f96bb
db6daf9a10758a12dccf898471cf89ca67d4732f
35625 F20101112_AADGDK kane_k_Page_36.pro
bd20cb7572808e21ae957725e1df27e6
8e59dfb1518b03571dcdc54580ad1741be731a4b
25095 F20101112_AADGIJ kane_k_Page_84.QC.jpg
7016daf2f84b1ceaeac32644c1160d7b
d90e3e890e78bd0d26d19fe90950fd874ccf5224
10331 F20101112_AADFWO kane_k_Page_03.jpg
7b08518de10ce6425335d2dddeaa3ded
be4bee6d752aacdb2e8c76c77bc2fb94d5ccb1e2
1000326 F20101112_AADFRR kane_k_Page_63.jp2
43d67abd47e3f8a5bad260d1a5b177dd
95af72d4762c350148efdef0a04f31de3e22c63b
38124 F20101112_AADGDL kane_k_Page_37.pro
be4f10af9d3e88fb77aca6deba7a0a50
d20fa8c653872626c1f0b352daa32aa162255435
32187 F20101112_AADFWP kane_k_Page_07.jpg
ae61da4fac992251e7d03ed12529c069
80ac1b38c9a406b0003ddf0869ccda2d1c9de55a
67694 F20101112_AADFRS kane_k_Page_49.jpg
170a81f4d5f7b02542f47ccc84d1b987
798f61171769035783d906762d76ad66ffaf1fa9
36449 F20101112_AADGDM kane_k_Page_38.pro
48a176c685e85775a4d664761483daba
f2ac06a273d9640e8bbbe63ea845a2a576367a72
20962 F20101112_AADGIK kane_k_Page_35.QC.jpg
2143047382fa166354ea9722c6b40ca9
ad1996b68f366061f760f3bc1b58ffeffff21b17
88907 F20101112_AADFWQ kane_k_Page_08.jpg
7684769109a75a0d736059fc011da58e
338b91611cf4b08afc15de42b25e54b32f10ae94
4029 F20101112_AADFRT kane_k_Page_10.QC.jpg
4c9db573364f8618d61bf24e994cd194
88fe216e5c8610820fb470e812b772440fb51fb1
33973 F20101112_AADGDN kane_k_Page_39.pro
7c5c2e45ea141e3c73da3456530b9359
1e908fa9a09abd4518e8ca699830203cb6362be9
21820 F20101112_AADGIL kane_k_Page_71.QC.jpg
61f24f4568f11e8effa2b16101f80d30
cc34bdbc0c8a1329484b028a8564ed2b5e3c7648
73284 F20101112_AADFWR kane_k_Page_11.jpg
2f72bc3f71d5f8d824d9fa83d3ca41aa
dd74cc4886073230e04fe62488920e442571ec19
35035 F20101112_AADGDO kane_k_Page_41.pro
50db30fd2b08c7279b0960f1996ad32e
20fd4da71a4e7be7d7e28f1502e714f97ca4718a
19271 F20101112_AADGIM kane_k_Page_28.QC.jpg
08798ade8d61fb5aa101344563d7b536
03522008c727eed8b382267cbef3ef6d63cac69c
93733 F20101112_AADFRU kane_k_Page_62.jp2
470dc0691c4c3623d108f152c8651376
d33236c6ef58b3cb4e9ab0f061861082d9976a04
32727 F20101112_AADGDP kane_k_Page_42.pro
960bbe3924125bea22a9e79c03f475f0
dedeac6d5d51b76cd48dcf3a3c25b0d0decce2cd
70705 F20101112_AADFWS kane_k_Page_13.jpg
4fcab0418abbe3e632bfed3ba38f5ef2
50cb0cbc638c0ea4045ee82c71b9081f93e0977e
4895 F20101112_AADGIN kane_k_Page_46thm.jpg
c5a88d655df8835f7b20d579d3a47aa1
0c4490d3052b2713991da9310ea982cc31d4be28
49769 F20101112_AADGDQ kane_k_Page_43.pro
a7ad75dce41f1e7b5b7d8d333c7b0528
3f4d2b71888d1f5acfd963afd732f9379f8b3bee
69441 F20101112_AADFRV kane_k_Page_39.jp2
fab59bea700d704a4c518fdc66777880
4c473e600a82bb13ef756dda80639226e34b5292
73815 F20101112_AADFWT kane_k_Page_14.jpg
f317f6ffdbc2b9924bcf158c9b8c028d
ee609f22eead4e0b5417d2acecd93f5ee2748868
11240 F20101112_AADGIO kane_k_Page_86.QC.jpg
3821bbba4fc0e12e40f872a2f651d071
12ec1133e29fb5d488032f2b14992773372ae897
54234 F20101112_AADGDR kane_k_Page_45.pro
a4f1f5d5424198ad55c24a6cede91d42
d6c86ce94a769f476b332c17315205596e6ea29c
41942 F20101112_AADFRW kane_k_Page_20.pro
46b7b98b4dc0cdac73404335e095dd15
158e9722039b56697f5731fd55cdc1e28c5b2e54
74921 F20101112_AADFWU kane_k_Page_16.jpg
effd6c910000c5fd2d792a7d5d49ec67
539d7cb13f02c3b36a3956f937386653270133c2
24225 F20101112_AADGIP kane_k_Page_22.QC.jpg
0196d9deb34d3319bda5cabb6fdc84eb
806054e439219050a3db606326dba0ad844d13f6
15118 F20101112_AADGDS kane_k_Page_46.pro
bab68de98d5c0509ac11ba76bebda119
525068ad049624000eddd1fec23dcb9a952e6306
62125 F20101112_AADFRX kane_k_Page_83.pro
ff3c8f038f2b9e57b34958332d5b29ef
f13904a8b8b514ea2bc4e117d66fff9f0cc06f88
76650 F20101112_AADFWV kane_k_Page_18.jpg
ef0f866b5ca6b110395bb48f6678a9fe
b4467d170c70d83d9ba2ab90f40571997c2c3f3d
15590 F20101112_AADGIQ kane_k_Page_12.QC.jpg
cb258e8afbbd4f0954719ecb37550041
abf7a8c6534c1da375b3e242ac7677bbdbd1319a
51548 F20101112_AADGDT kane_k_Page_47.pro
0c4c94e175c532442b31a0b69f0818cf
0e5f8e55e7523c68b73a30c49f2ce4d92b099f92
7016 F20101112_AADFPA kane_k_Page_72thm.jpg
2f89ddfdbd5c49a5e9359b5406b128d8
ddcd160fe11ae81d5204412cbe75a2f53f5aa472
16783 F20101112_AADFRY kane_k_Page_77.QC.jpg
63c5166c4e5aa814c01d85c7d1f5f855
17d996fb337146d1c9ac9aa953158058f17bdf79
78172 F20101112_AADFWW kane_k_Page_19.jpg
d142a814fca4bee2e15ecd5b210e567b
c499c0a7cb2f96a4bb41e3cbda9cf9e6c7e452a5
3179 F20101112_AADGIR kane_k_Page_03.QC.jpg
536e6f8c28f23e1e8fce53ace70859b0
1924f2e91506fc2e783a13d15648d4cf9e9368bf
32560 F20101112_AADGDU kane_k_Page_48.pro
581733ef87dbc4c87a861a6f375afbc4
383898afffc5af22baed57471c620778dd1f208d
9465 F20101112_AADFPB kane_k_Page_07.QC.jpg
8cae814b7b3aee4e68e55cc5e93d2dcc
f81cd9b021a22e58633f3858b21296733aaa8649
32341 F20101112_AADFRZ kane_k_Page_26.pro
2aa72f4a17a654128574e19ab4b886f1
f4a6ccceab381a1528a817080ab6df599500b091
60794 F20101112_AADFWX kane_k_Page_20.jpg
460be895fa1c2e8c27e3ee625fc3396c
7cf8d728f28614cf0eb479f8996b45cfd8ce1c3a
5056 F20101112_AADGIS kane_k_Page_39thm.jpg
c1f3e72e4b1f9e153f3247536752ff8f
87ed6d8b92bdf9efbdc93dfbb51f2d97386eedb4
47827 F20101112_AADGDV kane_k_Page_51.pro
43ccf2a9a71c38e2f6d051dc8c1963c5
952c89da0050fde4842a88a6a306fbd59c630b52
134914 F20101112_AADFPC kane_k_Page_85.jp2
6828a89e2247a388327ec209b1724251
5eac1c6650b2653057a2bde650c46127846aba1f
63584 F20101112_AADFWY kane_k_Page_21.jpg
cbbf596469476226f972aebdfa5e65b9
3cd192b398b1e5d401e21465468de82c880053af
1314 F20101112_AADGIT kane_k_Page_02thm.jpg
66624ec37b2e4662ff4c6de0754c8e1e
88db57d5d79525b4e7a85e02119e9c5b53dd878c
29056 F20101112_AADGDW kane_k_Page_54.pro
748eedeb15a416802f5ca967c13a014f
ae9549d3c527dbbf3410f6a80aafd7b7ceae64c0
118171 F20101112_AADFPD kane_k_Page_57.jp2
57133d91377ef31c3d2214441eba4551
32dcb1d6c09ff1e590cd030a757cf4c370a3ac2d
979884 F20101112_AADFUA kane_k_Page_59.jp2
2e04860b848441deba2097f23c224f58
f0a30fbda667a0af05de4e1e35eeb67bbd63622e
18222 F20101112_AADGIU kane_k_Page_42.QC.jpg
a85f449d331c0ca5a34dcacda1ab968b
80c6b6a9b7851be9a8a63213c8f0d0da455a16a6
52564 F20101112_AADGDX kane_k_Page_55.pro
caf215066a12cfd1031aaeaed1f59838
37443b2bf400d2b0791fa2bc568a86636b79462c
494 F20101112_AADFPE kane_k_Page_01.txt
4686d772bc92aa59463649ac57575be3
fc86bbd48d8856a97593f427b4c1817a9f01c434
50161 F20101112_AADFUB kane_k_Page_23.pro
51965bdc19d9350be6485c72f5bef0b2
93deb88a32651ec55be5e669faea9bea52d01d16
73289 F20101112_AADFWZ kane_k_Page_22.jpg
418fb2cdd9a6bcd497c6edc0e86d7ee1
2d5d5a4cf1abb15d32691f59e686194ac3e27020
6750 F20101112_AADGIV kane_k_Page_50thm.jpg
1d51eb409e3bfc36a2338141dd646f7e
965a382bd433921fe1372febe2e13b96388da352
57387 F20101112_AADGDY kane_k_Page_56.pro
62c79a31b9bca186c982dfc53a16e876
d6c9ab5cfc813c0fb63fa119de985c32d6b923ca
F20101112_AADGBA kane_k_Page_29.tif
ad26c08fde5a3bc11e6fcc49199096c6
0af56d2afc4d13761de9c9f0917b5eb7bcbff047
1153 F20101112_AADFUC kane_k_Page_34.txt
7b28c4744bc5b856423327398dd4da08
1ccdaa52a7f52a1dca437058d683909758111558
84353 F20101112_AADFPF kane_k_Page_65.jpg
e3afd131984f20f147c3481c7d4254f6
2e4112c704e4eba33c08c68e4f9450061708309c
6179 F20101112_AADGIW kane_k_Page_58thm.jpg
59310df3d1a2a3557afb03ced7430c4c
4ed770ace8630737f71764ef14f6c8ab852c263f
55424 F20101112_AADGDZ kane_k_Page_57.pro
6230ae075cd07affa3e24e65d75d7dd7
504fe220aa8225977408243cbea6aac59c6e7f4e
1051970 F20101112_AADFZA kane_k_Page_24.jp2
8bfb6c6c9d922209ed5e31d24594443e
5d8cff50fb1227508541ed68924c61e9637fd8fd
F20101112_AADGBB kane_k_Page_30.tif
6f8abe3d28d2544f7edb961412093adb
17e000eb99a25f0333dbcf26cb93c56bec401b7b
2084 F20101112_AADFUD kane_k_Page_53.txt
9cfcbf88b281ae283907654b685921fb
50663e375ffb5b4aa9a9d9386298bbcdea33350b
25184 F20101112_AADFPG kane_k_Page_08.QC.jpg
4dbf16a5d29dad1df45494c34f6bbb1c
9b431b1ee758928f05dfccc42ad5da67e9ab7fe6
6771 F20101112_AADGIX kane_k_Page_22thm.jpg
39d187bd0f2b1b08302e150fd71546ee
ee0117d58d85083df60b58e5ab7694cf3d9106e9
976287 F20101112_AADFZB kane_k_Page_25.jp2
649a3ce9eb18fe8cbf8d95076a2cd86a
8e66ab7cd68d15e064eab53313465bad01237200
F20101112_AADGBC kane_k_Page_33.tif
7b8e8907bab906c2e93db3bbe645ae73
db9e893bb226d8fe18d2671c32fdc9447743cfa1
77097 F20101112_AADFUE kane_k_Page_33.jpg
2e91ac7ca71d8f68d83f144ba4b9df80
8abb5e5518b6b9a3a74197fa894868218cb64eeb
120327 F20101112_AADFPH kane_k_Page_56.jp2
c9d26242182bc200a7fe9d2aeae02549
fd1774e330b8abdc87114cfc3d8d52210c4e6f23
4428 F20101112_AADGIY kane_k_Page_12thm.jpg
ebeec478c72a1ed89d1e6b004432d818
7081407f943401ebb2a920dc664a149e6eb034d6
2334 F20101112_AADGGA kane_k_Page_54.txt
8634e4b786c5bd20f44195d30b2f290c
20db72be2abaa71cd327ad84518ce6582105413f
1051964 F20101112_AADFZC kane_k_Page_26.jp2
46e0f86a4d0207e249ca75411f1c9a28
2ea77434a5c2151a53788711855802a0980772ef
6847 F20101112_AADFUF kane_k_Page_16thm.jpg
5277cc6acb19bae0d8214c3f2611c528
8c232ef1077ff9e14818e2a5a6d985dbdcfaee8b
22643 F20101112_AADFPI kane_k_Page_62.QC.jpg
fa70c11911c50cd3a99db0d8df026e01
f2f7206f527610f5964fb9edae1b28dc774e3e9c
5511 F20101112_AADGIZ kane_k_Page_05thm.jpg
3b89bbdabe41675b2b8f9c0aaedfad5d
efeaf8395a0ba3436f47e10b03c7b58537336d1d
2251 F20101112_AADGGB kane_k_Page_56.txt
100658cd780cb047dd22f5032cf453b1
b491a3c0d4b703b9bb6da62e45cc0a920b8caaf1
924850 F20101112_AADFZD kane_k_Page_27.jp2
83e47d16ed50e2baf5a96a467ffc7934
5706e706beeaa5c4024379cad81892cd3ff2f3dd
F20101112_AADGBD kane_k_Page_34.tif
eab270d183f73980ed099757d9cd0da7
0aa202e0d27de980fe6258df62e649c8dc802682
F20101112_AADFUG kane_k_Page_83.tif
8028d814bfc14e9c1a5251359d0f8447
106d2ea009391aec2f992609977a0f25f3c8a88d
F20101112_AADFPJ kane_k_Page_01.tif
90ea323a95e232a4488d5b1b328fcd20
62e11bcdbd8e50ed50e2ca27055d6b251180260a
2171 F20101112_AADGGC kane_k_Page_57.txt
09046563ddccfe18d3ac17d215a503d9
8ab659b9fe5fa942ab0652b7f359b60075f0281b
115089 F20101112_AADFZE kane_k_Page_30.jp2
a977617d690549ced5c3f303fa62293f
98c98891c9b7c5595f1ce5b5b82be62aed26e3ce
F20101112_AADGBE kane_k_Page_35.tif
63f3b61ded455b8e8b75a3df9e471284
318ab58fbca8396f383dc18eecdb63b77eafa6e2
19677 F20101112_AADFUH kane_k_Page_37.QC.jpg
8d2769a4c977b2df873d5814a804c6c0
372fe63b280aaaad21f850bc247502af357a5194
73301 F20101112_AADFPK kane_k_Page_79.jpg
53580918cbf1d848972f73be8aa11293
95a56488275dc233e6d51cfbf6326fa6752d2640
6168 F20101112_AADGLA kane_k_Page_48thm.jpg
bcf719cacd0955982178aeae4a5d1a25
401a32250582c58dd634563c7738faf5e53a1f40
764 F20101112_AADGGD kane_k_Page_58.txt
e978c2791c5d75f6e875c6f040b4208b
8bb22a4e072d69266f24aa205c8d1f41e84ffcb5
79578 F20101112_AADFZF kane_k_Page_31.jp2
7940a14be876f6a5988b14a5b653e30b
9c5979a3abbe3e59f3733fdd9ce7b33830fe744f
F20101112_AADGBF kane_k_Page_37.tif
98a5bd76895edf311c010ab332255421
791fe55534a436416f858bfb70f7e1184251a3c8
1365 F20101112_AADFUI kane_k_Page_52.txt
56cc8f042bcd57eed98548724cd25f51
3f7b1617dc54b2653af53624dc7bfca6741ce625
138861 F20101112_AADFPL kane_k_Page_10.jp2
34c1dac29f988f48cc1a6d48153fd6fe
d09ef165f24b16e95c24500da81f220f4bc452bc
23009 F20101112_AADGLB kane_k_Page_51.QC.jpg
a591862078ff3b5227dc0b5503f47d01
33d5d2a94d5c53ba0239633e8bf1624fdcaaef9b
675 F20101112_AADGGE kane_k_Page_59.txt
62de4af09dacc962fc673833e8e68d37
cea7de0eb351ed3228653692028ca3fc8b0aac2c
945651 F20101112_AADFZG kane_k_Page_35.jp2
b4dc4b4c22ea6d8566771cdaf38ffd94
a2c8b0b4f789dc3c9db1b6b66d3ae765f513ba6e
F20101112_AADGBG kane_k_Page_38.tif
b3daebc51ebcb4dd0f9e113cf2d6cabb
35d617553b554cfad2a7349c34798906e585de11
1899 F20101112_AADFUJ kane_k_Page_51.txt
f6b10b6d0f5a4570b34c5ddb651619ed
7897cec6782840738e7e8ccefcd7651d22790288
24086 F20101112_AADFPM kane_k_Page_64.QC.jpg
4496fc503647da66f05e995a930a871d
51708dda6a24bff10c69839179ed1a3c129383d8
6972 F20101112_AADGLC kane_k_Page_56thm.jpg
49edadb8a6dd7495182e1977da7f9acf
b2bf3c61065539db35f71a2d70c64bbcfe239bb3
2099 F20101112_AADGGF kane_k_Page_60.txt
d6f6fbb13d66eada65867d7b6c46b3aa
cfd49b4b756378520ebc75ebb3cafaddae13fdaf
969039 F20101112_AADFZH kane_k_Page_36.jp2
28e083b9a51d0c28c1c95b5408330cd5
610b1a3c7188b645a84ea57e654f1af7e02c4467
F20101112_AADGBH kane_k_Page_39.tif
ae535d85e71bf66a2d3786a5ad9b4fed
aab257c394bbf3a8cfe16ae01e9c05c6bbde20a7
73057 F20101112_AADFUK kane_k_Page_64.jpg
4cb6aca870e0dfe54df31c965664977a
6c50bd4fc88fc926a5773ad0e350792dfe9fbc30
F20101112_AADFPN kane_k_Page_36.tif
bd9a14d953f77124331b403b2ff291ad
03bd622a01d9017c452712c20068b866912b91fe
20344 F20101112_AADGLD kane_k_Page_59.QC.jpg
c30dfaf560366cc37474dc439ede3268
a3432a959363c4e68c0d1b2a8127345250133519
1286 F20101112_AADGGG kane_k_Page_61.txt
a6164ce216fac89c016a46c14bc6702b
964d8dfeb06c6a123acf1c625aba46935cf5dfed
1036533 F20101112_AADFZI kane_k_Page_38.jp2
00757129e86be20d5692592d9ca65340
2162dd37e1ffbd394743910944dcb3ab9a33d2f2
F20101112_AADGBI kane_k_Page_40.tif
ce839baa12b223ee26ecacc61e7c2085
d408ae67ea56a552d64fa7dde70ea8d0cf9b703e
784626 F20101112_AADFUL kane_k_Page_58.jp2
c4f1cb6abe8ea28884d36ca2825ae5a1
4bd1931adeea4306096df5d8971262b9a38076b9
78343 F20101112_AADFPO kane_k_Page_68.jpg
ffa85feac54fe9c174f2df05d81b9e22
d3971135a363c4b6754ddb6aa96e282fcd4455f5
24691 F20101112_AADGLE kane_k_Page_70.QC.jpg
b6792a3641dddeb71ce71778a3e9260a
c55a1fab1d9045b5b2f12d505a850294c756705d
1970 F20101112_AADGGH kane_k_Page_62.txt
47fe5a634b4d099f6c4ff3b49af9956a
db3d36445130480f8cf7648cb9680a7d65f1c111
762334 F20101112_AADFZJ kane_k_Page_40.jp2
4afc65699dbae7d0220ea6a67809efae
278dc05db6b3e4d3f6b354d13d94850068e872c2
F20101112_AADGBJ kane_k_Page_41.tif
886b919e3b860d0701390729874784be
ccee1e0c8ff6712cbef3f3c596d759c49dbad88a
110453 F20101112_AADFUM kane_k_Page_64.jp2
722a45fbdf242abb603a408a6180442a
aaba7cb89d55274dbdd6f95782a399c7f92e9660
14746 F20101112_AADFPP kane_k_Page_46.QC.jpg
e794ee2ad0e8a7b2271c1813c97b2520
5cecb140ebb86163ddf281ccfcc854e3aa7661f4
24465 F20101112_AADGLF kane_k_Page_76.QC.jpg
ec4a26e684ebb63df6a9cafb13ca1c5c
d9211983fc08a6b8b346969e359f44abf5534cc6
969957 F20101112_AADFZK kane_k_Page_42.jp2
e5b4c8be35c8fa31571d3581920ff17c
2518a46955c365a199113d4ba1d57e36a04da928
F20101112_AADGBK kane_k_Page_43.tif
0a3ba90c90d2def444b9d14b61aeaf1e
0d7eea3705fef4d18932a49ad4a4f11aada55968
6645 F20101112_AADFUN kane_k_Page_66thm.jpg
dfc31e6fd3d3ea44eb4c7f41d5c5837b
002e376ffecc3f6128f8f5784f804347bdc2a8a1
F20101112_AADFPQ kane_k_Page_05.tif
59012e9405114d28afef003651707e0f
bbe84260741b14316488f7d531fd74d572e017d0
7549 F20101112_AADGLG kane_k_Page_76thm.jpg
0729945a11c8d1d179e7e2aa680e3155
fd8447689330ce84245a71116a00d4a667595e07
1743 F20101112_AADGGI kane_k_Page_63.txt
cfc486d3fc287376ff75b055e18054cc
e7e1d690d35d5352e329f5d348d7bdbd9221e74c
116338 F20101112_AADFZL kane_k_Page_45.jp2
373cca5443cdcc14c2e6ed387945240b
03729ec744bc7a15bd3b19d8aa985d3cedd45821
F20101112_AADGBL kane_k_Page_45.tif
9247e81c3b6e6cbf7314cf4b15ade9a7
1de42e039117024ec2557cbbee19a503210f5405
F20101112_AADFUO kane_k_Page_12.tif
94c23c8b4f52e3cfcbdeb3c19cb67e1a
420e8e4021398dd02acec18e246ccdfb1c0044e0
36248 F20101112_AADFPR kane_k_Page_82.pro
f6c89789ee35401b285fecbad56c92a8
835040f4d21df987e3665921481374368706c200
23857 F20101112_AADGLH kane_k_Page_79.QC.jpg
00fca2cfd7441c32e2bd3bc6ff3a92a6
cab72abf7e257a9b7f8afcd7a1a7e65b8336e100
2098 F20101112_AADGGJ kane_k_Page_64.txt
b95ac06090356adf4152f60fa2d983a3
4900569a78bf31339cf5fda0e74fbafa8a4577db
766069 F20101112_AADFZM kane_k_Page_46.jp2
0d1823083cd1b32454dc60ba1a8c736c
eb54a928c578d7218e9d2704dcfe7364e3d0c2a8
F20101112_AADGBM kane_k_Page_46.tif
681677442ffcb0b7a12f1a76cffa0e08
2a6e7825b8243a4b3b2760dfa6c0a59268f54a72
25446 F20101112_AADFUP kane_k_Page_18.QC.jpg
43fb484c62993b1913c4f5c0c444b993
f46ae5b9f1bf081f1eb2b7423f0cf624486f8a17
6617 F20101112_AADGLI kane_k_Page_79thm.jpg
a15d332e799d3623fe3d434e7cce58b9
a6597847ca2bb2fb99f4e342b377fe49a295b4f8
1787 F20101112_AADGGK kane_k_Page_65.txt
12684234d070c4764c136fa3f48b0c25
a185b43c7bd71cd0ce1e50550258b5925204437f
110202 F20101112_AADFZN kane_k_Page_47.jp2
eca8d00fdfc22dda9961bf5978dbc37f
6018455f5fb28b14cf34f0caf70baefa4356cca6
F20101112_AADGBN kane_k_Page_47.tif
4833c0e88e1026abd1d2db8e2910c2fb
e0b2558a842c5e6bf49e7557173cc31b0ccc27ea
779 F20101112_AADFUQ kane_k_Page_02.pro
67b1b51df1b72c12546e54d5e4ce118d
98fa0abc9169e424d8d24a4e7dbd0235ff7fddac
19813 F20101112_AADFPS kane_k_Page_54.QC.jpg
b70a67cb0de00e2f1a76fc3c0cc81101
dfec36c1e4c08667d7640055bd54168d00e879b2
9261 F20101112_AADGLJ kane_k_Page_80.QC.jpg
dfcf3e6032ea5b1c485ad3c0c9b65cd3
ba3cb379c98157d0e6eba942b1ee54b764c1a843
1588 F20101112_AADGGL kane_k_Page_66.txt
27905b542b8ab24ffe4d0fb92fc717f5
e66772c4007e333b6f8d16cfd7e63a205f6fb2cf
F20101112_AADGBO kane_k_Page_48.tif
e169ce5fef1419fc7e79ce20ae4bb974
f8ae08977d5c03f3bad4057de495c33ea96e0c4b
5943 F20101112_AADFUR kane_k_Page_41thm.jpg
9e5d716dfe06c178cde100ffe6d2dfba
8549b62e85b25e3c034e9b3bdd165b8f24ff4121
6976 F20101112_AADFPT kane_k_Page_63thm.jpg
b052293e27c5852cc8fa1aa81bb5fbc9
d3f6551d7d07bbdfd4f80f8efb4cc09ae0d7b412
963539 F20101112_AADFZO kane_k_Page_49.jp2
0dde758cf11975fa6c58b5b52dfb6726
b06e642f50daff16988eb9509402627cac0bee0c
7166 F20101112_AADGLK kane_k_Page_83thm.jpg
7d59a738421108b9f12344a2eaf07a0b
ac283a51311842b82bb8f284221e16741995210b
1593 F20101112_AADGGM kane_k_Page_67.txt
a7ad10fcb0ca5fc705991f93b6423825
f2735d8aa843131f4f1afa2579891a1ade626959
F20101112_AADGBP kane_k_Page_49.tif
817dc7b918efaaea2c6a529b00b7e346
e7e93ab031fbe23e47f1e368b5659f1e15ae8fd1
8733 F20101112_AADFUS kane_k_Page_01.pro
86eef6f38d72c01f7adad51c0a63c335
40537ec6eae10a172ffb349ae76dd70cec8c1d03
F20101112_AADFPU kane_k_Page_81.tif
5b90570d508f7bf541e1b6e14f661cc8
50d501264f27afacaa1a787cb49749bd01f38e57
105386 F20101112_AADFZP kane_k_Page_51.jp2
1fee1518ea09904078d3cc8a13b1414c
94c0f93a3ab0cd8e6e2744fc24d75064c1150f23
7043 F20101112_AADGLL kane_k_Page_84thm.jpg
a8db82fbc03e81d9b19c4d96f3b46b32
3f762a99bc8f1209cc24489acef9980aed807874
2161 F20101112_AADGGN kane_k_Page_68.txt
59c17614b898db09eef0ca56c053972e
e78f940b2ba9f3b2a2d57c9d10cf12e57e4681d3
F20101112_AADGBQ kane_k_Page_50.tif
3d2c1e8ee13383acecf34553e381a691
f503478c467c61dce7ed45604f5b900b8bb0ceaa
F20101112_AADFUT kane_k_Page_80.tif
01defe14a0a655242be5e02a891373e0
da43df1e19791bd6d81b33b4b1d519c183f4005d
6426 F20101112_AADFPV kane_k_Page_52thm.jpg
fb812e89a04cfc0f0aefe64a02e7712f
76ae07aa2f53bd84b9ea447fa162a21b9c5c72ea
976017 F20101112_AADFZQ kane_k_Page_52.jp2
096694e4b48bab64b622177ef1eab1e0
b5c75e58e64da2a91073cc6adac4b1e2c7cfee50
26914 F20101112_AADGLM kane_k_Page_85.QC.jpg
bade5fd95fa01a19ac95bd13c0da1bde
5055af71dba54f6c05f854b932da93e2e56692e3
2256 F20101112_AADGGO kane_k_Page_69.txt
1dbfbe9f49a4505236537d2f5fe66300
03fdd0962e12357b8616adbf77078f68946ca4f5
F20101112_AADGBR kane_k_Page_51.tif
4ab60f30a06664bb1e379d31b4383def
ed562d553a19dfc9a8dd38b73f3dbc6a41eb54ef
114103 F20101112_AADFUU kane_k_Page_16.jp2
beb6b0c507aa290273911f87dc256e58
3bab9d38111e29cc741ec652880a81501e99bb58
65721 F20101112_AADFPW kane_k_Page_52.jpg
a03dc2a3235f99d47348344ff23864bb
a3e551411fa8666369108294759eb8c977be05ec
106102 F20101112_AADFZR kane_k_Page_53.jp2
c7865dd40a676b8137271aadb1c0335f
e9d89a2b9d754bcc7d364a6a4af110521bff9efa
1965 F20101112_AADGGP kane_k_Page_71.txt
a7561824f0bc65dfdbc1975c7f039aaf
8c7d0caa7feb8e07d83d63f1e2221fffb3d80c37
F20101112_AADGBS kane_k_Page_52.tif
495eef7cb20243f57a0ca37fe457a4df
bee147f7a8922529eaaaf61ed5d4a5c25d0e2923
6572 F20101112_AADFUV kane_k_Page_81thm.jpg
43b5b6cc98c560720bdfe200712223dd
21a1481da24e17c17c6a3f83dd303a2cb703c012
117613 F20101112_AADFPX kane_k_Page_33.jp2
48ef84d333d7cf3900cd2ae42c48b775
fabd730a04bd0b00d22efee288a93f3cc4b1ac1b
954735 F20101112_AADFZS kane_k_Page_54.jp2
eb54e46ff9df94658cea29a165e5352a
6239bb90b244dbc4aa658c82c189975af835a090
2267 F20101112_AADGGQ kane_k_Page_72.txt
7d984d874dd770e2adfcd8308be0f1b0
b67a4c1378ad09d7fd914f910e2344b20e373d00
F20101112_AADGBT kane_k_Page_53.tif
a3b824c1e8ef749f7587051d088ef027
31603e894f051b287a7472cf1786136f62800e9a
20228 F20101112_AADFUW kane_k_Page_20.QC.jpg
83ac4c3aa2194bcf7729f4321a20ecd9
5ced9eb433dbd7cc813f999d58e0f85fa4ab28c2
24399 F20101112_AADFPY kane_k_Page_50.QC.jpg
d32dd76e0be16aadc52aefe141e44c67
cedea27b304b85a46ce409b7770ee7fdd1f23b33
111956 F20101112_AADFZT kane_k_Page_55.jp2
7c784ea5e8389af6718783c7e25d275f
b6d7d33b3be819d85a32897c8b17b737e2ea51cb
1198 F20101112_AADGGR kane_k_Page_73.txt
7fa04d3ebcf48d391f9086d1e30ea3b0
a34048f41e5e2388f68825bb841d45b7744221b5
F20101112_AADGBU kane_k_Page_54.tif
53112a19115c4f272ccb10b2cbabd746
cecbb8bba1f684836b57554f31c2b8b9355d4e56
F20101112_AADFPZ kane_k_Page_06.tif
9932a60fafa26a4f44b724295f3a2af1
267c34d6cbf4038903f6cfab7f07dae33aa226d9
106273 F20101112_AADFZU kane_k_Page_60.jp2
0327815359ba1f8137f9c418ee3f0142
920170f400fe68bcd04edda76b9a978dd15b8c21
2275 F20101112_AADGGS kane_k_Page_74.txt
1b07ad3a0681febd39c657139668dcdb
ce3c2df9551d17f43ad381339613ff05605fd329
76257 F20101112_AADFUX kane_k_Page_76.jpg
7bb1ba9d78140673c756a127682201e1
8cd6ff451fc48a0c7aca3aafec2e029e7e17a069
856141 F20101112_AADFZV kane_k_Page_61.jp2
fa671840b3650deda92ff28f0ffd88f5
9ba2282c21f3a4507a04f75a3a255dce53b427ec
F20101112_AADGBV kane_k_Page_57.tif
dfa0c991184c0b7b6eaec763bdb45f8f
86699f1e5a508e5992a6bc24cee94429c16b80d8
2214 F20101112_AADGGT kane_k_Page_75.txt
635ede3360723c233ea8ebde655f9a61
b55309e8dceb56314e334d7d23e360350880babf
1018678 F20101112_AADFSA kane_k_Page_32.jp2
a678f4a3d319b433337b00297f3a6643
97cd5a4c1308d27b5f26007056d90ef9ec670b63
1051954 F20101112_AADFUY kane_k_Page_05.jp2
9346d35bdef7b58b414d4eb7bbc5c0ed
191e10db70a6de96724af3de94dde02f5db9b49c
1051962 F20101112_AADFZW kane_k_Page_65.jp2
099eb73646a80624a7355e08adc998c9
8986945026f329d667ff3af51ec903def237635e
F20101112_AADGBW kane_k_Page_59.tif
6772838475eed9ee6b93ef4a034d41c1
c49c79f86eda5ec096828c72a3c74de6ca40f403
724 F20101112_AADGGU kane_k_Page_77.txt
5f51b54aefde5aa6a8c959d262424689
4c553d9ecba5bdef8482b19337f88d48e4fd2b4f
F20101112_AADFUZ kane_k_Page_18.tif
4d089f899f82bc29b6d908e311f6187e
3b9e610fcb1767a72d5e6a6099fc4cd8cb1ada10
1044450 F20101112_AADFZX kane_k_Page_66.jp2
29a83d2fe1bab9fc97662acf58b083fa
52570b9e574365d585b8c45b08903d6aaa652e39
F20101112_AADGBX kane_k_Page_61.tif
29784488e540751e85251f7269b2c02b
69e4f9a922da69e4c4d59d73d899c64fa743b576
F20101112_AADFSB kane_k_Page_03.tif
a2d1ec3fb3d68e741d1f2f8e780897ce
45a86c717f61a285c4a946d571fbf35730e0b6f1
859 F20101112_AADGGV kane_k_Page_78.txt
e814b174eb166324f95100d4249f5f46
c7f07ef4ab016d09a682e580c2ee7cf859537cf3
942954 F20101112_AADFZY kane_k_Page_67.jp2
c24c656e57b1ee40674adbc38725ef59
e3ba4a2107a7ca68adff6740fc24f4004996eab7
F20101112_AADGBY kane_k_Page_62.tif
8faf7094113b720983cda68a484b29a4
e86b3f0294f622b177a8949600246ef897866dab
2173 F20101112_AADFSC kane_k_Page_45.txt
671b6394a03fc874ddc4b53a5f4cef02
9d5186408e5a229d9762b8a9c0f8655041e47366
2020 F20101112_AADGGW kane_k_Page_79.txt
22553027d13ead5749b4cf65eddc5958
a05d771fab76e0a0a38f600f15db604568a9c8da
71695 F20101112_AADFXA kane_k_Page_23.jpg
8aaa8b5a994ec56e6069be06a6381c5d
597d2c0a5a233ac01f134bfb550d74363ea56c12
89475 F20101112_AADFZZ kane_k_Page_69.jp2
d3f3b6e00652a78f63cdc35a96a574fa
098eaba913db8710d8c4a7e1b38ba34bda7cfd32
F20101112_AADGBZ kane_k_Page_63.tif
603acedc4d78f6ea212b25d454960390
38610603ff979507224ab708242dc65234622677
51936 F20101112_AADFSD kane_k_Page_06.jpg
0bf927ff978d8b6269b683bd9c7f715d
4224738272db2b4fcb1d6bc6cfdee0c7f54926d7
624 F20101112_AADGGX kane_k_Page_80.txt
d19b5fa7895a4e9bff8355577cd85283
267392daf82171d87336a0b21aae2e4e4f13a2a0
65505 F20101112_AADFXB kane_k_Page_24.jpg
071a878e94756e2eadd0f05dff2efb4f
86f28fdb6113f4a8e1cb4704c57ab7f5db595aaf
1832 F20101112_AADFSE kane_k_Page_37.txt
63fc141807a3d2049081682ed583d9b2
090f430c6ecad55ff3dd916ad109ce64897c10a2
2085 F20101112_AADGGY kane_k_Page_81.txt
a477aebc712afc4d08121d29bf7048fb
58fcf4bbac6524f189415abeda73142c7d1bc492
12658 F20101112_AADGEA kane_k_Page_59.pro
400d7bfcee75a8ef7573d8e5a994d515
ba2f213ef55abbbee7be6926af21ad021b0c5500
66929 F20101112_AADFXC kane_k_Page_26.jpg
7844cb0bea49c37ee2f38625ff4248fd
bd305d24dbb59909cdaeb9199b1ab835fa39e6e3
25072 F20101112_AADFSF kane_k_Page_30.QC.jpg
438ff1edc087eea8a556654b320638de
085a7d2d9831d43973952610ae657d7c297ad86d
1490 F20101112_AADGGZ kane_k_Page_82.txt
d73d46e889f22a16e452072fa15dbb16
506034cc22e5631091c73b0aa9ea64899b65f229
50137 F20101112_AADGEB kane_k_Page_60.pro
23e3c6bf0d981aed1b45d041fea7a0b4
4731ed89e41d2e2e988c9f40b43cad07b8a3d417
67377 F20101112_AADFXD kane_k_Page_27.jpg
e3a5e68ba9e440eef8a3cd74f6529164
dc365f6804065ca72aedb1addec382cf3234ab27
700 F20101112_AADFSG kane_k_Page_46.txt
76faa9decc8817556d8b7ddbd55416f0
149bac242c3fdb4210e9320223561914954492ae
31325 F20101112_AADGEC kane_k_Page_61.pro
98f8dbee9e0cdcbbc70cbb0e763b6d67
29d2676d06792db5e643c2c9e067b76b5a2af9e6
58658 F20101112_AADFXE kane_k_Page_28.jpg
11f96062174db15b992bc8381b3e535f
767d61ca8d3b4f954ac66fa5b18cefd667a85841
2140 F20101112_AADFSH kane_k_Page_22.txt
3f4efa883a6da86a5ad8e90354258999
371b0117672ef692746e57db5bc4227796e6df3d
23056 F20101112_AADGJA kane_k_Page_53.QC.jpg
51ec6e20ccfbd45b87bcc161c04bdc63
fc6e8db4a959b59f6ed3f5c75c7fc9f117414797
44929 F20101112_AADGED kane_k_Page_62.pro
c6e55994314e3a76b1637ff39a1bf725
6c9be9f71eedcb3fa9c60bdc4ea54a48a50623a2
74626 F20101112_AADFXF kane_k_Page_29.jpg
573e7dbd01c6324b67b4c3ab06637029
58a295f891c0cde6bd02ae07b15063e46eb3d87a
F20101112_AADFSI kane_k_Page_76.tif
66d67d14f7894429ac448aa9d5db86b4
f083ceb46b662ee54b84e0e95a0b7467f6de1909
6628 F20101112_AADGJB kane_k_Page_13thm.jpg
31881f052314b00e97e4148dacfe5926
ebdcda87f41998e122f4b42d37bd8d9f92e34321
52088 F20101112_AADGEE kane_k_Page_64.pro
612c76887fc3699136233e60f5ab701f
e5d595bc8697bdf744814bce849da6591262ae91
75613 F20101112_AADFXG kane_k_Page_30.jpg
e50d6441c236ebc8680b9abed87a37df
c0ee5205e1e280a52676f600d3a917e887da17fa
23070 F20101112_AADFSJ kane_k_Page_43.QC.jpg
0c2ea8e4f962037edb8281326c9b803f
41d1c433b3de5348dbdb103c281a6b7f84e5d78f
17702 F20101112_AADGJC kane_k_Page_82.QC.jpg
91d49109bf8e9e2d64034b4c4eab7e59
c1ae19f6e1b2215132ec649d34ab7e96a9e1efb3
39188 F20101112_AADGEF kane_k_Page_66.pro
b65884bc4b9851bbe37a191d0e5387ce
a8101749bf6d1db11c9d659c8c806fba9bbd3820
58410 F20101112_AADFXH kane_k_Page_31.jpg
5e3d8b0acbfecb1916c84384c6a0e3ec
d5eb198dca9c7e9cd1a1d5998c779baa4ed19574
47696 F20101112_AADFSK kane_k_Page_12.jpg
e72f0e237dc54e1790a997e64132bc8f
aab45b6c1bd530463ffefaa5255d2409b58f1722
6705 F20101112_AADGJD kane_k_Page_78thm.jpg
552ebb63a98f0dd1ce15fccae7f0c23e
3bba360df46a298f95a5615183545af67303c975
47644 F20101112_AADFXI kane_k_Page_34.jpg
426dd01edca59617ac0f12c9fc3259a4
11984df7ae6aa3bf6ce1efa517704cc942c11ff3
106325 F20101112_AADFSL kane_k_Page_44.jp2
e501502ad04f30d3b76153166b9635e8
b0f0df0106556a3443ab95d26b5c343faf8f0693
6019 F20101112_AADGJE kane_k_Page_37thm.jpg
a3e5bc33f75c58a38344530af061e988
d43c96d5a8ee2b190c93feae32f8fa948bd4412c
54890 F20101112_AADGEG kane_k_Page_68.pro
6825b694d42e3bd9d0e8ad9a91cf94f0
3649d3fe7fb9e26454eb42b0238a648fbc2f456f
65338 F20101112_AADFXJ kane_k_Page_35.jpg
2c5670a344939e7e3588283890a7c82b
bc86bf181abe481d6949f07bb02f86c1907c9280
50431 F20101112_AADFSM kane_k_Page_73.jpg
5c541aa7c961400d0747c179d1c6a038
565b006c6868fd316cafb897fd182b2c30a03a40
5435 F20101112_AADGJF kane_k_Page_74thm.jpg
effbf0a42925dbe3130aa20e959d3a36
c865aabe45d49d486d54c39e1d9c1d1a2861dcf3
46741 F20101112_AADGEH kane_k_Page_69.pro
831d8c1e5d4d55219f7feeff0d4f56e1
bb55fdec08e3da88fd5ba0c394b5e0d908b4a8aa
63359 F20101112_AADFXK kane_k_Page_37.jpg
efeca1510f2489ab5b0303c2c3700c09
6fd6e8393d0d32cfcb4ed59c985a9a782d6b9d03
F20101112_AADFSN kane_k_Page_44.tif
e1a6b0c154297afac8882dd12ed8e1c5
8fbd9939d0f19ea18a777c0c97bea8486111f74a
7412 F20101112_AADGJG kane_k_Page_09thm.jpg
6fa160ad456c996cb6be89980a23fd57
d10f2bdaf6effdc5f079720519eda0c1da40da02
54848 F20101112_AADGEI kane_k_Page_70.pro
b7ec91af9584f73169a5bef639f98444
c43fe0af542f69b939cba0b9d5100ce307319c3e
65701 F20101112_AADFXL kane_k_Page_38.jpg
b5a135100b79dbda7a2e01c2ced57ad8
9df0cbf3ebbea5028b23e3dee86621e620cc4476
12788 F20101112_AADFSO kane_k_Page_58.pro
0048ca2dcc8d7b99ad39b89ad4ef3035
945da538c0eca163587ca5e03e1e66b82be05f10
19663 F20101112_AADGJH kane_k_Page_41.QC.jpg
72116bb5606b6e3c962bc6a694e5a30c
86637ace0cf8a36c37eb574288a809f4e4203267
39254 F20101112_AADGEJ kane_k_Page_71.pro
7ff5d9f85ed39bbbeff6bf473f20af7a
ecd36866451dcfccc4850d09be41ce4ac87b8cc4
46092 F20101112_AADFXM kane_k_Page_40.jpg
225181b0e10b5c70dcab5a0d604b17ae
792c94527860bdbac9b6b9e9af6764b2a8a636d1
6821 F20101112_AADFSP kane_k_Page_60thm.jpg
c1597ca403b488a0d8491a6c99b93425
5d0597f155ed037e34845e0fdcd65a8c034bc301
16869 F20101112_AADGJI kane_k_Page_39.QC.jpg
11d91f7c848751068da0cc7f2732df34
7731d16a51e7dc872885275fcfb990523f2d7d9c
23922 F20101112_AADGEK kane_k_Page_73.pro
9ac031ec66f17bdac3d37ea2eea573d0
943c392f67f3c9d3c2e1918d9066b7884f43a146
61066 F20101112_AADFXN kane_k_Page_42.jpg
198fda1d6afb19a36768443e6b18d3ea
b6c7e6f35194d6bfc5e99dc07bcc77562142eaf4
15962 F20101112_AADFSQ kane_k_Page_04.pro
1625cae369ea4118ced987b3f5801a5a
0355fa38eac86e0666c889dde224e92e88bdc6e5
25572 F20101112_AADGJJ kane_k_Page_33.QC.jpg
e53332397eda7d0601f36db791072b5d
4beb91ed9d13f3172403d57caad56419dfbfb2db
47573 F20101112_AADGEL kane_k_Page_74.pro
ff20cfa6bd7ba5fdc3b78cbc7bfa8607
a906254ce7621302698b8273164e411c800cf276
71611 F20101112_AADFXO kane_k_Page_43.jpg
f1938ae4076bca2bcb5d53fd0b88504c
9b93dd7223df1f4bcb2e4da1c79393cc9295ca17
49921 F20101112_AADFSR kane_k_Page_53.pro
5ab2fe327da0a81d5b1491c3b114154f
6603ffc8acb0d41ec6ef9354237df9dfb6626f28
19037 F20101112_AADGJK kane_k_Page_31.QC.jpg
25b03c88e83bddcd25a5371e60c188c6
e37be4876797515cb5e86c4be990cefddf3b83c3
19785 F20101112_AADGEM kane_k_Page_76.pro
8086d0c8a1528f7d8ae65cffbe8eb8b8
6cd9c05070503b572abe32641b603edd392332bd
71402 F20101112_AADFXP kane_k_Page_44.jpg
842d424a0d1e3feff7f7285c481b09db
4c2500dc4672b4e6d77ff7e8caf0fd7a8edf317b
1371 F20101112_AADFSS kane_k_Page_03thm.jpg
7b9eb3bf145fd3421e7b04eba5b5f284
1c7fb362b0b941952306a73b87f0b3d5920c80d3
11764 F20101112_AADGEN kane_k_Page_77.pro
f31af8fe7a513506b7676f5dc19b0353
946e4b927b95f65c5aad9de1d6f0f8e425558da5
76085 F20101112_AADFXQ kane_k_Page_45.jpg
290c20b2de8601c24d7459aadbb93903
d579a51a3afa9cbdaf830877a841fc6e29d306e6
26695 F20101112_AADFST kane_k_Page_17.QC.jpg
626dc99d93fcd25ef295ff09cd46be58
39fe250331e62efa9405ab60c9c77dd1fe5817da
F20101112_AADGJL kane_k_Page_27thm.jpg
318b749374051aa42d0e5d5f89667880
7d7af4deb9d1d78192774136a6ee31c9d1fcb99e
16571 F20101112_AADGEO kane_k_Page_78.pro
78efafb0ac6fd09dd578b384f429bfb4
1fc2ecee3d3f4f8dba84bcdd5fec45923fc77b41
45991 F20101112_AADFXR kane_k_Page_46.jpg
e57cdf32fe46424b3c2b3bec7eb30791
d93e2ab0323212482fa23bd5c59611ecf937714e
62078 F20101112_AADFSU kane_k_Page_41.jpg
fadff9ac8195b8c5d69db207c2e197a3
fb95cde9352cf3dc83981b610b791ce4c7e7f16e
21887 F20101112_AADGJM kane_k_Page_05.QC.jpg
84d2132ad07574ef65b5679a107a771a
6fbb1321a2065ee1356e3f2b24d2ad939b615da9
51254 F20101112_AADGEP kane_k_Page_79.pro
598cfad1a3c9ea2f14ada84c839fc8bf
509ed262331a49e15077724c338025d440b5e7e2
74453 F20101112_AADFXS kane_k_Page_47.jpg
fdfbbe5a7bedd00e476d7bc8b58987aa
4ed0bd6b3bcb23ae0acab60721a7eb86a7192bcf
25434 F20101112_AADGJN kane_k_Page_45.QC.jpg
2e94e5fe0313173c8295d1240fdcf825
77f27a6888396d3a5373e92b1f171bcdf046ae19
50605 F20101112_AADGEQ kane_k_Page_81.pro
8bc52b4fd8858cc388296ba43cc3d00a
dd8c2c21f1001fa4b960a9b453c194b80ce4dcb2
68563 F20101112_AADFXT kane_k_Page_48.jpg
0e131bc136a4b29c21fd2e4ce69da296
c65f081a803c472cdcb327ab8b17ffed7a74c1ab
F20101112_AADFSV kane_k_Page_82.tif
1e02ca2b34edcd43e7d84678047c5390
a1daec07a1bd3bbddc11c8a02bc6d2e8c83907da
21374 F20101112_AADGJO kane_k_Page_48.QC.jpg
537a8b351cfa71486da3eb3a06617a51
cbd051ad6979df4c31a7b5af3047719aa13063c3
64685 F20101112_AADGER kane_k_Page_85.pro
7d3fad755df6d3f7cac0fcb227023fd2
0325dd3f4f164e063a047e867d720da958f6d03b
82911 F20101112_AADFXU kane_k_Page_50.jpg
c069bdc52169c33a1137a9fbf7f88fe5
d2a824b1a96b5bb6bccf99c03a9294d29d8d9a2a
2159 F20101112_AADFSW kane_k_Page_70.txt
c496521e44c459e4cdb060962be56f95
66e97413608494f15585e1dce1f39c6d1810eadf
21615 F20101112_AADGJP kane_k_Page_52.QC.jpg
3d281a7db0061cca01a68e3686a7f4cd
4feb040c17fe596a963a6a8fe3ce76aca52b0f72
18801 F20101112_AADGES kane_k_Page_86.pro
fd3271297b7977d446cf4ec9683a6577
47497e242affe7ab22e8dc43b81fb0d00e927762
69289 F20101112_AADFXV kane_k_Page_51.jpg
d144f9368b57314742992e2a9918b237
8fe1870276611a46904c51f2e646570543022a8b
34358 F20101112_AADFSX kane_k_Page_86.jpg
27d9fa4fb87915beae66f5edb1ee53f3
802a54951474edd911909d0c3d09f7abbb22bd2e
21597 F20101112_AADGJQ kane_k_Page_49.QC.jpg
c7972b94a9740fe167ece5d74323351c
24e77142083cf1744107de769a8674981dceb903
102 F20101112_AADGET kane_k_Page_03.txt
11b6490c3fc44a3b07e39fd16b670d8b
2fef747bab5233fdbe5126d81f6eca955fe8ae33
68952 F20101112_AADFXW kane_k_Page_53.jpg
1343826a1d10a571992ac8c78629300b
20d781ed63a9df196acaa7ab238c3f17a3edd5ca
6765 F20101112_AADFQA kane_k_Page_14thm.jpg
5dd0ac1738a0f1dff84324609fd1ed9d
a61a17d7d5eff3c24e368c064c002e31f2c9089b
122279 F20101112_AADFSY kane_k_Page_17.jp2
d2123356fbeedd815666765403131f80
a80fc519add9340af835131d3a8b0453d14187b4
5696 F20101112_AADGJR kane_k_Page_69thm.jpg
ce4fd798e2cbaecb442b6fe0d3891c60
8609ee0c1a307c5ec4a71c7aa02f957a6a497bca
3696 F20101112_AADGEU kane_k_Page_05.txt
dd27f4378a055dcd1c7ebfcb1debcd42
6bdcae4184b13d23c254413fbca32d81649abf3c
60673 F20101112_AADFXX kane_k_Page_54.jpg
8483166a6301854263fdf3e98478abea
37ec74323923b5c43895bbe1984fb63c06abe084
84323 F20101112_AADFQB kane_k_Page_74.jp2
dde95f91648d67388a14aa1ac9227698
21b121722d6e65881d6a693b3b4bc2f9ea1393e2
1057 F20101112_AADFSZ kane_k_Page_76.txt
f91f2a37e9661494bf59c8a7c02477c4
3ca45df80e4f4de231351c8ff829c0f5cdf51a0a
6338 F20101112_AADGJS kane_k_Page_43thm.jpg
b2e6811af73714245536c2ce83f8e9ab
0a9f8fdbb1e43d062c6a2fcbd6596407eafb2b3d
2303 F20101112_AADGEV kane_k_Page_06.txt
0dd2a5c6967ca7c640f440df2e65f7a3
7c9fd8e5d17140d6673c705a0fc3af87a308905b
74223 F20101112_AADFXY kane_k_Page_55.jpg
c83b10260d49b087534661972bb90f62
58407420d9140de873aaa9189f6b331f2a359649
4874 F20101112_AADFQC kane_k_Page_34thm.jpg
8278cc9c8ea905847506ea52d371b4dd
6d94339b546364921471e523bf3b9f182af1f619
5743 F20101112_AADGJT kane_k_Page_25thm.jpg
3a77e5a706ce0ccfba9c51c446c4b377
19a0a405f98a59067aae9dc1e4afd0bd66f3b499
963 F20101112_AADGEW kane_k_Page_07.txt
c311252778f78bdcb26044836e9d32be
0f629bf7a984b199393d31d95b5621f4f75d7cc4
56475 F20101112_AADFXZ kane_k_Page_58.jpg
828a0e9bd3eb020b884ddad70a4c36fb
a9df5df13813d142cd094fea51cbc2d2ed21248a
727104 F20101112_AADFQD kane_k_Page_34.jp2
faae22147ef94177bd6036d032c4ca19
62e533c312a4a3bb8d104a94929da15f3701a2b0
1636 F20101112_AADFVA kane_k_Page_38.txt
84312939a69a97ff146449d6dd9f2f86
b7b985c60460c15fe4500a55e8a2062757622644
6072 F20101112_AADGJU kane_k_Page_32thm.jpg
cd3f6b5491a272ee768300b250825b2a
a9d16fe1d853e38d3a785ea03e8c7c02c614a3ee
2965 F20101112_AADGEX kane_k_Page_08.txt
271ad79480d4a56d9f783bfa0acf58b1
03d414b8590f96f571699f1c044954aba955809a
55105 F20101112_AADFQE kane_k_Page_75.pro
3914fe1c2804215d92dff69baa0b4ff0
7f1972271077b78f1f76e9f1e5741282a3b01d24
44982 F20101112_AADFVB kane_k_Page_65.pro
e2f349fe54989ca08a460030389bce54
b1a83dc7bd813fae8a24457ef813397ced399e7d
3380 F20101112_AADGJV kane_k_Page_86thm.jpg
07dd7befe25cd97eaba64903c727b245
07151bb6288298585cb78ac4198231e3d7f3ac83
115 F20101112_AADGEY kane_k_Page_10.txt
82e849d6a3c3be37daa7f7b587d73806
d91d766e37bb4655b531ab8e06f626bdd1dec4dc
66285 F20101112_AADFQF kane_k_Page_36.jpg
8f00520b316e7b575221ddce42cce8b6
43849a3b9bfd06dae19cf4e44e135681c7f65c7e
F20101112_AADGCA kane_k_Page_64.tif
c2d2921b7bd3ef7f3d0c7cc3e4e16d1a
3b5434ba0a4e642ea0367850e11b9ea0ead8fb26
6795 F20101112_AADFVC kane_k_Page_29thm.jpg
fd569536e73a3b79e09b3d8df79e8fa6
bc77013ffae3cc7f5b4af4a6e947c13df11071bd
5005 F20101112_AADGJW kane_k_Page_82thm.jpg
80f00b5696fef3c2e483e31ef8ad7591
2409ab9b4c37a2a84552c2688a063920f30b8762
2138 F20101112_AADGEZ kane_k_Page_11.txt
a8d08707e8ff0c6ecb13db93525bde0a
ad0ac9f9a2102f0147d2098f56c69f90ef1bf684
88130 F20101112_AADFQG kane_k_Page_85.jpg
c73b0d476b955d8608da0c98d5577134
6b7ceb17c4a7e93e5966f8e6606bc57249153916
F20101112_AADGCB kane_k_Page_66.tif
d4941db0dbe183ca8c8f0f6699e8a199
692bb205ed2378184e66e205966264372e29f72c
108436 F20101112_AADFVD kane_k_Page_43.jp2
218c817eab1acdf87aa31c97c8201d5b
db7fc3dc37eb9e3d99d0a443c4441a21794b05c5
19445 F20101112_AADGJX kane_k_Page_78.QC.jpg
385947a7444435ae4ee6a38cf80753f5
41dc737a1d1f708426750b2eabd06f1d78b3fccc
F20101112_AADFQH kane_k_Page_58.tif
64450795b6dc27136d046977661da653
5ccffb100c38a6ac564ddb79df616727359bb92f
F20101112_AADGCC kane_k_Page_69.tif
64bce8a765fe2c85942d434904ed8b81
d76dcf4f473cddeee7edca5b112ce714be807e82
2266 F20101112_AADFVE kane_k_Page_01thm.jpg
d49cb6f72b790193d224bb013adc8f46
9493e7c3eb055842dcd775cb5b8d25da7c967dec
23855 F20101112_AADGJY kane_k_Page_44.QC.jpg
2be63c2a7bc1b2e636f4b7ccd279f379
e7ef93c290f5f091213b98ece7d05908c836debe
2471 F20101112_AADGHA kane_k_Page_83.txt
0596d0d787decc57dfbb7e4ac67e89fc
f741d85ee9890b37484975c39484787d629380ac
58418 F20101112_AADFQI kane_k_Page_84.pro
f9946c10fc718436cb52c332f096c231
67ce31ebcc1dbb6ee11069fc2e6d87ef857fea7c
F20101112_AADGCD kane_k_Page_70.tif
d0bcc60c4f8992be6363d05cd37d493a
820f1ed40a5de2a1a2e85f40d06d76af65700c7d
26066 F20101112_AADFVF kane_k_Page_56.QC.jpg
80efd0dc53ec4ef44eca18696ad01496
a310b87c2b224c7e813d8f21ac5645ad48380520
26312 F20101112_AADGJZ kane_k_Page_57.QC.jpg
52bec5f557c0a8d67248af649bac7ff9
813fde1128122351883e8405f0429712df6edc13
2347 F20101112_AADGHB kane_k_Page_84.txt
c90f73bc26ee8fb984c9b6b49b4d9cc5
119df8a7096abf0ced43e7a8b2a019423f4e8bfa
20797 F20101112_AADFQJ kane_k_Page_24.QC.jpg
8b8b78da3a46d7955023940f98014aa6
8fa88f115959b4365a68cb6cbfbfbf5ca978b4f8
89646 F20101112_AADFVG kane_k_Page_05.pro
533625b6d8b1a91e493b9c51915caaca
6189605bb4b54d60a32fb3cb663bb907a250a3b5
2589 F20101112_AADGHC kane_k_Page_85.txt
a917deee4f5aa6582e190f39068ab588
ede01f19311786906a087571f85d69a493bd81d8
90 F20101112_AADFQK kane_k_Page_02.txt
43e5c2fd1ebfd4fe65ff482c696f2753
36ea3cf1ffbb64a0d529128e1b33f370ddb9c55b
F20101112_AADGCE kane_k_Page_71.tif
8c3b59403ada069e0a3b6413ec0712ad
f275bc4b2be6d1f71596a4c4b3f83244e870b0e3
F20101112_AADFVH kane_k_Page_85.tif
c7ebbf442bb9b37c14275573042c52fa
47e1b9f9b725b134b6b2ef328d3d62bba47eeb86
783 F20101112_AADGHD kane_k_Page_86.txt
b8c82df58e657da21da78405582e6101
e1829535d679aa8187f48eafbbe80982723ce243
2776 F20101112_AADFQL kane_k_Page_07thm.jpg
40f3352bb72ed95a986f250c474fb05e
343fe6210ba7ac085099fc778416a1ce8ac92817
F20101112_AADGCF kane_k_Page_72.tif
2349e361aec3ee415926fde6240e9d30
895bffbff1ca73932aad751f2e2eac3f47a42af0
2123 F20101112_AADFVI kane_k_Page_15.txt
1b3838a35f8200a4d0c4a4ac465ac462
3fb1e1862487f646addc21b93cdc7932b878303d
3503 F20101112_AADGHE kane_k_Page_06thm.jpg
8d5f01f61267a9949f3376a19f49fe3d
15b2bbc5849edeefc3615a0f0763d0b32ed59cef
1214 F20101112_AADFQM kane_k_Page_12.txt
046c2929c70498c92e07e02131804d62
878492a65693ffba9a6a346fee7bf58cfa8da101
F20101112_AADGCG kane_k_Page_73.tif
cbc8aac9669064c009ff2edaca57cdfa
00b8ff4e8403664077a7b5f423da580c9d008bea
73597 F20101112_AADFVJ kane_k_Page_81.jpg
9bbc78e8e50b79b4aba858614cf2c01f
7b61232a5c60c8cc9ad628b902b1d7638874f798
15759 F20101112_AADGHF kane_k_Page_40.QC.jpg
865cc207982d4f83b0db7f21b27efaa5
6e6eccbbc02561b62beb6918e9063218cfe45323
24319 F20101112_AADFQN kane_k_Page_47.QC.jpg
8c392c6780124355cf3d486c73d5c5b7
e6d07be90df17cd47c3fdb1d12955481ea484ead
F20101112_AADGCH kane_k_Page_74.tif
d56961d74f542ccd7f80d18bdf26fff5
847b6e9ccda56f139956f17a7db9cce61cf2fade
5550 F20101112_AADFVK kane_k_Page_42thm.jpg
855429eb6ba8d92880ee0ca0f048b84a
9f6d9cdade83f2e88d1970e825d1ebc8848edfd9
24474 F20101112_AADGHG kane_k_Page_55.QC.jpg
4a0e5c251fa0cbb6b100c3bf57f5c8af
8c6c46fea8e2ce7568e1ec8e23b791bffde55ae5
19977 F20101112_AADFQO kane_k_Page_67.QC.jpg
ec18d24b55b84864db948a8e7c94aa31
eb8e264300ae299da3d6aa4cd4f98251dd9eeba5
F20101112_AADGCI kane_k_Page_75.tif
3847577f2ff4d98f51a059439b7448e5
723644e2f5722b0bacf58956508dc96db356ca62
1230 F20101112_AADFVL kane_k_Page_50.txt
402a9d1c85ce5ebd823b09ebe4b26551
21fef145793c83e905686c1eeb988f27796532e4
6281 F20101112_AADGHH kane_k_Page_51thm.jpg
15bf4e3829d4119a1256c1c9c8afbf6c
c58d6603e499c67091c0421104c38fe1b3b4deab
5598 F20101112_AADFQP kane_k_Page_20thm.jpg
00d0f339cbe7585bddba1371f0c1124a
344be48410231c6d58337610dd1ffd1968267778
F20101112_AADGCJ kane_k_Page_77.tif
ddb161d8ddd03ee900cf8aeeac103c08
0801fb4d27ab29e54e2fb4146a890f5d1fa471fa
1051947 F20101112_AADFVM kane_k_Page_50.jp2
fb86f65670ff7fb91c82fb694734c5a8
a00c1f6caaa3cb7cd4240c62791ae0cb1ceec5e4
7093 F20101112_AADGHI kane_k_Page_75thm.jpg
c9e909ef7d99eee651364289b3127f70
7699f7a7afabe81d84751b77b87b48798df45a34
64550 F20101112_AADFQQ kane_k_Page_61.jpg
0c53657099192361eacb3033a579a6e3
8077e57e1f2a9b1ab3a57dd500d1245eebb7b2b8
F20101112_AADGCK kane_k_Page_78.tif
ef58043b340cc32fd5226afeeea81bcf
6b18ba79a46fd6deac5e97144efe9b2de57993ec
18826 F20101112_AADFVN kane_k_Page_74.QC.jpg
a726cf3e7ab3054177102d1d9bb4005c
ae9dee32bd77241c08c940ec8a4724b4fd83657d
79839 F20101112_AADFQR kane_k_Page_56.jpg
2d8582e754bc9b5a3ae6ff8791030819
d596904d2e7be40adf88320aa43afb4d0f9e7d9c
F20101112_AADGCL kane_k_Page_84.tif
123a53aaebed2c96cc88c42906666ff6
bed0aac34a6ab0e9ac6a871aca1bc6a8d2721d99
71745 F20101112_AADFVO kane_k_Page_63.jpg
6b0bb1453413ccb960f4e3b6ccdbb2e6
697b4549f809162b8bcecb939a6a6ecaf99d6607
5989 F20101112_AADGHJ kane_k_Page_49thm.jpg
964c8f189b7c1bf48923f4d57b38cbad
a227a65ac533d2ed3da1cf2273136548d7b0ba48
98713 F20101112_AADFQS kane_k_Page_09.jpg
0974497d06741c18e0e845dd8de67c05
77c0371fbb1234b8b67226dd9c4b595afcfac070
F20101112_AADGCM kane_k_Page_86.tif
be38e5b341a1089deb72e8daf19e66bf
c2f5da98deeb083c8a1a49c7da1c015544102361
80545 F20101112_AADFVP kane_k_Page_72.jpg
ef5b516110b55864a111f7e66e8b4df3
584d28c6d469d0bae166352e080853f1c9ad7a82
20606 F20101112_AADGHK kane_k_Page_21.QC.jpg
bcb66b44ddf97f089a64e39fe71346fa
16351246b9030a586f35f9c9c6fc9ec22838813a
1245 F20101112_AADGCN kane_k_Page_03.pro
667e8cffc74aadd5c074663c46baf419
4061e315da56fe65e21eb85188442f099091a38c
23232 F20101112_AADFVQ kane_k_Page_63.QC.jpg
7011eec95d80ae678de984767193133d
63d826e462902409a18e04954ad19a7a04dc1b16
18852 F20101112_AADGHL kane_k_Page_58.QC.jpg
a09c163362267a7a4f29c5e3a63296e4
1abc838268a2f35617e9b76a4cb582a23f436e24
5893 F20101112_AADFQT kane_k_Page_03.jp2
0936dc632de5088e4673fe106be71955
6988814cbfd9dc4dd9c04c713b71a6d528f4fff4
23148 F20101112_AADGCO kane_k_Page_07.pro
4ccecb7a36f0d3eb5115d558537e65d1
504997a4ccb46ddceaacf2b12f93899730166ba8
953641 F20101112_AADFVR kane_k_Page_48.jp2
cf4b37119d6c986ab6cd2e809ff5e196
82e2df01ee6618d6e5a2fa40d70d78fdc83bdda3
4948 F20101112_AADGHM kane_k_Page_40thm.jpg
14c07ba8e33a7e3e0d5c588050d34a55
8c68e20591de69cdecc632034deea6cf6fb79225
673 F20101112_AADFQU kane_k_Page_04.txt
efe084c285b0727359435dccbea180a8
5396bbd7bedebcb06da960f0020aff7b7d9692fa
78974 F20101112_AADGCP kane_k_Page_09.pro
3c8d9a866d57c39b09ee12f76af23c0c
176960ff78c79c2e74f6fe10e7be7a92c2760f64
811021 F20101112_AADFVS kane_k_Page_73.jp2
8ef27e8e0ecc2ff359ad3b268b16a637
e510f39fc3785fa2ded1dd2d0fbe93c250c871f3
6175 F20101112_AADGHN kane_k_Page_62thm.jpg
12f2a78b6595d3418f9cd3fc78cfb0e4
f5dd5c6ab7b1dd477ffac7d84ebc64e89101e48b
5870 F20101112_AADFQV kane_k_Page_35thm.jpg
e9161a609ef9a34666be0727085f5bb0
de80ee892612f5176e4a56fd6b8d28dd70deb3d8
2793 F20101112_AADGCQ kane_k_Page_10.pro
cca2225c70d14ac2db922899687c0e9a
ff035963dc6195bdc9360f5ea6fdd900f57637be
38267 F20101112_AADFVT kane_k_Page_49.pro
70710ae873b88ddd25b7c05b37d031a0
e6f9a943e9a0c2380c45634c2cac84599d5dde47
6632 F20101112_AADGHO kane_k_Page_08thm.jpg
c81d06b91a6bfee54d7ded240359ec14
63bdf970b1b84855d3719a62c6bb28fa96791d29
1051972 F20101112_AADFQW kane_k_Page_09.jp2
571293b992c779c4126a5094fe9e61af
617e69b54a59625f90ae8087154e37bf25a51486
49323 F20101112_AADGCR kane_k_Page_11.pro
ddfe320e870477976bcfd2ef4ab6958a
730ab2e99fe58b47fce61d41225e35f588ba10fd
54883 F20101112_AADFVU kane_k_Page_19.pro
991bfd422627618960f7f2ebac60edc0
b42f62393fd9d8cd1491de5fc9f5edb1fe29db01
25786 F20101112_AADGHP kane_k_Page_68.QC.jpg
58d8623dc205d4e2eae39be03b5c3bea
66e400a14ed3f3f8ee272135ff4de95d21baedcc
F20101112_AADFQX kane_k_Page_79.tif
f1e9092a7f4196ea5a0203462367e339
9743eceb705d343cf6fbd6c278181f12d5711868
52716 F20101112_AADGCS kane_k_Page_14.pro
67f7fa7361e1182a6e4d438dc93b0c5f
4b02c90db8d6e962c9a4c05303dc335fa519c1b6
6830 F20101112_AADFVV kane_k_Page_55thm.jpg
1ad0b3fc16fe471a5cc1fde66d28e904
c0459e5af91a608cd66a5517b4da5fdb19badec7
5940 F20101112_AADGHQ kane_k_Page_71thm.jpg
b60f58f7d7c31cc69d095005c4f011d9
f2220c12684f1b4b594e03b9c3dcf7303af36a28
F20101112_AADFQY kane_k_Page_55.tif
92302efa33a8913853e0233dffbed3a5
b42abbb1596ac4135e3306f3218c51230b62a0f5
53787 F20101112_AADGCT kane_k_Page_15.pro
7a7142784be9b8d33217b1caf4270e6a
d1d2185697fb77d50c18279e30011f0f740e63dc
87140 F20101112_AADFVW kane_k_Page_28.jp2
a7ba11f3e9b5c558de7352c343accaef
1330b44041f0c13ccea77f04676d42e91236ef72
6924 F20101112_AADGHR kane_k_Page_64thm.jpg
ea3f2922306051552dd678c36568ca0a
1d0a10ada1b1626baad4013e4c6c987999703c68
53464 F20101112_AADGCU kane_k_Page_16.pro
a81232762acd000e62fdb9f0b075c172
c02862beaf162d0d8e524b9e720e0557b6274247
6863 F20101112_AADFVX kane_k_Page_18thm.jpg
816761e90739af26e32ae3a999544222
051bb186897f9a69817b29a20307557097c08d17
53770 F20101112_AADFQZ kane_k_Page_77.jpg
2b039f1da4083a64884f722196982fb5
050f11a024a03a0352c81eb4f17c0d2aebdb3185
20006 F20101112_AADGHS kane_k_Page_69.QC.jpg
b22378722e2f5a11a7e329786aa7c3bd
4ff991f0e1d1d6ef9706738ec1b9878f7060111a
56653 F20101112_AADGCV kane_k_Page_17.pro
7257660faee47e8c56f13ddf4384a0e7
9127eb2d5dac16f6479fd95655a2b7e9bbb42315
27129 F20101112_AADGHT kane_k_Page_83.QC.jpg
fed0cf3bb64319afcafef82fc8c844ab
179df49ff55901ef32c10df5f7eedcfc95cc8ebf
54633 F20101112_AADGCW kane_k_Page_18.pro
3b7099973060993159f7646ac2a46eab
add51d957fc7309b480835a22995eaa8f575d471
77643 F20101112_AADFTA kane_k_Page_57.jpg
6cb437ad8541b12317300cdab81935ec
5d20b76af6d79cd4e0c0b6fe4ff29d9c12a0a38f
130433 F20101112_AADFVY kane_k_Page_83.jp2
c067fec752806f348424fe37d2a132e6
8f775dced4462af5363cee1a618210a93b4cbb25
21189 F20101112_AADGHU kane_k_Page_61.QC.jpg
707da639340d5e550bd6081139e87a2c
24bb028820885892d2efdba02e2b386d535bb309
43259 F20101112_AADGCX kane_k_Page_21.pro
3330dd5f3eddbb2187735b13e6b9135b
2f06b1f90b9b240262480fc872256e67582c95df
F20101112_AADFTB kane_k_Page_32.tif
ff9c2c7a2e979f3ac5bf61537ba5fdeb
4c0ea0f48a165827e22f294d0facc396f8ea08b7
1643 F20101112_AADFVZ kane_k_Page_49.txt
c943657badd73a9b311d721a41dd5271
eb9304d04fe66679953d21e9edbcde19c49c596c
15531 F20101112_AADFOE kane_k_Page_80.pro
6707bb8bef9ff6af161794a83770e510
06a70315fe604fc2b16034d523075be3fe3cb2ce
7115 F20101112_AADGHV kane_k_Page_85thm.jpg
f32bd5ad1eb413e598b67b140ac95b5d
fd7c8e45d26a9f503ed8bce906da0a2681bf52e3
114633 F20101112_AADGAA kane_k_Page_70.jp2
c5c384a4de8efb8c7effceefec85ce6f
324ea567d7ac39bd5806dd1f407718bdff2aeb05
51114 F20101112_AADGCY kane_k_Page_22.pro
f35bd33e1f9b306eca27d268daac01ad
4382361d3e8e3aa20257f6508278110150ef2c38
2189 F20101112_AADFTC kane_k_Page_55.txt
78a24733ef15397e4b8ceff856c4d921
aa2ef40237634e765d073584450a1895f58ccb1b
5262 F20101112_AADFOF kane_k_Page_77thm.jpg
9abbdd8976bcf482c54e27897be1170e
42515e23befc096751940e7a80b5ac76f0fe5432
5795 F20101112_AADGHW kane_k_Page_54thm.jpg
efa3ceadd590aa89abac5892995b0bb3
5155d20da4d903eef9579735b9b75800408757ef
35389 F20101112_AADGCZ kane_k_Page_24.pro
6484432db1ad7c0389dc4e7aa5069df2
71b565e5ae0a0a07c0ea90339d6d704b01189f07
117465 F20101112_AADFTD kane_k_Page_68.jp2
05de0b0a0b67d7dd8f81ed53bd62de79
a39588345827c24d0914637de7251a2fdd8ec925
30059 F20101112_AADFOG kane_k_Page_04.jpg
dbf0b781658e2b61a30153ca4d1de8fd
3a468a7a31981bfba0b25461c845422f85aa2c12
64873 F20101112_AADFYA kane_k_Page_59.jpg
ee302ae31cf4698cd21eba7362117491
8ea52ba76e2e242b923efc76dce4da59686154a7
900480 F20101112_AADGAB kane_k_Page_71.jp2
ad5ddf6679db9d6a561dcee52f6d343e
4e2b73703b9b9ffa94cf4ba8a1a3860fe6d3151f
4514 F20101112_AADGHX kane_k_Page_73thm.jpg
9bf167c091b2900e97cf9a86940136a1
32fde2f6957330ca2a3e1be9e227d390636f090c
12296 F20101112_AADFTE kane_k_Page_10.jpg
0d73bf70c8f6753c5a114acd6b44235b
c6bda695d5b5c82bd2082c82596c86485ea5f10a
1051979 F20101112_AADFOH kane_k_Page_06.jp2
52b347b248c0ab002491113ee48814a4
f842fd1e4456033f1b4504c691b6e57c232e0dbf
71002 F20101112_AADFYB kane_k_Page_60.jpg
3cb813935a1beadfeaf2ba23f5882263
dc9c38f74ef2f4449020b81fa76f06df470d1c7a
23042 F20101112_AADGHY kane_k_Page_60.QC.jpg
8bdf2f6c009a08075630d11d0ab40b9d
a08f8f87768063861f2325a4010837526f3690a7
2112 F20101112_AADGFA kane_k_Page_13.txt
0b933252b0d23302154f6fa91cc8f931
9de4ea9b162d5b2ec6b154d586aead61542b033d
6206 F20101112_AADFTF kane_k_Page_24thm.jpg
20d46da72fb7fefc14b477604995f1e1
ed3d34da47a7e3af43a31b3de6dc212352674878
57827 F20101112_AADFOI kane_k_Page_72.pro
76660873c4b94889407cb8877cce3e9f
734fa37d0a1b78cbb854b520bcda95646621a625
70265 F20101112_AADFYC kane_k_Page_62.jpg
05ce6fcab02b8d71bfe603db4271e7cb
d086e67e264d75a7eaed002088edd447e82fbf88
120492 F20101112_AADGAC kane_k_Page_72.jp2
b345c8705b3ac0a452c35e1d681f0082
52bab2a2a61c08f473a0d4c8cad343eba252e3a7
23785 F20101112_AADGHZ kane_k_Page_23.QC.jpg
a5e53012b767429845befe6758e07900
ea3a6dc0050f8bee7724208815fbcfd026cc24d2
2073 F20101112_AADGFB kane_k_Page_14.txt
2c9ff0e19aca0e7567b3870ebd195423
4dd1d66d76c7b8dd7354ea9485d42bb870cc4b73
F20101112_AADFTG kane_k_Page_67.tif
786453bea1316755ff7d7aca9ef578c2
260de880796061cef43cc8ded40f2ccfb1a7fc7f
3142 F20101112_AADFOJ kane_k_Page_09.txt
1ef5b37b4ce581e195b2874a69b9381f
3cf8544596f2ad38f1496cb39df0f827e46a095d
71878 F20101112_AADFYD kane_k_Page_66.jpg
d77babe83dd2ba335681029f27e3ce19
622c89ada5f1b641c37a73d0cd6addd04374a464
115808 F20101112_AADGAD kane_k_Page_75.jp2
6abd8650574f61ed7fcce3d1c074cfa2
7060a51f4b9bcbea218663e602d84642f386a969
2106 F20101112_AADGFC kane_k_Page_16.txt
f300f163e48116af549032507830c903
c81bf42a1cf17b2a438f054f574f9bb7936c0ea6
49959 F20101112_AADFTH kane_k_Page_13.pro
b8b149d4aa25e6fbb2334a0c6329e980
a8167e95522a5ae91f01d70f3b15f454d3c0166a
30494 F20101112_AADFOK kane_k_Page_12.pro
cf504fe7b68844416f5179f59f386c6d
838df63348d29cc7fdfcad335410d561c061fb36
60890 F20101112_AADFYE kane_k_Page_67.jpg
b0dea3855236f849e6ea026d1cc0f3af
22488549f6c6b6d23978db28f2511b9ed5af164a
1051985 F20101112_AADGAE kane_k_Page_76.jp2
1ce294a9b75037e081b443339e6a78b5
08113182983b84223b4b125a7cbfb9177b80f710
22358 F20101112_AADGKA kane_k_Page_27.QC.jpg
3970fc18bd3035f356e7c116c342d2cb
ecd3092de09b758426a76c8aaab8094ae96c15f1
2151 F20101112_AADGFD kane_k_Page_18.txt
35c7f994b30086c2e8a0bc8bc0b96eb0
eadb7903bf65d1f67516b31507ef0888f54a4c9e
2978598 F20101112_AADFTI kane_k.pdf
9369342a8b2bb1025ea2e4de0ef8c484
de232d143e3f83394a776900a595c6d4b010b97c
52976 F20101112_AADFOL kane_k_Page_39.jpg
8660d94694b8e143797d5bb815cde469
ecd8db1d767820f9ac4a0169518c65e459c88739
75296 F20101112_AADFYF kane_k_Page_70.jpg
ba65cc2e78cab16cd2a54e490d98db48
7dc28262601e676406378c755c8f9d44174262c9
762901 F20101112_AADGAF kane_k_Page_77.jp2
7fd5ad283e6c84a02f6abe0b9e385db2
ee66ceac6ca1ee291d25a7f4861fde14c5cb5d78
2945 F20101112_AADGKB kane_k_Page_04thm.jpg
dcdce21cb879540c75ed82681fbd7f66
6cf26a0f2b2bfd8acd6789530c0e8db0604ed3a1
2148 F20101112_AADGFE kane_k_Page_19.txt
8272881e47a8b231a99bba0a57b457ee
0c0fd9fb5bf5e1ff2894dbc5fec027e5bdd39c0a
59284 F20101112_AADFTJ kane_k_Page_78.jpg
b75a271e921fafaf3ee36b7fadc5b336
69be44a2c5b11e2c98ee959795a00c50a9566172
5430 F20101112_AADFOM kane_k_Page_31thm.jpg
f95003d27b03a2413d0052fb161f45e5
2c83598d98d973e82c341aad9fbda4a951074429
71327 F20101112_AADFYG kane_k_Page_71.jpg
8650a6e51d80885b8fe01f4432ea48b8
76b828e07c638882a8b512de569d731e6ca88ac7
853638 F20101112_AADGAG kane_k_Page_78.jp2
9093fb820e7bd2dfef7213218828f3e8
778c9499ff069631163f35436cfee30deea2f220
6388 F20101112_AADGKC kane_k_Page_11thm.jpg
b346c640d9f4f26b0986010a36593628
be78718b57235290939f8a5a78b94cd2eba267b1
1664 F20101112_AADGFF kane_k_Page_20.txt
f450ba4f2f104b71d90233fff082d758
67f88e1d74c3849f04847fba510c979ae9b359af
23528 F20101112_AADFTK kane_k_Page_66.QC.jpg
2753a0c436700015c406f6ca64911123
68450c6c3f1789ae8c8aad07bcd43a882cbb5119
31880 F20101112_AADFON kane_k_Page_67.pro
22cd802bc231821506687cf6e2614f2f
1ef4a6d1b282addcc8b9a643f188b84c06f6c809
58408 F20101112_AADFYH kane_k_Page_74.jpg
c8cd98d92cdab74cc28ae693655e91b2
cd7367945cb7427794dc9250fe5b675dfd6feb8e
111449 F20101112_AADGAH kane_k_Page_79.jp2
3d851095a1395ef33bc8d9ff77e64644
40e86ee49f7ed7c085b62dced6db485fe7d4617d
6682 F20101112_AADGKD kane_k_Page_47thm.jpg
72b017da9fdc8e3dd2990bfe81caaad3
a4786b94f0dfb4b4f0f339f977960039a93b4457
1852 F20101112_AADGFG kane_k_Page_21.txt
ff91ad40858413f23a0f646b293a0e64
4246e0db387c7e0838d703e4c286fe2b9a8ed1f1
105927 F20101112_AADFTL kane_k_Page_13.jp2
bb89ac22bb16efc5272e2569b70533d3
2c35d58d05775298c1aa7aa52be308ce725e1156
25813 F20101112_AADFOO kane_k_Page_19.QC.jpg
a863e2b0e5d06a80c167994a4013818f
8316f977a11c3f20bf95267c7f0c233f8b8f1b5a
77435 F20101112_AADFYI kane_k_Page_75.jpg
4fa164e8194af15cd584690b22b83385
50b4c0c11ef3f8993073708dfa9727677291a6c9
36567 F20101112_AADGAI kane_k_Page_80.jp2
cf58b83c7c92e2a71edb9c216885aad3
0ffabdcf4220834394f939f0a48ec54b0dba2f5f
24888 F20101112_AADGKE kane_k_Page_16.QC.jpg
2cf9357d466eafc055a1261c67af4a1e
3596f2b11cdb23cc3eff3a94d7f3bbe61adebfdb
F20101112_AADFTM kane_k_Page_10.tif
8f35c5b8e0053859a62278f34eee8509
278cacae7c9e25fcc59ebc4c3bd9f77c80760b3a
8423998 F20101112_AADFOP kane_k_Page_25.tif
51baf192d252be9306dc70d480633ff5
50f790630df2d19524938ce9c08bab7969c1b30e
27241 F20101112_AADFYJ kane_k_Page_80.jpg
ae0a31db109ec9510d05def0024bb184
5e396d7137b5b38f23ff497b89059bf749465fca
110609 F20101112_AADGAJ kane_k_Page_81.jp2
9799e96aa0c2928913c99c53fb82ea37
842ddae622b702a9b85736655f2dd479d95ed091
6601 F20101112_AADGKF kane_k_Page_26thm.jpg
d318a9dd839ed09d6375e55d7362667d
43f65bb094d49eddeca25325c37d1f642d553aa3
2010 F20101112_AADGFH kane_k_Page_23.txt
04b0edc7cbacd1dabb2651e511c61729
5032772294c2041cd880cbf4bfe029e29b616c77
F20101112_AADFTN kane_k_Page_15.tif
74658cb820939175aace4f13b534c2c9
6757846d117520e57cfcb90b909002683e8266a1
F20101112_AADFOQ kane_k_Page_56.tif
214ccbc89c88a1be8937b6fb3d40f7b0
896d2d696ad383c39991e85aa7734afaded4ad41
55032 F20101112_AADFYK kane_k_Page_82.jpg
ed498e6e9034b9c18e31a01b6b110957
e0826eddf522c54be2d7d952f8dd3df2669a1fed
79460 F20101112_AADGAK kane_k_Page_82.jp2
843cda80e7d8a9bb5f49ad2815fb9f3e
d3019ad9f33a38bbdd830783b218c74d59daaf5c
6413 F20101112_AADGKG kane_k_Page_44thm.jpg
37e8b2f9a689269fc7df8f7f1c0e384a
6e525ee36482e2758090b69b1371cb394bca5560
1613 F20101112_AADGFI kane_k_Page_24.txt
d9478234725028bc488f4a8a234d5aaa
0cbf230b566df772143b7eaee0b583c188761a31
6966 F20101112_AADFTO kane_k_Page_68thm.jpg
b6c0132aa190f826de9b2aebb1ba738c
1f382d70c6ee6f9a0cad88a263d56de6d181cbed
92043 F20101112_AADFYL kane_k_Page_83.jpg
fe5c5c3a9692420171a2ebf5e95a161d
f635e4376196ee188dd26f2de8bc9b9bc0c2ef2c
45264 F20101112_AADGAL kane_k_Page_86.jp2
ba7b42ce2c6f5632fc2ac7d75beaf884
d51a5fe52edfadaab6a7e6feaf99213caf9ae7f8
7052 F20101112_AADGKH kane_k_Page_57thm.jpg
c77007e839ebd967eedd0d4988080701
a66c23408e5b5de89617d02b8b183ba372c34aa7
1702 F20101112_AADGFJ kane_k_Page_25.txt
4c1a04f41a1ba70e26d4350c3bbce9d2
69bc38d0b49a6663bf94e7b67dbb0b52b1edf406
F20101112_AADFTP kane_k_Page_13.tif
fa5cee80c50f6b55eeaf06980f2a059b
b2180b9d8e4b38c5d7f1e19875187dfa94eda8ad
80440 F20101112_AADFOR kane_k_Page_17.jpg
998eae6f6361c2068c6bb592b8883193
ed8699ec5a9462fabc85205cb1a05aa92a85ccc2
26563 F20101112_AADFYM kane_k_Page_01.jp2
2993887561eb268ee3423959aa7f9980
cd84e03a6222ba56cf2ae954e42fad1ac9e3c4a2
F20101112_AADGAM kane_k_Page_02.tif
60e9071cca34208aaa69cb05a1729b2d
53fe8664bd6aaf65bbe29001c71e6bc7dad24dfd
128624 F20101112_AADGKI UFE0020581_00001.xml
bfa12124b9bb15b43f9a71054b5b1d84
e7bb628dbf71fa28271598dba947157fd034cba8
1818 F20101112_AADGFK kane_k_Page_28.txt
9e0bc446b374d29c6e38ec255b3374d4
13427c6c3f8854ea10a480e744354f71208d7a5a
73654 F20101112_AADFTQ kane_k_Page_08.pro
09e176b7cc0e4ee9262e905c3f6558b1
4c05d48cdd2363ce8a186050abaeb9413ec3cef1
F20101112_AADFOS kane_k_Page_21.tif
70153bd6e3b27b1609eb0ef191ef63ee
7c03cfd859a0060a8c9006aaa28b2dab7ffc7370
4885 F20101112_AADFYN kane_k_Page_02.jp2
8f05a2ac702c1250bfa0c23d08349d32
356a4ba816a563fde1455bdf5a5e62c7b3c9e8e4
F20101112_AADGAN kane_k_Page_04.tif
2a34a93faf67f3fee64fccdfa9eabc0e
b54b1908b017e8f56f6c12a5c33c3f4ed249c7de
9611 F20101112_AADGKJ kane_k_Page_04.QC.jpg
ca601a16066e199feafaac32a1fed747
e763d9d3aa248ad9954ae7957db41dd1539d402a
2077 F20101112_AADGFL kane_k_Page_29.txt
194e2943a0b7a8373900af3c22ae24b1
16018c92c6fe1cec22fc8b64e08e96d39e321e0f
18150 F20101112_AADFTR kane_k_Page_40.pro
f2f17827dd7f61048a303e838bce37c7
2d7942c7d74eef31477d1eab02142899d45faea9
29778 F20101112_AADFOT kane_k_Page_50.pro
6918af4e5e9d4df61701a21779e0a354
7b1bddf38048808e910139997fe207bff886545e
38765 F20101112_AADFYO kane_k_Page_04.jp2
b4f38fab98776d982ab34f7c9b7bee61
d2500b0794513ac2fc3e7ace6e0b962e3698250a
F20101112_AADGAO kane_k_Page_07.tif
87ecd7c3d509628fc05d39a0d7b8ad57
cf1ff1c974a8805929bc4d8582f4cf6d33318495
13356 F20101112_AADGKK kane_k_Page_06.QC.jpg
04b0798e4eaf06c353f5f1c8e74bd3d6
95856575cd04dc45c76177e8af20c497c7652fb2
2168 F20101112_AADGFM kane_k_Page_30.txt
3818a584893dfc45ba908706405b732b
32567b7035846ee3e62490d5cf53a47b1419b324
78039 F20101112_AADFTS kane_k_Page_15.jpg
8b41775356ec53a5edcc1b6b38b62175
f653c57fe9f4f2fa04860023e7ae77dd9483401d
31780 F20101112_AADFOU kane_k_Page_52.pro
9ff2d010259d94ba4e6d88501be2c3cf
852cb32dcf07267ea86095282de2470a46915871
1051958 F20101112_AADFYP kane_k_Page_08.jp2
964cdccbea977abf95ed4065a776d4a3
2ba6d09f6db31746a2600ceb797a3e44d6677a6c
F20101112_AADGAP kane_k_Page_09.tif
cb35cd1a786a803508104fe157d7f470
1eca1039ff8da8af50a704fd869de1d263d4aab2
27533 F20101112_AADGKL kane_k_Page_09.QC.jpg
76f9ac51d232132898733a3c99873cd7
1cfad7ed8baa4241e67fc669807963b1a8f5e1ec
1528 F20101112_AADGFN kane_k_Page_31.txt
3666486ba70d1ccc612e4a338156654e
32aaa3f8a5415f29ddfb23f9bcf5eed8f8eeb3c8
F20101112_AADFTT kane_k_Page_65.tif
aa31aeade80690506683938882a3746c
b8586dd9a16f0cd35bb1d23b3fbc29fdd4a3de65
67091 F20101112_AADFOV kane_k_Page_25.jpg
9f0a7d636d9ce6b2309e4b9bc3211e3b
d8c7921e7af7c4c9e72b708abd5bc00ad5a4411c
106151 F20101112_AADFYQ kane_k_Page_11.jp2
716de3295749ad00bcc68b7de3531245
7ff3ae29088a014efde1e721257b9297fef9ccbb
F20101112_AADGAQ kane_k_Page_11.tif
a218f39db8354b8d30f7b95b68368520
e1e1a6c9416bf3f4575db15fa6f94459dd174fb7
1331 F20101112_AADGFO kane_k_Page_32.txt
ad6ed3b09e6e2d8bb0b1cdda0b9f2752
28e7e6cf565c64e82898f728f5e9ca0b6591c573
F20101112_AADFTU kane_k_Page_31.tif
5c8c66249c1833ae8f8d22b48e3ea3ee
a8a7a5f914d9d1f61276154cdcd369b4f993cd2b
25557 F20101112_AADFOW kane_k_Page_15.QC.jpg
63edebbcdcc82445342555a81dd7e443
c66876cc3eb168d8d72244d32a34916d970194cc
68233 F20101112_AADFYR kane_k_Page_12.jp2
f38a2494ac8702b90c15c9ae3947741c
13cbd66216158f7f6ab354b86816da50262b9b50
F20101112_AADGAR kane_k_Page_14.tif
ec1ab7b581025e9d51e37aafa656afd3
21b6301c679fdf56576db8533bb9d620d4ac9f13
1571 F20101112_AADGKM kane_k_Page_10thm.jpg
6eb309a75e18f44766526087fc4915eb
a771517ce80970a8f2db3caf4e8b8f3f01e7f89d
F20101112_AADGFP kane_k_Page_33.txt
56f3986c5ee25dacab94c6b0c62c4a7b
009328da5bf53c71e1262c8ffb3898ae5274b0be
80711 F20101112_AADFTV kane_k_Page_84.jpg
face978df581dacdf5251fcb9b844286
a2b2b09ae8e08be7ba4fc2d73cb36c672731765f
83458 F20101112_AADFOX kane_k_Page_05.jpg
2a09ed034385231fd7815fa9bdfce4d2
4384a904fc9b2a2dc1d626f27404fa53a33d01c1
113078 F20101112_AADFYS kane_k_Page_14.jp2
6344012ebd7017aee89b354211695002
b47ff343989c38e2d0e54bf10bff6bba7c83f92f
F20101112_AADGAS kane_k_Page_16.tif
096c1378314bc03003bb5d00e5d00a0a
69685ae04647a2ec55cd3e882a798538192cb412
22829 F20101112_AADGKN kane_k_Page_11.QC.jpg
8cdb1cc285e0fb9b142a84f1dda70015
35357ab2d1a82b2cb78fcd42b22591b4ded68b3e
1773 F20101112_AADGFQ kane_k_Page_35.txt
06361eb8baf8a9b0cbce1021fa30f1ff
d8fb26ae88bc7e8bf65f1cb7675b191bb0b20b68
7503 F20101112_AADFOY kane_k_Page_01.QC.jpg
ff32b0b4ab11230dc6c65338aac0a689
b88e24fd21ff75e637ffaebde306982df20650ee
118219 F20101112_AADFYT kane_k_Page_15.jp2
e919ef151b29f8e40d4b7e417129797d
114e39a3285a5cf3ec6d5befa1283682ba423e8f
F20101112_AADGAT kane_k_Page_19.tif
8577a5e539d28199257c47a5ef855a7b
993a11b249b66e52b50d22ef2e1343ef4bfa3c50
23342 F20101112_AADGKO kane_k_Page_13.QC.jpg
187188290e882fa093fc33490ff5d393
b307112ddb93f510f9379db0293d4c343fa36a67
1569 F20101112_AADGFR kane_k_Page_36.txt
37769d1bd8c2570c505bf9502803f357
31e508afaabe3c2999144d8307f5244bc62554ef
2012 F20101112_AADFOZ kane_k_Page_27.txt
991061f3da5c22ddfd7c43734b58b23c
1c5a5be901105b6b0d08d121d5a62f61693f5b1a
116223 F20101112_AADFYU kane_k_Page_18.jp2
4768573612716a41906f0e9b4de3e7d1
faa1e601ef4e9c2293f2212038a52819ecf98c79
F20101112_AADGAU kane_k_Page_20.tif
2856de95530c61b5669bb2cbe056cf67
be65a78642c6eaf9d45c8b925361b0c2148faef7
6772 F20101112_AADFTW kane_k_Page_70thm.jpg
a8fbc560320daa1fcaf7ba954ec42668
27239e95e6bde2ac8b558852326aa650d88af103
24847 F20101112_AADGKP kane_k_Page_14.QC.jpg
b56534b448438eaa2231455d0a50db41
cf4b3b7739b47d6c59c060da82776b127cf87991
1810 F20101112_AADGFS kane_k_Page_39.txt
d18b681deb45c0a0decb6b2ac9af371e
7e55aed66f14a370553e073630522e0de7b2b5a9
118755 F20101112_AADFYV kane_k_Page_19.jp2
1e354b5e25166f073424fa4f543916fa
6e96b050d7b49d4fc8e1c6e7050c2b64cc8d303e
F20101112_AADGAV kane_k_Page_22.tif
e0244042f8ffa7bb82ae34047b86a2b0
befe9a0b3a0e2d4df5a9f8c3ea942bb427a8362f
6140 F20101112_AADFTX kane_k_Page_67thm.jpg
3a0fa2581ff19004e8d398dd1f15d8ac
416f99467b50d602634d999e5577bf452080d5f3
6906 F20101112_AADGKQ kane_k_Page_15thm.jpg
fbdcc16c6f0d4af1b8733576cf71dd89
719b5c4e3a1f433fe225bd49ac874f14f3bf723e
917 F20101112_AADGFT kane_k_Page_40.txt
2f17be7c35e2e40191e5a484349d851a
beada1ac441312c14657e651a39d10ebba48828c
91121 F20101112_AADFYW kane_k_Page_20.jp2
1b4bbf875dfbcdadb5b14c7c32450579
6fb5cb3c5d8c8e93f861bf6356019d37536215fb
F20101112_AADGAW kane_k_Page_23.tif
5aa6702472a37514beac4b307325fa9d
fe3c26f879a10f768f55088ba0d8f99eae95737f
3041 F20101112_AADFRA kane_k_Page_02.QC.jpg
66872541b82f8ecbbdbdc09f26cb2416
3956de4167157366fad0fa0230bbac4895a3f4c1
F20101112_AADFTY kane_k_Page_33thm.jpg
9bc7d6899870b7061e8fca4115d9427d
74b5fb360f7d121578f449a5679c82e96cb182a0
7212 F20101112_AADGKR kane_k_Page_17thm.jpg
e2390706f4acc9a22117860613ad257b
1809408705688d45ae78f18d4160ac506e68d5e6
1508 F20101112_AADGFU kane_k_Page_41.txt
cab29ad95863bcc9d65a2f4ff4dc4186
782d7da35560ebf94ab54714197a606feb6f96c1
95256 F20101112_AADFYX kane_k_Page_21.jp2
e9bf4a781376b8f11790c0b794511e06
cac49fede756795445c36a161225a56f9374fbe1
F20101112_AADGAX kane_k_Page_26.tif
d270d1aab5881c040fa4c617b65c1e40
877e3f3c614328d7d92cae5d8d9257cfdf45a4de
2216 F20101112_AADFRB kane_k_Page_17.txt
a54b1fe42ddf3ecff05563ccf7c1dbbb
36c78a2d9d6dd642ade7c23185752cb26ca8a911
24822 F20101112_AADFTZ kane_k_Page_29.QC.jpg
96b2f336e82e4c0fada9055a06af96c3
1262239813061a2d18fb0eeeccb2719c486b7fae
5714 F20101112_AADGKS kane_k_Page_21thm.jpg
2180e59289c584739745d5fd25c14314
495403e751f0a50cef8481382aa2db044f6be08e
1555 F20101112_AADGFV kane_k_Page_42.txt
5cce2a2c39589e0062af4d7b26181d71
18a3862845943178a3f75fb689d763e7e6b0524b
109826 F20101112_AADFYY kane_k_Page_22.jp2
52381685986db0f89b730cf0d0b5c3df
a936c84066f747cd9da300e081657ab16f0985c9
F20101112_AADGAY kane_k_Page_27.tif
0536707f38a74d249d502c092d7c9822
35b1698cf5029d9584c93245cc773bc03596fd1b
912015 F20101112_AADFRC kane_k_Page_37.jp2
907ba8dc1bfce5f74c78db247b919e4b
6110b107087532c9b99c10a923fdf7b510bc9f8f
21416 F20101112_AADGKT kane_k_Page_25.QC.jpg
bfb8fcf46c1305ccbba363d781186f8d
9ce5104c8a04771bb34cf22dcd721b865fe702ae
2013 F20101112_AADGFW kane_k_Page_43.txt
91e00e63d441c29621e85e0a11972096
43fbdde6eeceae33c682969b170bd544f54747cc
61832 F20101112_AADFWA kane_k_Page_32.jpg
79d9f0d147126d26a4d407e82765dad3
3972e3f01869da3886c39c26149b17a75666a260
109158 F20101112_AADFYZ kane_k_Page_23.jp2
aa20ae3e3e64afa5d5d3ea81772bcb08
431e315c42df61fedaa490ff6f4e4596707b2318
F20101112_AADGAZ kane_k_Page_28.tif
21a29439dc2de4df3e78c4292a55c6ed
77ad6150d99c52ec455ec67462c37e7f68414ecd
F20101112_AADFRD kane_k_Page_68.tif
6aee48c2c81b831f3b973565d0e7cc1d
124a8008a1d2d03819707cd1f47f8fc02637c364
21899 F20101112_AADGKU kane_k_Page_26.QC.jpg
4452dbe19c957735ee05dd48c790e0f0
544bae7631e14c4c88ccca85c18c954e65262dc3
2040 F20101112_AADGFX kane_k_Page_44.txt
05d7408532c705dc20282aedd1119da5
ae1b4454127494950bddf21c62b7916a2ea7b837
6187 F20101112_AADFWB kane_k_Page_38thm.jpg
2fbaf0311193e68cfd51bea6e17b3f94
c6f26b9f8edf1e0c822cbd124a6bc777a4d165d9
61137 F20101112_AADFRE kane_k_Page_69.jpg
e39389d36dc391ca59f33e2dd4a6c0df
f0420ce283c59680ef5e71380f9d41c5e91d69b2
6997 F20101112_AADGKV kane_k_Page_30thm.jpg
87524b408bb24e776a0f511347eb1ef8
f4d8bb5e77a7f7ee515c248ff0a046dbae6a1445
2031 F20101112_AADGFY kane_k_Page_47.txt
0dc8da7b89187ce070a925ab8ff52af6
7a0e332f17cff8b328c5f0783c636218ebb9c16a
114580 F20101112_AADFWC kane_k_Page_29.jp2
9dbe5ae20977c102613df9499c032d56
14d10e82df0607c363392105ba9b8c18ca57885f
F20101112_AADFRF kane_k_Page_17.tif
348f33fbc0b2599c60b8009e2bc0f1db
cd50bfa6c10b37c4b93dfc0d6194063a5d12b125
40266 F20101112_AADGDA kane_k_Page_25.pro
9c93112bb740de0fc0bbaa9fb5872eae
2934926ade82c4353907d843fab498aab12ce72a
19490 F20101112_AADGKW kane_k_Page_32.QC.jpg
f6a773babee05087b1b0854ccd96380d
034966070e786a377775dae7ce79399cd7a2633f
1410 F20101112_AADGFZ kane_k_Page_48.txt
f5d2fec6cb3a2fb8dccde140373354f0
30af0cf905de65485076bbcae6f79bcd970c2067
48567 F20101112_AADFWD kane_k_Page_44.pro
fb62b29ecfc3e1c2d1e86d4b99d021f0
81a1630175dc33fe6907b39757efdb89491a523a
713567 F20101112_AADFRG kane_k_Page_07.jp2
3e243840974c87c6782ccdd074350999
db2c89ad7b6e2a0ecce0c75a9a495da4f2442fb9
41689 F20101112_AADGDB kane_k_Page_27.pro
35fb12298760dbafe8eac8814e8bb5e9
f361c3e7c19aa0edae3c49d8d67401ad01a4983d
15547 F20101112_AADGKX kane_k_Page_34.QC.jpg
2f96bffa3d7176f038a6d02e38a854d3
4e201a1f16b9e14860edc0426e35b45ef7df08e0
26888 F20101112_AADFWE kane_k_Page_65.QC.jpg
c7c972d1fccb5bb083016a2056526195
f4ee2de1e26e0841ccee3f794bef5615332dc396
6570 F20101112_AADFRH kane_k_Page_23thm.jpg
ea57b7f74cafbae487b56b95809a8fb3
fecf84debb4a010d341e147a2f390f23716525f7
40420 F20101112_AADGDC kane_k_Page_28.pro
ee7573e9787a3bf4190e11934710fe8b
dce79a3c16b37ab27e649b5579df3aeecd560550
21141 F20101112_AADGKY kane_k_Page_36.QC.jpg
1e12bf012e8a9dfbd1203919ae533f57
54a604825f7c4edae09dca742810a1be4fbf6b93
2812 F20101112_AADGIA kane_k_Page_80thm.jpg
be4750230cd5caeda2cfbd18a831c212
c884ae6c3f37c85be1b6d17f39ece9c679d83157
7678 F20101112_AADFWF kane_k_Page_65thm.jpg
cef35b58816d9f50375b941978ef57fb
d21e094110503ea99eda4655e1c4a253ddd50582
54042 F20101112_AADFRI kane_k_Page_06.pro
e40652bf0ba39bea5ebff6058e5ceed3
5f5f0660cca151edfd41c40c7dec29b55d602cdf
52930 F20101112_AADGDD kane_k_Page_29.pro
d6835ac2c804a0f4850dd12eae0f7c6e
d2101ba6e7ad06e85254913f4add2e1cf8a64878







IDENTIFICATION AND CLASSIFICATION OF GREEN CITRUS BY SPECTRAL
CHARACTERISTICS FOR PRECISION AGRICULTURE




















By

KEVIN EDWARD KANE


A THESIS PRESENTED TO THE GRADUATE SCHOOL
OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT
OF THE REQUIREMENTS FOR THE DEGREE OF
MASTER OF ENGINEERING

UNIVERSITY OF FLORIDA

2007



































2007 Kevin E Kane

































To my loving wife for all her support









ACKNOWLEDGMENTS

First I would like to thank my major advisor, Dr. Won Suk "Daniel" Lee, for his support

and encouragement, both in my research and in my life. I would also like to give thanks to my

other supervisory committee members, Dr. Thomas Burks and Dr. Arnold Schumann, for their

advice and suggestions during my research. My family and friends were a great help and

sustained through the good and bad times of this educational journey. Above all however, I

would like to thank my wonderful and caring wife. I would surely not have pursued and finished

graduate school without her by my side.









TABLE OF CONTENTS

page

A CK N O W LED G M EN T S ................................................................. ........... ............. .....

LIST OF TABLES .......................... .........................................7

LIST OF FIGURES .................................. .. ..... ..... ................. .8

A B S T R A C T ................................ ............................................................ 1 1

CHAPTER

1 INTRODUCTION ............... .......................................................... 13

1.1 Florida Citrus .................. ................... ............................ ......... 13
1.2 Precision A agriculture in Florida Citrus..................................... ........................... 13
1.2.1 R research history and trends........................................................ ............... 14
1.2.2 Y ield m monitoring and m apping......................................... ......................... 15

2 O B JE C T IV E S ................................................................................................................... 2 1

2.1 Identification of Critical W avelengths....................................... .......................... 21
2.2 Green Fruit/Leaf Separation Using an NIR Camera............................. ...............21

3 CITRUS CLASSIFICATION BASED ON DIFFUSE REFLECTANCE .............................22

3 .1 Introdu action ...................................... ................................................... 22
3.2 Literature Review ............................... .. ...... ... ................... 22
3 .3 M materials and M methods .......................................................................... ....................23
3.3.1 Green Citrus and Green Leaves....................................... .......................... 23
3 .3 .2 D iffu se R eflectance ....................................................................... ..................2 5
3.3.3 D iscrim inability ....................... ..... .......... ...................... .............. .............. 26
3.3.4 Fisher Linear Discriminant Analysis (FLDA).....................................................28
3 .3 .5 T raining and V alidation .............................................................. .....................29
3 .4 R esu lts................. ..... ...... .......... ............ ........................ ..... ..... .. ............... 3 0
3.4.1 Reflectance Characteristics of Green Citrus Fruit vs. Green Citrus Leaves .........30
3.4.2 Reflectance Characteristics of Citrus Varieties................. ............................32
3.4.3 Spectral Growth Patterns of Maturing Citrus Fruit..............................................35
3.4.4 Training and V alidation R results ...........................................................................36
3.4.5 D iscrim inability U sing A ll Sam ples........................................ ............... ... 37
3.4.6 Fisher Linear Discriminant Analysis (FLDA) Results............... ...................41
3 .5 C o n clu sio n s............................. ................................................................ ............... 4 3

4 CITRUS INDENTIFICATION BY NIR CAMERA SYSTEM...........................................44









4 .1 In tro d u ctio n ................................................................................................................ 4 4
4.2 Cameras Used in Citrus Yield Mapping .............................................44
4.3 M materials and M methods ............................................................................................ 45
4.3.1 NIR Camera, Optical Equipment, and Hardware...............................................45
4.3.2 Experimental Location and Environment................................... ...............49
4.3.3 Im age Processing................... .. ........ ................................ .... ...... ........ .. 51
4.3.3.1 Im age processing's m ajor stages........ ..... ............. ....... ....... ........... 52
4.3.3.2 Im age Indexing stage ............................................................................ 53
4.3.3.3 M arker C checking stage........................................... .......... ............... 57
4 .3.3.4 R esult analy sis........... ......................................................... .... .... .... .. 62
4.4 R results and D discussion .......................... .............. ................. .... ....... 64
4.4.1 Im age A acquisition Problem s ........................................ ........................... 64
4.4.2 Quantitative Identification.................... ......... ............................ 68
4.5 C conclusions ............................................. 75

5 CONCLUSIONS AND FUTURE WORK.... ...................... .. .......... ..81

5.1 Conclusions of Research Objectives ........ ...... ............ .......... .......... ............... 81
5.2 Future W ork .......................................................82

R E F E R E N C E S ..........................................................................83

B IO G R A PH IC A L SK E T C H .............................................................................. .....................86






























6









LIST OF TABLES


Table page

3-1 List of dates that variety sample sets were harvested. ........................................... ........... 24

3-2 Discriminability results of the training data set, Methods I and II. ............. ............... ..37

3-3 Discrim inability results using all the data............................................................ ........ 39

3-4 Projection vector equations found by Fisher Linear Discriminant analysis ...................42

4-1 O optical band pass filter details ................................................. .............................. 46

4-2 Validation im age citrus fruit pixel results ..................................... ......... ..... .......... 69

4-3 Validation image fruit marker and manually masked marker results.............................74









LIST OF FIGURES


Figure page

3-1 Samples from August 9, 2005 of A) green citrus fruit and B) green citrus leaf................24

3-2 Growth chart of the average fruit varieties' size, as measured by fruit diameter .............25

3-3 Green citrus fruit positioned for diffuse reflectance testing with the Cary 500
spectrophotometer with integrating sphere............................ .......................... 26

3-4 Histograms of example fruit and leaf reflectance. A) Weak discriminability at 1650
nm B) Strong discriminability at 881 nm ............................................ ............... 27

3-5 Green citrus and leaf average spectral reflectance through the fall 2005 growing
se a so n ................... ................... ........................................................ .. 3 1

3-6 Spectral reflectance of layered citrus leaves and single "Orlando" Tangelo...................32

3-7 How multiple leaves/surfaces create added diffuse reflectance. .....................................32

3-8 Citrus varieties average spectral reflectance .................................................. .......... 34

3-9 All citrus and citrus varieties standard deviation values for all wavelengths ..................34

3-10 Citrus varieties average spectral absorbance. ....................................... ............... 35

3-11 Reflectance changes over the growing season of Hamlin. ..............................................36

3-12 Fruit & Leaf (Method I) with Fisher projection line (solid line) and classification line
(d a sh e d lin e) ...................................... ................................................... 3 7

3-13 Fruit & Leaf (Method II) with Fisher projection line (solid line) and classification
line (dashed line)..................................... ................................ ........... 38

3-14 Discriminability results of all the sample citrus vs. leaves............................................. 39

3-15 Tangelo & Hamlin with Fisher projection line (solid line) and classification line
(d a sh e d lin e) ...................................... ...................................................... 4 0

3-16 Tangelo & Valencia with Fisher projection line (solid line) and classification line
(d a sh e d lin e) ...................................... ...................................................... 4 0

3-17 Hamlin & Valencia with Fisher projection line (solid line) and classification line
(d a sh e d lin e) ...................................... ...................................................... 4 1

3-18 Scatter plot showing all classes: Valencia, Hamlin, Tangelo, and Leaf..........................42









4-1 Spectral sensitivity range of InGasAs technology (Graph courtesy of FLIR Systems,
Inc.). ........ ....... .......................................................46

4-2 Three band pass (1150 nm, 1064 nm, and 1572 nm) ...................................................46

4-3 Solar spectral irradiance reference from the ASTM G173-03 of "direct circumsolar".
This data was gathered by the American Society for Testing and Materials (ASTM)
and government research laboratories. ........................................ .......................... 48

4-4 Percent transmittance for each of the band pass filters.....................................................49

4-5 In field experimental setup with gasoline generator, personal computer, monitor, NIR
camera on tripod, and Teflon disk (from foreground to background).............................50

4-6 Representation of an im age spectral block. ............................................ .....................52

4-7 Histogram stretching function; raw image and histogram............................. ..............54

4-8 Image Indexing stage of validation image number eight....................... ............... 58

4-9 Image Indexing stage of training image number 11 .......................... ................... 59

4-10 Marker Checking stage of validation image number nine. Input binary image of
possible fruit m arkers ........... ..... .. ......... ... ....... ... ...................... 61

4-11 Complete image processing procedure. ........................................ ........................ 62

4-12 Manually produced image mask showing the difference between citrus fruit pixels
and non-citrus fruit pixels. .............................................. .... ........ ......... 63

4-13 Location of the target fruit and some of the leaves has shifted between the two raw
images. The Teflon sheet and some leaves remain in the same place. ..........................65

4-14 Lighting condition extremes (blackout and saturation) from inside and outside the
citru s can opy ........................................................... ................. 66

4-15 Changes in sunlight and the effect on image histograms................................. ..........67

4-16 Pixel classification results for all validation images and the overall total.........................71

4-17 Manually masked fruit pixel count versus image processing fruit pixel count results
with linear best fit using all pixel count data. ........................................ ............... 73

4-18 Manually masked fruit pixel count versus image processing fruit pixel count results
with linear best fit with one outlier removed. ........................................ ............... 73

4-19 Complete image processing and results check of validation image number five.............76

4-20 Complete image processing and results check of validation image number two .............77









4-21 Complete image processing and results check of validation image number three. ...........78









Abstract of Thesis Presented to the Graduate School
of the University of Florida in Partial Fulfillment of the
Requirements for the Degree of Master of Engineering

IDENTIFICATION AND CLASSIFICATION OF CITRUS FRUIT BY SPECTRAL
CHARACTERISTICS FOR PRECISION AGRICULTURE

By

Kevin Edward Kane

August 2007

Chair: Won Suk "Daniel" Lee
Major: Agricultural and Biological Engineering

Citrus production is a multi-million dollar industry in the state of Florida. The crop's

economic significance is the driving force behind current developments in precision agriculture

technologies assuring citrus growers can minimize their costs, protect the environment, and

increase overall yield. A real-time citrus yield map while the citrus fruit are still maturing

provides information to growers, giving them time to be proactive at improving the groves

growth and planning ahead for the harvest. Recent research has shown cameras a image

processing techniques have the ability to identify and count orange citrus fruit in the grove.

However, there is a desire by the industry to have citrus yield maps earlier in the growing season,

a time when citrus fruit are green. In this case, a traditional visible spectral camera can not

accurately identify green citrus fruit against their green tree canopy.

The objective of this research was to use spectral information from the near infrared (NIR)

reflectance spectrum to identify citrus fruit while they are still green. To begin this work, 540

freshly harvested samples of green citrus fruit and leaves were gathered and measured their

diffuse reflectance using a spectrophotometer. The resulting spectral curves from 400 nm to

2500 nm were analyzed using discriminability calculations to find critical wavelengths for

separation. Fisher linear discriminant analysis showed the wavelengths of 881 nm and 1383 nm









provided perfect green leaf and green fruit separation. This research provided a foundation for

the design of an in-field NIR camera system for green citrus fruit identification.

A highly sensitive NIR camera outfitted with three optical band pass filters (1064 nm,

1150 nm, and 1572 nm) were used for natural in-field image acquisitions. The 256 bit

monochromatic images were studied using spectral indexing and image processing schemes.

Using training images, an indexing and image processing algorithm was developed and tested on

validation images. A 90.3 % correct pixel classification result was obtained, proving that NIR

camera images can successfully be used in the identification of green citrus fruit in the grove.

An R2 of 0.746 for fruit pixel counts verse a manually masked fruit pixel counts was achieved.

Despite these positive accomplishments, the research has also revealed problems that are

prohibitive to this identification method. These problems included the high financial cost of an

NIR imaging system and the difficulty of proper illumination and multiple image alignments

when working in a normal Florida citrus grove environment.









CHAPTER 1
INTRODUCTION

Florida Citrus

Florida was first introduced to citrus in the 16th century by European explorers; with the

first commercial production of citrus beginning in 1763. Over the next two centuries the citrus

industry grew to a peak of over 350,000 ha of citrus in Florida (Sevier and Lee, 2003). In 2002,

there were 322, 658 ha of commercial citrus groves spread throughout Florida, accounting for

more than 74% of the total citrus grown in the United States. Among the citrus fruit varieties,

the orange accounted for 81.4% of the production, followed by grapefruits at 13.2%, and other

specialty fruits (tangerines, tangelos, lemons, limes, etc.) making up the last 5.4%. (Florida

Agricultural Statistics Services [FASS], 2005) With an "on-tree" value of citrus estimated to be

$1.1 billion in 2005 (Florida Agricultural Statistics Services [FASS], 2005), the citrus market has

become extremely important to Florida's economy.

Currently international trade is putting pressure on the Florida citrus market. Cheaper

labor in other countries, such as Brazil, has made the cost of staying competitive in the United

States difficult for many growers. The need to increase yield per acre while lowering overall

production costs is vital to the survival of the Florida Citrus industry. It appears that through

technology and smarter grove management techniques like precision agriculture, the citrus

industry will continue to thrive.

Precision Agriculture in Florida Citrus

Traditional whole field management approaches treat the entire crop production area in a

uniform manner ignoring natural in-field variability. This can lead to under or over application

of field inputs at different locations. Precision agriculture is a managerial technique of using

high-tech equipment to monitor and then treat smaller sections of a larger area on a site-specific









basis. This entails managing inputs (fertilizer, limestone, herbicide, insecticide, seed, etc.) to the

farming land based on the variability inherent in the field. The goals of precision agriculture are

to reduce waste, increase profits, and maintain the quality of the environment (Morgan and Ess,

2003).

Due to the "input-intensive" nature of citrus production and Florida's volatile trend of cost

per unit area, precision agriculture is becoming a vital technique to manage the cost of

production (Sevier and Lee, 2003). The Florida citrus production industry lends itself

"perfectly" to the saving that can come from managing and controlling expensive inputs to the

grove. Cost of production from input applications can be brought to "manageable" levels by

precision agriculture. In addition to minimizing input costs, precision agriculture offers better

managed groves, a protected environment, and increased crop yield and profits. Proper

implementation of a precision agricultural system requires knowledge of in-field variability; in

the case of this research, in-field citrus yield variability.

Research history and trends

According to the Citrus Research and Education Center (CREC) at the University of

Florida, in early 1996 a meeting between citrus industry personal and the University of Florida's

Institute of Food and Agricultural Sciences (IFAS) was held to discuss "decision support systems

for citrus". This meeting resulted in the creation of the Decision Information Support System for

Citrus (DISC) group at the CREC in Lake Alfred, FL. The goal of this group was to pursue

research in the area of precision agriculture, with an emphasis on: mapping citrus yield, canopy

volume, variable rate application and uses of the Global Positioning System (GPS) and

geographic information systems (GIS) in commercial groves. This group's research has lead to

advancements in precision citrus grove technologies and numerous published papers over the

past decade.









Sevier and Lee (2003 and 2004) investigated Florida citrus growers' adoption of new

technologies, including: sensor-based variable rate applicators, prescription map based variable

rate applicators, pest scouting/mapping, GPS, soil variability mapping, water table monitoring,

and yield monitoring and mapping. Their findings showed the most commonly used precision

agriculture technologies were sensor-based variable rate applicators and soil variability mapping

at usage rates of 18.6% and 18.0%, respectively (Sevier and Lee, 2003). The least used

technology was remote sensing data from plane or satellite images, with a usage rate of just over

four percent. Findings also showed that grower's age had a negative correlation with their

likelihood of incorporating the technologies investigated (Sevier and Lee, 2004). That is

younger growers were more likely to adopt the new technologies than older growers. This can

be understood in a social realm by the younger grower being quicker to learn and use the

computer technologies they have grown up with. The study also showed growers with higher

"in-grove variability" are more likely to adopt new technology when compared to growers with

less variability in their field.

In the Sevier and Lee survey (2004), respondents were asked to provide a reason for not

adopting all new technologies. The leading reason for none adoption was that growers were

"satisfied with current practices". They also described themselves as being somewhat reluctant

to adopt, and "normally wait to see other's success" before implementing new technologies.

These trends show the Florida citrus industry is moving slowly in the direction of new

technologies, while continued research at the university level is needed to show growers the

"success" they desire before adoption.

Yield monitoring and mapping

Knowledge of variable yields within blocks of citrus groves may provide information that

a grower can use to find the cause of the variability (Whitney et al., 1998). There are many









elements within a grove or block that can lead to variability, including tree sizes, age, health,

spacing, soil type, fertility, water availability, fertilizer application, and more. With proper

knowledge of the variability in the field, grove managers may act not only in the right locations

but also in a timely manner.

Fruit yield monitoring techniques may offer benefits for forecasting the number of fruit

and quality at the time of harvest. Apple yield information together with ecological, cultivar, and

price parameters can predict future yields, which in turn allows management to plan for incomes

and calculations of profit (Welte, 1990). The European apple and pear industry only uses the

'Prognosfruit' Forcasting Model for their estimations of yield quantity and quality; however, this

is a long tedious process requiring counting measurements of required parameters. This also

limits the ability to predict individual orchards' future yields (Winter, 1986; Stajnko et al., 2004).

Similarly, future citrus grove forecasting systems will also require yield quantity and quality

parameters. Thus, the Florida citrus industry, like the European apple and pear industry will

benefit from the use of fruit yield monitoring systems.

There have been several yield monitoring and mapping concepts proposed but few have

found their way to actual development and field research. For the most part these systems can be

grouped into three basic methods: 1) counting and mapping citrus tubs during manual harvesting

2) tracking citrus flow during mechanical harvesting, and 3) using automated computer vision

systems before harvesting.

Citrus grove yield maps can be created during manual harvesting by marking the locations

of filled tubs of fruit with a GPS unit in the grove (Whitney et al., 1998). This discrete data can

be interpolated to create two dimensional yield maps. Using this method, yield maps can

provide more meaning to growers by being presented in a boxes/acre format. DISC's early work









was with GeoFocus Inc, the only company at the time to offer a commercially available citrus

yield monitoring system. Their Crop Harvest Tracking System (CHTS) required the truck

operator to press a button 'marking' his/her GPS location every time a tube was lifted and

emptied into a 'goat' truck. This seemingly easy system however had problems when operated

with untrained or ill-informed drivers. Sometimes the operator would forget to push the button,

or push the button multiple times just to be sure it was done. Other times the operator would

remember to press the button later on, only after moving the truck location which means

improper markings. All these button pushing mistakes diluted the data and resulting yield map's

accuracy. Advancements to this system included automating the GPS system by way of a weight

threshold switch on the truck's lifting arm and pressure transducers mounted on the 'goat' truck

for weighing the citrus total (Whitney et al., 1999; 2001). Currently GeoAg Solutions located in

Lehigh Acres, Florida, is the only company to provide citrus yield and monitoring solutions.

They have incorporated more value added incentives with the technology such as monitoring

worker progress, chemical applications and managing payments based on harvest counts.

GeoAg Solutions' main system is called CitriTrack with three additional sub-programs

available: HaverstMap 2.0, HarvestPay, and HarvestWatch. The CitriTrack system can handle

payroll, grove mapping, and tracks real-time harvesting progress. CitriTrack uses GPS and

wireless technology to connect the grove manager directly to the grove. Up to date information

is collected by a computer (also available for purchase) that is attached directly to the harvesting

equipment. Similar to the GeoFocus system, the equipment operator logs each tub and its

location and links it with the appropriate picker. HarvestMap 2.0 uses the yield data information

collected from CitriTrack to produce informative full-color maps using GIS technology. GeoAg

Solutions claim a grower can "track the daily progress in a block down to specific boxes, specify









variations within a block for chemical applications or even target and apply products based on

return on investment potential". This data is collected in the field can also be routed to the

HarvestPay software. The program tracks the information and can provide reports about worker

productivity and hourly wages. The final sub-program, HarvestWatch, can allow the grove

manager stay connected the crews with "up-to-the-minute reports" no matter their location.

A second method currently being explored is the use of fruit flow tracking technologies on

mechanical citrus tree canopy 'shaker' harvesting equipment (Grift et al., 2006; Chinchuluun et

al., 2007). Currently a canopy 'shaker' dislodges a citrus fruit from the canopy by violently

shaking the citrus canopy. Freed citrus fruit can either fall to the ground and be harvested later

by hand or the fruit can be caught by a conveyor mechanism, lifting the citrus up to a ramp

where the citrus rolls down into a collection bin. Grift et al. (2006) have expanded previous

research in estimating particle flow rates, more specifically fertilizer, by observing and

measuring the space and time between "clumps" of particles. In this case, the "clumps and

spacings" is measure by a laser and the particles are much larger and slower moving. Another

more precise, but ultimately more complex solution, is counting the individual citrus by high

speed cameras and using image processing techniques (Chinchuluun et al., 2007). Although

current research is focused on fruit catching equipment working in conjunction with tree shakers,

it should be noted that other forms of mechanized citrus harvesting equipment would produce

even finer grade citrus grove yield maps, such as robotic citrus harvesting equipment. Such

future systems could provide not only an exact count of fruit per tree, but also a size and quality

of each fruit.

The final method explored for citrus yield map creation is the use of cameras with

automated image processing and assessment techniques to count the number of citrus on tree.









By corresponding the image's citrus counts with their acquisition location in the grove by a GPS

receiver, a highly detailed citrus yield map can be formulated. What separates this method from

the pack is the availability of the yield map before the citrus is harvested. This provides time to

grove managers to prepare more precisely the logistics of that harvested season and a head start

at improving conditions for the following growing season. The research discussed throughout

the remainder of this thesis is with respect to this on tree fruit counting citrus yield mapping

method.

Agriculturally based vision systems have been used for the identification of everything

from apples to weeds (Stajkno and Cmelik, 2005; Lee and Slaughter, 2004). On tree fruit

identification using vision systems is not a new concept with original concepts dating back

nearly forty years (Schertz and Brown, 1968). Research in fruit identification systems using

machine vision have had mixed results with wide variations of camera systems, experimental

tests, and image processing algorithms (Jimenez et al., 2000). Resent research at the University

of Florida in the area of on tree citrus identification includes Annamalai et al. (2004),

Chinchuluun and Lee (2006), and MacArthur et al. (2006). In these cases, a digital camera took

visible light images of trees and then used image processing techniques to separate

yellow/orange colors from the green canopies. In Chinchuluun and Lee's (2006) research,

multiple cameras were used and the system was pulled behind a truck while passing through the

grove rows. This allowed the images to be acquired from the side of the trees in high detail,

which permitted morphological image processing techniques to work due to the closeness of the

targets. MacAuthur et al. (2006) investigated the same concept but by using a remotely-piloted

helicopter. This gave an advantage of flexible mobility throughout the grove, but additional skill

was needed to gather high quality images. In both cases, citrus occlusion, shadows and the









separation of touching fruit were the most difficult problems faced. Research preformed on trees

with different water treatment levels provided an interesting observation. The healthier the tree's

canopy, the more difficult it was to count the fruit. A tree with thin or little leaf coverage made

counting fruit easier and increasing the yield estimation, while a tree with healthy thick leaf

coverage made the observation off all fruit imposable, thus low yield estimation (MacAuthur et

al., 2006). It has been suggested that a weighted multiplier based on canopy fullness could be

used to adjust these citrus yield estimations; however, no such work has been started at the time.

However, nearly every fruit finding vision system has used the visible light spectrum as the

only means to decipher fruit from the surrounding leaf canopy. One resent exception was the

promising work of Stajnko et al. (2004; 2005) using thermal imaging of apples, but there is no

published work with citrus imaging outside the visible spectrum. A major consideration for the

imaging of citrus is how the visible spectrum imaging research has only been preformed late in

the growing season only after citrus has changed from a green to an orange color. A gap in citrus

research is early season identification of citrus fruit, a time when the fruit is the same dark green

color as the surrounding leaf canopy. Solving these problems can allow precision agricultural

techniques to be used earlier in the growing season, thus compounding the benefits of early

information on yield, health, and in-field variability.









CHAPTER 2
OBJECTIVES

Identification of Critical Wavelengths

To properly identify and separate green citrus fruit from surrounding green citrus leaves

using only near-infrared (NIR) spectral information, an investigation into the reflective

characteristics was preformed. This involved the use of highly sensitive optical equipment to

determine the spectral response of green citrus fruit and green citrus leaves in the visible to near-

infrared range (VIS-NIR). Use of a statistically significant number of samples through the

growing season provided large amounts of data to reliably analyze. It was the first objective of

this research to identify the significant wavelengths needed to separate green citrus fruit from

green citrus leaves in a controlled laboratory environment (Chapter 3).

Green Fruit/Leaf Separation Using an NIR Camera

For future developments of early season citrus yield mapping systems, a test of in-field

green citrus fruit/leaf separation was conducted in a Florida citrus grove using a highly sensitive

NIR camera outfitted with optical band pass filters. Selection of the band pass filters was

determined by the results of the first objective. The NIR camera's resulting monochromatic

green citrus fruit and green leaf canopy images were post processed using image processing

techniques. Differences in the spectral characteristics were the only image processing tools used

for citrus identification. The second objective of this research was to determine if separation of

green citrus fruit from green citrus leaves was possible in-field using current NIR camera and

image processing technologies (Chapter 4).









CHAPTER 3
CITRUS CLASSIFICATION BASED ON DIFFUSE REFLECTANCE

Introduction

Before designing a remote sensing camera system for the identification of green citrus in a

green citrus leaf canopy, an in-depth study of the diffuse reflectance characteristics of green

citrus and green citrus leaves was required. While the spectral responses of citrus leaves and

citrus fruit had been studied many times in the past, the use of this information to find and

identify a robust separation scheme in the near infrared (NIR) range had never been published. It

was vital to identify critical wavelengths for separation, before proceeding with in-field camera

research. It is only through the spectral characteristics that proper judgments could be made for

the design and execution of the in-field NIR camera research.

Literature Review

Research in the field of near infrared (NIR) sensing technologies on agriculture has been

around for over a century. In most early research, the interactive nature between light and leaves

was studied (Williams and Norris, 2001). It has been a more recent adaptation to incorporate

NIR remote sensing equipment into the actual agricultural process, both pre and post harvest.

This can be observed in the new technologies to track and monitor crop products and personnel

(Whitney et al., 2001; Aleixos et al., 2002).

Citrus should be considered a late entry into the field of NIR sensing, despite the first

major study on NIR diffuse reflectance of citrus fruit being three decades old (Gaffney, 1972).

Since that research no real follow up work was conducted. Most likely this gap was due to two

issues: 1) lack of valuable uses of the NIR technologies in the industry. Agriculture is among the

slowest industries to research and adapt new technologies, as managers traditionally held a, "if it

works don't fix it" mentality (Sevier and Lee, 2003), and 2) the high cost of performing NIR









research in the past. Lower prices for better equipment and technologies are quickly eliminating

these concerns for both the researchers and the eventual end users, the grove managers. Florida,

and all U.S. citrus growers, will need to apply new managing solutions to compete with the

growing international markets (Florida Agricultural Statistics Services [FASS], 2005). This

means high-tech solutions are desired for pre-harvesting and post harvesting, resulting in the

increase of NIR sensing research over the last decade.

Currently the most commonly used NIR technology in the agricultural fields of the world

is multi- and/or hyper-spectral imagery from planes and satellites. This has lead to a lot of

research in spectral vegetation indices (SVI) for finding a great number of important field

information, such as disease, crop moisture, and weeds (Apan et al., 2003; Alchanatis et al.,

2006; Gumz and Weller, 2005). Despite this, there has been little research preformed with these

types of cameras and sensors on the ground. This research is a preliminary study of the

reflectance characteristics of green citrus and green citrus leaves for the conceptual method

design of a simple multi-spectral ground system, for separation of green citrus from green leave

canopy.

Materials and Methods

Green Citrus and Green Leaves

During the fall 2005 citrus growing season, June 2005 to January 2006, samples consisting

of one green citrus leaf and one green citrus fruit located next to each were acquired from the

University of Florida's citrus research grove. A sample set consisted of 10 samples (10 fruits

and 10 leaves) gathered on the same day and of the same variety. Sample sets were obtained

weekly from one of the three citrus varieties: Hamlin (Citrus sinensis), "Orlando" Tangelo

(Citrus X tangelo), and Valencia (Citrus sinensis). There were a total of 27 sample sets









collected: for a total of 270 individual fruits and 270 individual leaves. The dates and varieties

of each harvesting sample set are listed in Table 3-1.

Table 3-1. List of dates that variety sample sets were harvested.
Tangelo Hamlin Valencia
7/6/05 7/12/05 10/18/05
7/19/05 7/26/5 11/1/05
8/2/05 8/23/05 11/22/05
8/9/05 9/7/05 12/6/05
8/16/05 9/20/05 12/20/05
8/30/05 10/5/05 1/3/06
9/14/05 10/19/05 1/11/06
9/27/05 11/2/05
10/11/05 11/29/05
10/25/05 12/13/05

Figure 3-1 is an example of the green citrus and green leaves that were harvested during

the growing season. Each sample was weighed to the nearest 1/100th of a gram by a digital scale

(Adventurer, Ohaus, Inc., Pine Brook, NJ), and diameters (horizontally at the widest cross

section with the stem pointing upwards) were measured using a sliding caliper to the nearest

1/100th of a millimeter on the same day as being harvested.













A B

Figure 3-1. Samples from August 9, 2005 of A) green citrus fruit and B) green citrus leaf.

Figure 3-2 shows the citrus varieties growth during the fall season as measured by average

width. Only ten citrus fruit samples were harvested every other week, or more, which allowed










the variance in average size to swing up and down. Close observation of Figure 3-2 even shows

instances where the average size of the ten samples declined from the proceeding ten samples,

such as Hamlin week two to week four. This was due to no other reason than the random nature

of the fruit sizes harvested. The overall growth pattern, fruit sizes and seasonal developments

are clear. Growth results, as measured by weight, showed similar results but have been excluded

due to its irrelevance to the remainder of this thesis.

75

70







55 ---e--- Tangelo
-- Hamlin
50 ...... ...... Valencia


1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
Week Number (week 1 is June 6th, 2005)

Figure 3-2. Growth chart of the average fruit varieties' size, as measured by fruit diameter.

Diffuse Reflectance

Diffuse reflectance of the leaf and fruit samples was measured by a spectrophotometer

(Cary 500, Varian, Inc., Palo Alto, CA) with an integrating sphere (DRA-CA-5500, Labsphere,

Inc., Brossard, Qc, Canada) on the same day as harvesting (Figure 3-3). The samples were not

cleaned prior to measurements as natural "on tree" and "in-field" characteristics were desired.

The spectrophotometer measured percent diffuse reflectance of the samples between 200 and

2500 nm in one nanometer increments. When the spectrophotometer was used, the two lamps

(deuterium and tungsten) were allowed to warm up for one hour prior to testing to stabilize the









light sources. A diffuse reflectance baseline was measured using a 50 mm diameter

polytetrafluoroethylene (PTFE) sample disk, which was used to obtain the optical reference

standard for the system each day.





integrating
sphere



citrus Cary 500











Figure 3-3. Green citrus fruit positioned for diffuse reflectance testing with the Cary 500
spectrophotometer with integrating sphere.

Discriminability

Two types of classification were desired from the data set. The most important was a

distinction between citrus leaves and any citrus variety. This recognition would allow citrus

growers to identify all immature green fruit regardless of variety. This information is critical as

it offers transferable functionality to the research not only to local oranges and tangelos but also

to limes, grapefruits and other members of the citrus family. The second classification studied

the separation of citrus fruit into their variety classes: Tangelo, Valencia, and Hamlin. There

were normal minor spectral reflectance changes over the course of the six month growing season

due to chemical and physical alterations within the fruits' peel, based on the research ofNagy et










al. (1977). However, for the purpose of this research a normal Gaussian distribution was

assumed at each wavelength for the reflectance of all sample fruits and leaves. Histograms of all

fruits and leaves for wavelengths of 1650 and 881 nm are shown in Figures 3-4A and B,

respectively. These example wavelengths serve as a visual comparison of 'strong' and 'weak'

discriminability.


60 60
Fruit Fruit
50----- 50--
S Leaf Leaf
S40- 40-
O O
S0 0 30




10----10--

5 10 15 20 25 30 35 40 45 50 45 50 55 60 65 70 75 80 85 90
A Reflectance (%) at 1650 nm Reflectance (%) at 881 nm B

Figure 3-4. Histograms of example fruit and leaf reflectance. A) Weak discriminability at 1650
nm. B) Strong discriminability at 881 nm.

Discriminability comprises two mathematical measures: the distance between means and

the standard deviation of two probability density functions (PDFs). If we image the graphs from

Figure 3-4 to be normalized and have an area under the curve of one unit, then Figure 3-4(A)

shows a small overlap between the two PDFs, while Figure 3-4(B) shows a gap between the

PDFs. This difference is a result of the discriminability of the two wavelengths.

The mathematical strength of the discriminability between two PDFs with the same

standard distributions is defined by Duda et al. (1988) as:


d'= 2 (3-1)


Where, d' = discriminability

a = standard deviation









/i, //2= means of class 1 and 2

In general, a higher discriminability value is desired as it shows a greater separation exists

between classes. In the case of this data set, the standard deviations are different between the

two classes, thus the formula above can not be used. For this reason the following alteration to

equation 3.1 was made:


d'= (3-2)
(a, +a,)/2

Where, d' = discriminability

1, o2 = standard deviations of class 1 and 2

A/I, /2= means of class 1 and 2

Averaging the standard deviations of the two classes and replacing this for the standard

deviation of equation 3.1 allows the discriminability to scale with the standard deviation change

of both PDFs.

Discriminability is a simple and reliable way to determine which wavelengths have the

greatest possible separation; however, this should not be the only consideration. Reflectance

characteristics at one wavelength share common properties with those wavelengths near it. To

maximize the quality of multiple feature extraction, a limit on the minimal distance between two

wavelength features was set. This threshold limit will be defined in the Training and Validation

section that follows.

Fisher Linear Discriminant Analysis (FLDA)

Dimensionality reduction is an issue with many recognition systems using more than one

variable. The most well known and commonly used dimension reduction technique is principal

component analysis (PCA). This method searches for feature spaces in the multi-dimensional









data that contains the greatest separation between classes with regard to variance. By projecting

multi-dimensional data points onto a new calculated space, a smaller number of dimensions

maybe used for classification. In the case of two dimensional data, a one dimensional direction

is found. Fisher linear discriminant analysis (FLDA) is a common type of PCA first seen in the

famous paper by Fisher (1936). In our research the analysis technique was used to calculate the

direction, w, for projection. Duda et al. (1988) defines w as the "linear function yielding the

maximum ratio of between-class scatter to within-class scatter". A positive byproduct of this

analysis technique is a reduction of "noisy" directions. A one-dimensional projection direction

was calculated, reducing data from a two dimensional space, defined by two feature

wavelengths, to a one dimensional space. This methodology was chosen for this research as it

adapts well to future NIR computer vision and image processing research. By using this method

a stack of digital images at different spectral bands, know as a spectral image cube, can be

dimensionally reduced to a single two dimensional image. Each pixel of this image would have

its own uniquely calculated class likelihood.

Training and Validation

Prior to using the pattern recognition techniques discussed previously, a performance

evaluation was conducted using two-thirds of the samples as training data and one-third as

validation data. The 180 fruit and leaf training samples were selected at random. Using only the

training samples, wavelengths for separation were selected, by discriminability as was discussed

previously. FLDA was then used to find the best projection vector (direction). The remaining

90 fruit and leaf validation samples were classified after being projected onto the one

dimensional space. Only the discrimination between fruit and leaf was verified using these

techniques, as it was the primary purpose of the research.









For fruit vs. leaf classification, two methods (Method I and Method II) were investigated.

Method I included the use of the feature wavelengths with the two highest discriminability (d')

calculations that also met a threshold distance of 100 nm apart. This 100 nm threshold was

chosen arbitrarily to remove multicolinearity. Thus, the two selected wavelengths were allowed

anywhere in the spectral range as long as they remained 100 nm apart. Method II required one

wavelength feature having a fruit reflectance greater than that of the leaf, while the second

wavelength feature having the fruit reflectance less than that of the leaf. The reader might be

best served to glance over Figure 3-5 of the next page to better understand Method II. This

second method was tested for fruit and leaf classification as it offered wavelengths with

contrasting reflectance magnitudes.

Results

Reflectance Characteristics of Green Citrus Fruit vs. Green Citrus Leaves

Figure 3-5 presents the average reflectance spectra of 270 citrus samples with respect to

the 270 leaf samples. From the spectral reflectance characteristic curves, two traits should be

observed. The large increase of reflectance between 690 and 720 nm is referred to as the red

edge. This is a result of plants having higher absorption rates in the visible range due to

photosynthetic pigments. Absorption rates of the red edge region by chlorophyll pigments are

lower resulting in higher reflectance (Ding, 2005). The second trait is water absorption bands at

970, 1450, and 1940 nm. These valleys in reflectance magnitude are a result of light absorption

by water within the samples.

Differences between the fruit and leaf reflectance include a large magnitude change from

720 to 1120 nm. A simple experiment suggests this to be a result of the thickness of the citrus

fruit compared to the thinness of the leaves. This was verified by stacking multiple leaves and

testing the reflectance as the leaf count increased. The results showed that an increase in number









of leaves lead to an increased diffuse reflectance, most significantly between 720 and 1400 nm,

as seen in Figure 3-6. This is explained as the first leafs transmitted light may become a second

leaf s reflected light. Thus, the more leaves the more opportunities there are to increase the

reflectance (Figure 3-7). This theory is in agreement with Williams and Norris (2001) which

preformed similar tests on thin potato slices. There were no significant increases in the

reflectance due to leaf layering, outside this 720 to 1400 nm range. These results are supported

by Fraser et al. (2002) which showed light penetration depths in apples was larger in the 700 to

900 nm range than 1400 to1600 nm. They claimed this to be a result of the absorption profile of

water. A second important spectral difference between fruit and leaf in Figure 3-5 is the curve

crossover occurring at near 1150 nm. All citrus varieties showed lower reflectance than the

citrus leaves at higher wavelengths.


90

80
80 Fruit

70 .............. Leaf

S60 .

0 50 I

o 40

S-30

20 /

10

400 800 1200 1600 2000 2400
Wavelength (nm)
Figure 3-5. Green citrus and leaf average spectral reflectance through the fall 2005 growing
season.










90

80 Single fruit
Five leaves
70 -........... Three leaves
60 Single leaf

50

0 40

a 30

20


0
400 800 1200 1600 2000 2400
Wavelength (nm)
Figure 3-6. Spectral reflectance of layered citrus leaves and single "Orlando" Tangelo.


n Alight
< > source





transmission
increased diffuse
reflectance

Figure 3-7. How multiple leaves/surfaces create added diffuse reflectance.

Reflectance Characteristics of Citrus Varieties

All three of the tested citrus varieties are highly correlated throughout the ultraviolet-

visible (UV-VIS) to NIR range, as shown in Figure 3-8. It should be explained that the

difference in the visible light range average, 400 to 750 nm, was a result of the fruits maturing

during the experiment and the color brightness differences of those mature fruits. A difference in

the chlorophyll to carotenoid conversion is the main cause of this color brightness (Merzlyak et









al., 1999). A consistent separation of citrus variety is seen in the diffuse reflectance magnitudes

reflectance of the higher NIR spectral waves: Hamlin had the highest, followed by the Valencia

and finally the Tangelo showing the lowest (Figure 3-9). Another way to view the separate of

varieties is based on their individual standard deviations as compared to the standard deviation of

all the citrus samples, presented in Figure 3-9. It shows the greatest standard deviation of all the

varieties to be much higher in the spectral range of 1400 to 1880 nm with an additional standard

deviation range maximum around 2200 nm. Meanwhile, the standard deviations' of the

individual varieties remain low at those same wavelength ranges. This suggests that they are

prime wavelengths for variety classifications information, a statement which will be

mathematically proved later (Table 3-3). Note that the highest standard deviations are at 660 nm

due to the color conversion mentioned before. Because the standard deviation of the individual

citrus varieties remains high here as well, this would not be a good location to gather

classification information.

The average reflectance of leaf varieties showed a maximum less than 3% separation.

Those separations are insignificant when compared with standard deviations on the order of

3.5%, thus the leaf variety classification was not possible with our results.

The reflectance curves were converted to absorbance spectra by use of the Beer-Lambert

law, Equation 3-3 (Williams and Norris, 2001) (Figure 3-10). Although analysis for light

absorbance compared to absorber is not discussed in depth in this thesis, the reader should

understand that light absorbance by materials is completely dependent on the molecular

structures of the said material. More specifically the Beer-Lambert Law states, the concentration

of an absorber is directly proportional to the sample absorbance (Williams and Norris, 2001).

This fact is used throughout the agricultural and food industry in non-destructive examinations of











food. It should also be realized by the reader that the different absorbance curves of Figure 3-10


mean different absorber components and quantities exist among the citrus varieties.


A4 log (3-3)



Where, A = Absorbance and R = Reflectance


90

80

70

g60

o 50
C
0 40

r 30

20

10

0
400



Figure 3-8.

0.35

0.3

0.25
0o

c 0.2
C

c( 0.15


03 0.1

0.05

0-
400



Figure 3-9.


............. Tangelo
. Hamlin
-- Valencia


800 1200 1600 2000 24
Wavelength (nm)


Citrus varieties average spectral reflectance.


1400


.... Valencia
............ Ham lin
Tangelo
All Varieties





1880 2200
2200


800 1200 1600 2000 2400
Wavelength (nm)


All citrus and citrus varieties standard deviation values for all wavelengths.











1.4

Water absorption bands
1.2 1940


U)
1450
660

S0.8

0.6 .. o
f 0.6
0
_Q -. 1 /.
< 0.4- Tangelo
970 '"--- ........... Hamlin
0.2
S-- Valencia

0
400 800 1200 1600 2000 2400
Wavelength (nm)


Figure 3-10. Citrus varieties average spectral absorbance.

Spectral Growth Patterns of Maturing Citrus Fruit

To verify that the chemical and biological changes occurring inside the maturing green

citrus fruit would not significantly harm our identification methods, a brief survey of the spectral

growth pattern was conducted. Previous research (Merzlyak et al., 1999) conducted on the

maturing process of citrus focused mostly on identifying wavelengths signaling growth. In this

research, sampling ranges that show growth are to be avoided, to prevent maturing from harming

the accuracy of our identification. A solution that can identify both early season and end of

season citrus offers more robustness and/or could be implemented other future citrus harvesting

systems.

Changes to reflectance are dramatically illustrated in Figure 3-11, where early season

Hamlin sets harvested on July 12, 2005 and September, 20 2005 are compared with later sets on

November 2, 2005 and December 12, 2006. By mid December, 2005 Hamlin samples turned a

bright orange. To design a machine vision system that identifies green or orange citrus fruit










from green citrus leaves, the 500 to 750 nm range needs to be avoided due to these dramatic

spectral reflectance changes.


90

80





S50I
o0 4 I
I f II



--------------------11-2-05
20
lO...] .. .... 12-05 9-20-05



400 600 800 1000 1200 1400
Wavelength (nm)

Figure 3-11. Reflectance changes over the growing season of Hamlin.

Training and Validation Results

Discriminability results of the training data set, Methods I and II are shown in Table 3-2.

The training data shows separation between green citrus fruit and leaves to be the strongest at

863 nm, with a discriminability value of 7.84. The strongest separation wavelength based on

discriminability that met the 100 nm distance threshold was 763 nm. The strongest wavelength

for separation above 1150 nm, where leaf reflectance is greater than fruit reflectance, was 1389

nm. These second and third feature wavelengths, 763 and 1389 nm, respectively, show 'weaker'

separation but are still relatively 'strong' for classification purposes.

Scatter plots of the validation data using Methods I and II are shown in Figures 3-12 and 3-

13, respectively. Using a "classification line" passing through the centroid of the training data

and perpendicular to the "projection line", all samples but one validation fruit were correctly









Table 3-2. Discriminability results of the training data set, Methods I and II.
Feature 1 Feature 2
wavelength (nm) discriminability, d' wavelength (nm) discriminability, d'
Method I 863 7.84 763 6.72
Method II 863 7.84 1389 4.87

identified using Method I wavelengths, yielding R2 = 0.994 (Figure 3-12). Using Method II

wavelengths, all of the validation data was correctly identified, yielding R2 = 1.000 (Figure 3-

13). It can be observed that Method I of the fruit to leaf separation possesses the strongest

discriminability values; however, the scatter plot of Method II displayed stronger separation in

the two dimensional space, as shown in Figures 3-12 and 3-13.


90
o Leaf
E v Fruit
c 80-----------

-0
70 --
8 projection
v^ 9 line
<---
6 0 - - -

S/ classification line
o /
50 60 70 80 90
Reflectance (%) at 863 nm

Figure 3-12. Fruit & Leaf (Method I) with Fisher projection line (solid line) and classification
line (dashed line).

Discriminability Using All Samples

The discriminability results using all the samples data is shown in Figure 3-14. It is

important to notice the general trend of this graph and what regions show the 'strongest'

discriminability, 740 to 940, 1060, 1380, and 1570 to 1830 nm, and what regions have the
















COO O 0
o 40 5060
coto

0v

cIv
w a, t30 ---------- ---- 1 4V0 10 -

V
o Leaf
( V
'" ,0
20ptv Fruit
w tv
Iy V


10
40 50 60 70 80 90
Reflectance (%) at 863 nm

Figure 3-13. Fruit & Leaf (Method II) with Fisher projection line (solid line) and classification
line (dashed line).

'weakest' discriminability, 1140, 1930, and 2500 nm. Although the remainder of the

mathematical analysis in this Chapter is dependent on the randomly selected training and

validation sets, this figure will be referenced back to in Chapter 4 in the selection of optical

equipment.

The results of discriminability calculations using all samples are similar to the training data

results, with the strongest discriminability at 881 nm. Using Method I gives the second feature

wavelength at 781 nm while Method II gives 1383 nm, as shown in Table 3-3. Notice that

'weak' discriminability in wavelengths with high mean separations can be an artifact of the

multi-modal nature inherent in the total fruit PDF, for example 1650 nm seen in Figure 3-4A.

The difference between fruit variety reflectance increases the standard deviations at these

wavelengths. The three peaks seen in the fruit histogram of Figure 3-4A serves as visual

evidence of the three fruit varieties' multi-model effect.











8
881
7 Fruit and leave
"cross over"
6-
781 1383
5-
-- 5 t1

Cn
*.t 3 ,
/ /




0
400 800 1200 1600 2000 2400
Wavelength (nm)

Figure 3-14. Discriminability results of all the sample citrus vs. leaves.

Table 3-3. Discriminability results using all the data.
Feature 1 Feature 2
wavelength (nm) discriminability, d' wavelength (nm) discriminability, d'
Fruit & Leaf 881 7.44 781 4.98
(Method I)
Fruit & Leaf 881 7.44 1383 4.92
(Method II)
Tangelo & 1712 5.52 1392 5.44
Hamlin
Tangelo & 1417 2.96 1882 2.83
Valencia
Hamlin & 1711 3.20 1813 2.89
Valencia

Figures 3-15, 3-16, and 3-17 show scatter plots created by the feature discriminability

results in Table 3-3. When observing the fruit variety plots, most significantly Figure 3-15, the

difficulty in using features with dependencies on each other are seen. The result is a class cluster

of 'cigar shapes' with high variances in the same direction. Incase of Figure 3-15, the variance is

between the bottom left corner and the top right corner of the graph. This can create difficulty










with class separations by means of PCA, due to the dependence of one feature space on the

next.


o Hamlin

v Tangelo


16 20 24 28 32
Reflectance (%) at 1712 nm


Figure 3-15. Tangelo & Hamlin with Fisher projection line (solid line) and classification line
(dashed line).


29-- -
o Valencia
E v
c- v Hamlin / -
-- 25 --
co








S17




13
15 19 23 27 31
Reflectance (%) at 1711 nm


Figure 3-16. Tangelo & Valencia with Fisher projection line (solid line) and classification line
(dashed line).


E
c 26
C(
0)
CO
M 22

U)
0 18

4--
( 14
ry


10
12













E d Vill d l;lo
12 V Tangelo -
8- -------------- O----


104-W- -------

co





6,
8 10 12 14 16
Reflectance (%) at 1417 nm

Figure 3-17. Hamlin & Valencia with Fisher projection line (solid line) and classification line
(dashed line).

Fisher Linear Discriminant Analysis (FLDA) Results

The solid lines in Figures 3-12, 3-13, 3-15, 3-16 and 3-17, are the projection lines

calculated by FLDA. While several of the calculated projection lines, w, are intuitively correct

like Figure 3-13, others are counter intuitive such as Figure 3-16. The parallel nature of the two

'cigar shaped' class clusters is the cause of this odd looking projection line. The reasoning is

that the class means of the projected data on to the one-dimensional space is close in distance;

however, the "between-class scatter to within-class scatter" (Duda et al., 1988) is the best for the

most accurate classification, because of a lower variance. A look back at Figure 3-12 is an

example of the same tendency for the projection line to be perpendicular to 'cigar shaped'

clusters.

Mathematical definitions of the calculated projection vectors for the two class systems are

shown in Table 3-4. Notice the projection space is one dimensional and not defined by a line but










rather a direction; a line was included in the figures only as a visual reference. A complete

scatter plot including all 270 data samples of every class is in Figure 3-18. The x-axis uses a

feature wavelength of 881 nm as it displayed the strongest discriminability for fruit and leaf

identification, while the y-axis uses a feature wavelength of 1713 nm as it displayed the strongest

average discriminability between fruit varieties.

Table 3-4. Projection vector equations found by Fisher Linear Discriminant analysis
Feature 1 (x-axis) Feature 2 (y-axis) Projection vector, w
Fruit & Leaf 881 nm 781 nm w = (-0.1309, 0.0439)
(Method I)
Fruit & Leaf 881 nm 1383 nm w = (-0.0612, 0.0380)
(Method II)
Tangelo & Hamlin 1712 nm 1882 nm w =(-0.5119, -0.1092)

Tangelo & 1417 nm 1813 nm w =(-0.9040, 0.7114)
Valencia
Hamlin & Valencia 1711 nm 1381 nm w =(-0.1751, -0.0312)


40












SV Hamlin
Co


E 15 V T









50 60 70 80 90
0
Ref 20 Valenciance (%) at 881 nm
SHamlin
S15 O Tangelo
E Leaf

10
50 60 70 80 90
Reflectance (%) at 881 nm

Figure 3-18. Scatter plot showing all classes: Valencia, Hamlin, Tangelo, and Leaf









Conclusions

Samples of green citrus varieties and citrus leaves were harvested during the late 2005

growing season. Diffuse reflectance over the UV-Vis and NIR range (200 to 2500 nm) was

measured with a spectrophotometer. Average spectral reflectance curves show the opportunity

to use NIR sensors and/or cameras to identify citrus fruit from citrus leaves, but also to classify

different citrus varieties. Wavelength features for classifications were chosen by discriminability

calculations. The selected feature spaces were used to create two dimensional scatter plots,

which were then used to calculate the best projection line by Fisher linear discriminant analysis

(FLDA). Using two-thirds of data as training and one-third data as validation, an R2 of 1.0 was

possible using these pattern recognition techniques. As expected, separating green leaves from

green citrus fruit proved to be more accurate than distinguishing among different citrus varieties.

It has been shown in this Chapter that while using only two NIR feature wavelengths,

extremely accurate green citrus fruit from green citrus leaf identification is possible. A scatter

plot of feature wavelengths 881 nm (x-axis) and 1383 nm (y-axis) projected onto a one

dimensional feature space defined by the direction (-0.0612, 0.0380) proved to be the best

mathematical method for separation when using the data gathered in the laboratory. The

transferability of these critical wavelengths and mathematical methods to in-field systems have

some interesting difficulties to be discussed later.

This Chapter's research was designed to test NIR data validity for the use in a computer

vision system for counting green citrus yield in-field, while still on the tree. The in-field testing

of such a NIR camera based system is in the following chapter, Green Citrus Identification by a

NIR Camera System (Chapter 4).









CHAPTER 4
CITRUS IDENTIFICATION BY NIR CAMERA SYSTEM

Introduction

After completion of the spectral analysis of green citrus fruit and green citrus leaves, it was

important to use that knowledge in the selection of optical camera equipment. The selection of

wavelengths is of significance during image processing because an understanding of the

expected spectral responses helps in spectral index designing and testing.

Cameras Used in Citrus Yield Mapping

Fruit identification using machine vision systems was proposed nearly forty years ago

(Schertz and Brown, 1968). However, technology has been only recently advanced enough to

allow researchers to investigate their usefulness more fully. Major fruit identification studies

have focused mostly on apples and citrus fruit, the most common application being robot

harvesting (Jimenez et al., 2000). Traditionally the visible spectrum has been used for fruit

identification, which lends itself very conveniently to non-green colored fruits, such as red

apples and orange colored citrus. The biggest issues with these camera systems have been

occlusion and grouped fruit segmentation (Jimenez et al., 2000).

Citrus harvesting system research has used cameras with different forms of traditional

machine vision for many years with mixed results and a wide variation of algorithms (Jimenez et

al., 2000). Recent research in this field of study at the University of Florida includes Annamalai

et al. (2004), Chinchuluun and Lee (2006) and MacArthur et al. (2006). However, all of these

systems utilized only the visible light spectrum as a means to decipher fruit from the surrounding

green leaf canopy. This leads to problems with early season citrus identification, a time when

citrus are a dark green color, the same color as the leaves. By solving this problem, precision









agricultural techniques can be used earlier in the growing season, compounding their benefits of

early information on citrus fruit yield, health, and in-field variability.

Annamalai and Lee (2004) proposed a method to decipher green citrus fruits from leaves

by their spectral differences in the near infrared (NIR) region. This work was extended in

Chapter 3 by using a spectrophotometer to identify critical wavelengths that could be used to

separate green citrus fruit from green citrus leaves. The objective was to test a simple non-

destructive computer vision system for the identification of green citrus fruit while "on-tree" and

during normal in-field growing conditions utilizing the previously obtained results of Chapter 3

Materials and Methods

NIR Camera, Optical Equipment, and Hardware

A Merlin NIR InGaAs camera (FLIR Systems, Inc., Indigo Operations; Wilsonville, OR)

was used for all image acquisitions. The spectral range of this NIR camera is 900 to 1700 nm

with the light sensitivity dropping off to zero at each side of this range and a peak sensitivity at

approximately 1600 nm. This sensitivity curve is true of all InGaAs technology and is displayed

in Figure 4-1, courtesy of FLIR, Inc. The Merlin NIR camera recorded 640 x 480 pixel

monochromatic images (307200 pixels), saved in a two-dimensional grayscale TIFF format with

256 possible pixel values (0 to 255). Each image was taken with one of three optical band pass

filters (1064, 1150, and 1572 nm) positioned in front of the camera lens, permitting only a thin

spectral band of light to pass. Figure 4-2 shows the three optical band pass filters used during

the image acquisitions of this study. Additional technical details about each band pass filter are

provided in Table 4-1.

The two most critical wavelengths discovered and discussed in Chapter 3, for the

separation of all green citrus varieties from green citrus leaves were those of Method II (i.e., 881

nm and 1383 nm). However, the spectral range of an NIR InGaAs camera starts at 900 nm, and











Spectral Sensitivity
1.2

1.0- -




a.
0.8- InGaAs J "J "" "

ID

o 0.4-

- 0.2-




wavelength [nanometers)


Figure 4-1. Spectral sensitivity range of InGasAs technology (Graph courtesy of FLIR Systems,
Inc.).


Figure 4-2. Three band pass (1150 nm, 1064 nm, and 1572 nm).


Table 4-1. Optical band pass filter details.


Supplier/Manufacture
ThorLabs
ThorLabs
Andover


Model
FL1064-10
FB1150-10
1572.0 / 20.0 27539


CWL (nm)
1064.0
1150.0
1572.0


FWHM (nm) Transmittance
10.0 70%
10.0 45%
15.0 75%









even at 900 nm the spectral sensitivity is very low, less than 20% of the maximum sensitivity

range. Referring back to Chapter 3, the wavelength selection to best mimic the wavelength of

strongest citrus fruit and leave separation is around 1060 nm. This was the reasoning for the

selection of the FL1064-10 model with a central wavelength (CWL) of 1064 nm. The second

band pass filter, model FBI 150-10 has a CWL of 1150 nm and was selected because the

wavelength had a low discriminability between citrus and leaf, making it ideal for normalizing

images.

The final band pass filter was the 1572.0 / 20.0 with a CWL of 1572 nm. A filter selection

of 1383 nm was not made fore two important reasons. One, no filter around 1383 nm is

commercially available while maintaining a transmittance of over 50% and a full width half

maximum (FWHM) value of less than 50 nm. Two, and far the most importantly, the spectral

solar intensity between 1450 and 1350 nm is low, close to zero. It is desirable to use

wavelengths of higher solar intensity to present the best lighting conditions possible (Figure 4-3).

The range 1550-1590 nm provides the stronger solar intensity that is desired. Also referring to

Figure 4-3 the spectral irradiance of 1064 nm is a strong 0.65 W m-2 nm-1, while the 1150 nm

remains lower but still functional with a spectral irradiance at about 0.3 W m-2 nm1.

Figure 4-4 shows spectrophotometer (Cary 500, Varian, Inc., Palo Alto, CA) transmittance

results of the three band pass filters. The observation of most interest is the transmittable light

bands outside each specific band pass filter design range (i.e., the 1064 nm band pass filter

transmittance spikes at 800 nm, 1600 nm, and above; the 1572 nm band pass filter transmittance

spikes at 1950 nm and above). Most of these extra band passes are limited by the NIR camera

range of 900 to 1700 nm; however, the exception is the 1600 nm spike of the 1064 nm band pass










filter. While this was of some concern, it should be noted that the solar intensity at 1064 nm is

greater than that of 1572 nm. As a result, the transmittance of the 1600 nm spike are overcome


15-








C 075


05
CU

C 025-



0
300 500 700 900 1100 1300 1500 1700 1900 2100 2300 2500
Wavelength (nm)



Figure 4-3. Solar spectral irradiance reference from the ASTM G173-03 of "direct circumsolar".
This data was gathered by the American Society for Testing and Materials (ASTM)
and government research laboratories.

by the stronger solar light at 1064 nm, having only a small effect on the resulting images. In

contrast, the same transmittance levels for the 1572 nm band pass filter remain acceptable,

because the NIR camera mechanical iris can be opened allowing enough of the 1570 nm solar

light in. In fact, during image acquisition the iris was always opened more for the 1572 band

pass filter imaging then for the other two filters. The band pass filters were all bi-directional

with the same transmittances in both directions. Although this was verified by the

spectrophotometer, the results shown in Figure 4-4 are from the same direction as was

consistently used throughout the image acquisition process.










100
0 ......... 1150 nm band pass filter
90
1064 nm band pass filter
80 1572 nm band pass filter

o 70

D 60

50

u) 40

30
20

10
0 ..... ......................
400 800 1200 1600 2000 2400
Wavelength (nm)

Figure 4-4. Percent transmittance for each of the band pass filters.

An in-field experimental setup used a gasoline generator to supply a standard North

American wall outlet power supply of 120 Volts, peak-to-peak. The generator powered a

computer monitor, the NIR camera, and a personal computer (Pentium 4, 2.4 GHz). The NIR

camera was connected by coaxial cable to the PC, which was running image acquisition software

(Intellicam, Matrox Electronic Systems Ltd., Dorval, Canada). The in-field experimental setup

is shown in Figure 4-5. The use of a light weight table and wagon was to facilitate the easy

moving of equipment from one image location to another.

Experimental Location and Environment

In order to acquire multispectral images, the NIR camera was positioned in front of citrus

fruit and leaves and an optical band pass filter was manually locked into place in front of the

lens. Images were taken before switching to the next optical filter without moving the NIR

camera. Each time a filter was positioned, several images (three to seven) were obtained with

the same filter. A commercial digital camera (Canon Digital Elph S300) was used to record a









visible spectrum image of the NIR camera field of view to be used later for target scene

referencing. For the purpose of this thesis, an NIR camera field of view showing citrus fruit and

leaves will be referred to as a "target"; a group of spectral images acquired at a target will be

referred to as an "image", and a "raw image" will refer to a single monochromatic image.




I&

















Figure 4-5. In field experimental setup with gasoline generator, personal computer, monitor,
NIR camera on tripod, and Teflon disk (from foreground to background).

Preliminary test images were acquired outside in a controlled environment before taking

the system to the grove. This was to test the system and handling of the optical equipment

before being isolated in a citrus grove. Thirty-six images were acquired at the University of

Florida Citrus Research Grove in Gainesville, FL for in-field testing of the multi-spectral

imaging system. Target citrus trees were all Hamlin variety. Both the test images and in-field

images were acquired in November 2006 under sunny weather conditions. A total of 552 raw

images were obtained.









The predetermined experimental plan included a Teflon disk or sheet to be used for raw

image light intensity normalization. However, the brightness of the Florida sun saturated the

Teflon material in the raw monochromatic images (refer to raw images of Figures 4-9A, 4-10A,

or 4-14). Efforts to limit the light to the camera by closing the NIR camera iris and/or shading

the Teflon or target area limited the visibility of the fruit and leaf targets. Raw images with the

Teflon were used in the study; however, the Teflon was ignored.

Image Processing

Results from in-field image acquisitions were treated as a series of three-dimension matrix

values. The x- and y-axes formed the two dimensional image while the z-axis was the number of

images at the target. This z-axis is also referred to as an image block depth. The z-axis length

varied among images depending on the total number of raw images gathered from the target. It

was essential for this research that each image block included at least three raw images: one at

each of the band pass filters (1064 nm, 1150 nm, and 1572 nm). In every image block studied

one raw image for each band was selected for image processing based on image clarity,

alignment, and brightness. When only images of each spectral band are used to create an image

block it is known as a spectral image block (Gonzalez, et al., 2004). Figure 4-6 shows an

example of a spectral image block used in this study.

Image processing was conducted using MatLab 7.0 software with the Image Processing

Toolbox (MATLAB, 2006). Raw images were treated as 480 x 640 two dimensional matrices

consisting of 8-bit unsigned values. All image processing techniques and mathematical

calculations preformed between raw images were computed on a matrix point by point (pixel by

pixel) format.






















y-axis




z-axis


_x-axis

Figure 4-6. Representation of an image spectral block.

Image processing's major stages

The image processing steps used in this study consisted of two major stages, each having

sub-steps. The two major stages will be referred to as Image Indexing and Marker Checking. In

the Image Indexing stage a likelihood image was calculated of each pixel being one of two

classes, either citrus fruit or non-citrus fruit. It should be noted that not every pixel is either

citrus fruit or citrus leaf as images include branches, sky, ground, Teflon, metal pole or other

non-fruit objects; all these possible non-fruit objects are referred to as non-fruit, as it is the desire

of this image processing to not count them as citrus fruit. Likelihood images were created based

on pixel values of each band pass image. Using Otsu's method, a threshold value was found and

each pixel was classified as either citrus fruit or non-citrus fruit. In the Marker Checking stage, a

calculation from the original raw images was used to check the validity of the citrus fruit

markers, or pixel groups, identified as possible citrus fruit pixels. This calculation was based on


"









the regional variances which describe the complexity of the image marker. A more complex,

thus higher variance region is more likely to be leaves or branches intertwined while a citrus fruit

will be less complex. The stages will now be discussed in more detail.

Image Indexing stage

The first image processing step of the Image Indexing stage is a simple histogram

stretching in order to maximize the reflective light intensities of the monochromatic images.

This assures that 1) the maximum amount of data can be extracted from each image and 2)

images to be indexed together are at about the same image brightness level. In a histogram

stretching process the maximum (i,,,) and minimum (im,,) pixel value of each image is found.

Then each pixel value is multiplied by an enhancement multiplier, M, defined in equation 4-1.

The result is rounded to the nearest whole number while remaining in the range of 0 to 255.


M 255 (4-1)
I -I
max mm

Where, imax, imm, = maximum and minimum pixel value, respectively.

This histogram stretching allows overly dark or bright images to use a wider spread of

pixel values enhancing the quality of the image as displayed in Figure 4-7. This same process is

used again at a later step of the Image Indexing stage.

The second step of the Image Indexing stage was the use of an image smoothing filter.

Smoothing filters, also known as image smoothing functions, can be very complex and powerful

or quite simple. The resulting smoothed image requirements decide what type of smoothing

filter should be used (Gonzalez et al., 2004). Because clear grove images are desirable in this

research, a small 3 x 3 pixel smoothing filter using an averaging function was used. The filter is

defined mathematically by equation 4-2.

,,j = (P, 1,J P, + j I p,, ,, + 3p,,J + 3p ,,J+ + 3p, ,, +3p,+,j + 8p,,)/24 (4-2)












Where, p,j = raw image pixel (i, j)


P,, = resulting image pixel (i, j)



4000
3500
3000
2500
2000
1500
1000
500
0
0 50 100 150 200 250
A



4000
3500
3000
2500
2000
1500
1000
500


0 50 100 150 200 250
B


Figure 4-7. Histogram stretching function; raw image and histogram A) Before histogram
stretching. B) After histogram stretching.


The smoothing filter performed three tasks for improving the image to image calculations:


1) removed random noise in the images; 2) allowed image target edges to smooth the transition


between the background and foreground, helping resolve situations where targets did not align


perfectly between raw images; 3) extracted information from saturated pixels (i.e., a pixel of 255


located next to pixels of less than 255 was not considered as bright/saturated as a pixel totally


surrounded by other saturated pixels of value 255).


The critical image processing step of the Image Indexing stage was the actual index itself.


To identify the citrus fruit, an index calculation using all three band pass images was designed.









Indices in remote sensing applications are the mathematical calculations of different spectral

band images. Most often in agricultural applications these calculations are preformed with high

spatial resolution images acquired by aerial or satellite based systems. The most widely used

index is the NDVI or Normalized Difference Vegetation Index (Morgan and Ess, 2003). The use

of diffuse reflectance information from Chapter 3 was used as a guide, but image investigations

revealed that in-field lighting and target orientation conditions created very different reflectance

values than had been expected. Many index calculations were attempted without quality outputs.

Examples include ratio indices (such as the 1064 band pass divided by the 1572 band pass; both

normalized and un-normalized), indexing by way of finding class likelihood using Fisher linear

discriminant analysis (FLDA) as described in Chapter 3, and even more complex hybrid indices

(using multiple indices while solving for the average). When studying the common and reliable

spectral indices used in agricultural application the most widely used indices appear very simple

(Apan et al., 2003). Further study of the literature inspired the design of the index used in this

thesis research. The equation for the index is shown in equation 4-3.

B = (b064 / b1150) (b1572 / b1150) (4-3)

Where, b1064 = images with 1064 nm band pass filter

b150 = images with 1150 nm band pass filter

b1572 = images with 1572 nm band pass filter

This green citrus index functions by solving the differences between the reflectance of

leaves and citrus fruit at 1064 nm and 1572 nm, but first the images are normalized by the 1150

nm band where the leaves and citrus fruit reflectance are very similar. This normalization

eliminates the effects of dark images of one band being indexed with light images of the other

band. This index does not represent the ideal segmentation system that was calculated in









Chapter 3 but still uses an interpretation of it. The higher reflectance of the citrus fruit versus

citrus leaves around 800-1100 nm is directly compared against the lower reflectance of the citrus

fruit versus citrus leaves around 1500-1800 nm. Two critical band pass filters are used, as was

desired and a third band pass filter normalizes the image shadows and bright reflections.

The resulting two-dimensional image from the green citrus index can be understood as a

pixel likelihood for each class, where high values are citrus fruit and low values are non-citrus

fruit. One problem remaining in this likelihood image is that misaligned images can result in

pixel calculations being too high; therefore an upper value cut-off was set to 165. This value

selection was based on training image histograms. The pixels that exceeded this upper value of

165 were by default set to zero, therefore eliminating them from being classified as citrus fruit

pixels. A fine example of this can be seen in the index images (Figure 4-8B and Figure 4-9B),

where the brightest spots are not the fruit but the edges of objects which are not properly aligned.

The histogram stretching function was used again after this upper value cut off to maximize the

pixel value separation for the next threshold step utilizing Otsu's method.

Otsu's method automatically chooses a threshold based on the histogram of a grayscale

image (Otsu, 1979). The method is based on discriminant analysis, as discussed in Chapter 3.

The objective of such an analysis is to separate each pixel into one of two classes, Co or C1,

based on a threshold value, t. Assume Co represents lower pixel values {0, 1, 2, 3, ..., t} and C1

represents higher pixel values, {t+1, t+2, t+3, ..., L-1}, where L is the number of grayscale

values. Let C2b and G2d be the between-class variance and within-class variance, respectively.

The ideal separation threshold can then be obtained by maximizing the separation of the

histogram. The objective is to maximize the ratio of the between-class to within-class with

respect to the threshold level, t. This is found by stepping through all intensity levels of the









grayscale image and at each level calculate o2b/C2d. The higher this ratio is, the stronger the pixel

class separation. MATLAB 7.0 has an Otsu function built into the image processing toolbox

(MATLAB, 2006). The function returns only the recommended threshold value and has no

effect on the original grayscale image. Otsu's value can be multiplied by a scaling factor

lowering or raising the threshold and provided an opportunity to fine tune the method for this

citrus image application. The training images showed better classification results with a

multiplier value of 1.5. This is due to the citrus images needing a threshold for only the highest

valued pixels, as most pixels were non-citrus fruit pixels.

The final step of the Image Indexing stage was to improve the binary image, removing

holes and noise using a ten pixel sized 'disk' for erosion then dilation, also referred to as opening

(Gonzalez et al., 2004). The selection size often pixels was a choice based on the training

images. It is feasible to use a larger disk shape and this would improve some results by clearing

out medium sized errors, however some images would have all the possible fruit pixels

eliminated as well. Again, the MATLAB image processing toolbox has a function for this

morphological process (MATLAB, 2006). The complete step by step process of the Image

Indexing stage is shown in Figures 4-8 and 4-9.

Marker Checking stage

Examination of the final images reveal several extra large and small fruit markers, which

was defined as groups of possible citrus fruit pixels in the images. If each marker signals a citrus

fruit location in the image, then most images have too many such markers. The Marker Check

stage is a means of going back to the 1150 nm band pass image and checking if the marker in the

final image makes sense. This is accomplished by treating each marker as a separate feature and

checking it individually based on the variance of the pixels in the marker with respect to the

number of pixels making up the fruit marker. The 1150 nm band pass image is not the raw

















A












B C












D E
Figure 4-8. Image Indexing stage of validation image number eight. A) Three band pass images
1064 nm, 1150 nm and 1572 nm, respectively. B) After smoothing, histogram stretch
and the index calculation steps. C) After upper limit cut off and histogram stretching.
D) After thresholding using Otsu's method. E) Final binary image after the use of a
10 pixel erosion and dilation 'disk'.

















A












B C












D E
Figure 4-9. Image Indexing stage of training image number 11. A) Three band pass images
1064 nm, 1150 nm and 1572 nm, respectively. B) After smoothing, histogram stretch
and the index calculation steps. C) After upper limit cut off and histogram stretching.
D) After thresholding using Otsu's method. E) Final binary image after the use of a
10 pixel erosion and dilation 'disk'.









image, but saved after being histogram stretched and smoothed in the Image Indexing stage

before being used again in this stage. This prevents noise from negatively affecting the variance

calculation results.

The first step of the Marker Checking stage consisted of a simple separation of the markers

based on connectivity. If a non-breaking line could be drawn around the marker and connected

to the starting point of the line without crossing over another marker or being inside another

marker, the enclosed marker would be identified. The next step calculated the distance between

the highest and lowest pixel values of the marker and divided it by the total number of pixels, n,

in the marker. The result is defined as the marker value for that marker, defined in equation 4-4.

MV = (Pmax -Pmn ) / n (4-4)

Where, MV = marker value

Pmax, Pmin = maximum and minimum pixel values, respectively

n = number of marker pixels

Markers with an MVvalue greater than 0.2 were assumed to be non-citrus fruit markers

and were eliminated from consideration. The MVvalue used in this step was chosen based on

the trial and error and best results from the training images. It can be interpreted that for every

100 pixels of the marker, a 20 pixel level change was allowed without the marker being

removed. The MVused was for image processing can be considered conservative, as it permitted

some small non-fruit marker errors to be kept rather than risk removing a large correct citrus fruit

marker. This major stage also filled any holes in the citrus fruit markers, such as the one seen in

Figure 4-10. This MV concept was based on the major assumption that leaves, branches, Teflon

sheet edges, and other non-citrus fruit objects are more likely to have edges appear in the

markers. These edges will have a higher variance in pixel values ranging from the shadow to a









bright glare within a short range. Citrus fruit objects on the other hand tended to have a

smoother and less severe rapid change in pixel value. Additionally, because of the method in


C D

Figure 4-10. Marker Checking stage of validation image number nine. Input binary image of
possible fruit markers (A). The 1150 nm band pass image (B). Markers and the pixel
value intensities (C). The new binary image output after removal of most non-fruit
markers and filling of marker hole (D). Of the five output fruit markers four are
correctly identified as citrus fruit.

which the MVis calculated, only smaller markers are at real risk of being eliminated. This

means correctly identified citrus markers that include both the dark shadowy edge and the bright

sunny spot of the citrus fruit would not be removed as long as the marker size is sufficiently

large. On the other hand, small markers are at great risk of being removed and sometimes even

correct markers were removed, as in the Figure 4-10. The small marker on the bottom left is

incorrectly removed by the Marker Checking scheme, however, an incorrect marker was










removed as well. This type of scheme balances the pros and cons to make images

improvements, and for most of the training images the positives outweigh the negatives.

The complete image processing routine involved two major stages, ten sub-steps, and three

raw image inputs to produce one final output image. This included variables that were fine tuned

by the training image data set. Variables included: smoothing filter type, Otsu's method

multiplier of 1.5, upper value cut off of 165, dilatation and erosion 'disk' size, andMVlimit of

0.2. Slight changes to any one these variables would result in different binary image results.

Figure 4-11 is a flow diagram of the complete image processing major stages and sub-steps.


Image Indexing stage

histogram smoothing green citrus upper value histogram
stretch E= filter index cut off stretching
r-------------------------------i
r------------------------------ i

marker value marker I erosion and threshold
calculation (MV) identification 1 dilation (Otsu's method)

--------------------------------
J Marker Checking
stage
marker cleanup
and removal
L____________________--------------------------------------------J
Figure 4-11. Complete image processing procedure.

Result analysis

Of the 36 in-field images, two-thirds were randomly selected and used as training data,

while one-third was used as validation (i.e., 24 training and 12 validation images). Training

images were used in this study to permit the testing of image processing techniques on some

images while not being the same images final results would be taken from. This prohibits the









image processing steps from being overly designed and conditioned to the validation images. At

the conclusion of the image processing steps each validation image was checked for three

measures of accuracy; percent of correct pixels, fruit pixel counts and number of markers.

Percent of correct pixels and the fruit pixel counts were measured against manually produced

image masks of each validation image showing the citrus pixels (Figure 4-12). The measure of

fruit markers is compared to the number of fruit seen by human observation in the images and

the number of fruit markers in the manually produced image masks.













A B
Figure 4-12. Manually produced image mask showing the difference between citrus fruit pixels
and non-citrus fruit pixels.

Percent of correct pixels were measured on how many pixels of each image were correctly

classified and how many were incorrectly classified. The resulting classifications were judged

using Bayes error rate, defined in equation 4-5 (Duda et al., 1988). There are two possible

misclassifications: either a pixel x falls in R2 (non-citrus fruit) while its true state of nature is wl

(citrus fruit) or the pixel x falls in R1 (citrus fruit) while its true state of nature is w2 (non-citrus

fruit). Otherwise, the classification is defined as correct.

P(correct) = 1 P(error) = 1 [P(x e R2, w) + P(x e R1, w2)] (4-5)

Where, R1, R2 = regions 1 (citrus fruit) and 2 (non-citrus fruit)

wl, W2 = classes 1 (citrus) and 2 (non-citrus fruit)









The measure from Bayes error rate (equation 4-5) is out of a probability factor, meaning a

perfect 100% rate would be 1.0, while a 50% rate is valued 0.5. In the results section, these unit

rates will be discussed in percent correct and percent error.

Image processed fruit pixel counts were measured with respect to the manually masked

fruit pixel counts. For each of the 12 validation images used in this study a count of citrus fruit

pixels was made from the resulting processed image and the manually masked image. An R2

value was calculated for a linear approximation relating the two measures. This is a very

important measure as it provides a quantitative value of the ability to predict the density of green

citrus in each image using this method. This is a methodology future researchers could use in

determining green citrus yield.

Lastly, a survey was conducted regarding the number of citrus fruit markers found,

manually counted number of citrus fruit, and the masked image citrus fruit markers. This is

important as it has been the measuring tool of previous research for on-tree citrus fruit

identification (Annamalai et al., 2004; Chinchuluun and Lee, 2006). The results conclude with a

discussion about the value of such marker count information.

Results and Discussion

Image Acquisition Problems

There were many problem that appeared early during the in-field image acquisition

process, but there was very little that could be done at the time. These problems and their effects

on the results will be discussed before the qualitative results section. The most important issues

faced with this experimental design were target shifting, light/shadow changes, raw image

saturation and multiple leaf reflections increasing the expected diffuse reflectance.

During the time it took to change out one optical band pass filter for another (30-45

seconds) slight changes could occur in the uncontrolled environment. The two most prevalent









changes were what is referred to as target shifting and light/shadow alterations. While one image

may be taken under ideal conditions, subsequent non-ideal conditions could follow closely. This

includes the biggest in-field problem of strong breezes shifting the leaves and/or branches and

even swinging the target fruit(s) between image acquisitions. Without the fruit being perfectly

aligned from one band pass image to another, incorrect edge classification will occur. Figure 4-

13 is an example of a target fruit and surrounding branches and leaves shifting as a result of wind

turbulence. The two images were taken with the same band pass filter only two or three seconds

apart, however, the location of the fruit and leaves shifted dramatically. This problem can not be

resolved by a simple shifting of the image to align the fruit, because not all the objects are shifted

by the same amount. For example, the leaves just above the Teflon sheet in Figure 4-13 are at

approximately the same location in both images. These means a dynamic shifting would need to

be used, stretching and compressing different parts of the images to better align them. Such a

complex process would become increasingly difficult on such a complex background.















Figure 4-13. Location of the target fruit and some of the leaves has shifted between the two raw
images. The Teflon sheet and some leaves remain in the same place.

In addition to wind moving the target and the resulting shadows, shadow and sun

illumination changed during image acquisitions due to inconsistent cloud cover and the slow

shifting of the sun. These shifts in the sun's location in the sky creates more than a change in









shadow location. The shift changes the solar angle which may affect the magnitude of resulting

fruit and leaf reflectance (Gilabert and Melia, 1993). Because raw images at each target location

were taken within a short period of time these solar angle effects were not considered in

individual images blocks, however it may be a concern if future systems require images to be

acquired all day. The only solution used for this research during image acquisition was to take

many images quickly and select the images that facilitated better image to image comparisons

based on target fruit and/or leaf alignment.

Another major problem during image acquisition was raw image saturation. Many of the

images became too dark or too bright. This was most prevalent on the outer and inner layers of

the citrus canopy where lighting conditions could vary dramatically. Figure 4-14 shows one

example of the light saturation among the leaves and part of the citrus fruit. This was a critical

issue as information at these pixels was lost and comparisons to the reflective behaviors of the

other wavebands would not be correct. The Teflon sheet in several of the images became the

most susceptible to this problem. This is why the use of the Teflon sheet was suspended when it

became apparent that the material was too reflective to be used as a normalization material in

sunlight.















Figure 4-14. Lighting condition extremes (blackout and saturation) from inside and outside the
citrus canopy.









The effects of the saturated pixels are dramatically observed in the histograms; Figure 4-15

shows two images of the same targets using the same band pass filter. They were taken moments

apart when the strength of sunlight changed dramatically due to sudden cloud cover. The

histogram on the left has around 30000 pixels saturated with a value of 255. The histogram on

the right has only about 700 pixels saturated. Comparing these values to the total number of

pixels per image (307200 pixels), a loss of almost 10% of the pixels is experienced by the image

on the left as apposed to less than one quarter of one percent on the right. During the image

acquisition process the mechanical iris of the NIR camera was not precise enough to correct all

lighting problems.


OwU


r 1



n


n


x 103500
35 z 30000 pixels 3500
3000
3/ \
25 250:
I I
2 !200:'
15 150.
1I 100:
05 I 50o:

0 50 100 150 200 \ 250 0 50


Figure 4-15. Changes in sunlight and the effect on image histograms.


:: ("0 pixels

/ \
I \


100 150 200 \250 /









The last issue faced in the experimental setup was foreshadowed in an experiment

discussed in Chapter 3. The diffuse reflectance of multiple leaves stacked together were

measured with a spectrophotometer (Cary 500, Varian, Inc., Palo Alto, CA) and revealed as the

number of leaves increased so did the percent of diffuse reflectance. It was theorized that

additional leaves became a second, or third, reflective surface for light waves that passed through

the first leaf. After reflecting off the second (or third) leaf, the light would transmit back through

the first leaf again, and be detected as light reflected from the first leaf. The acquired NIR

camera images show a similar phenomenon as the 1064 nm band images showed only a minor

intensity difference between fruit and leaves. In addition, leaves in the NIR camera images were

not facing flat but had more variations in direction, which changed the reflectance among leaves

and altered the expected reflectance from Chapter 3.

Quantitative Identification

Using the image processing methods and variable values described previously, the average

correct pixel classification was 90.3% with respect to the manually masked images. Correct

pixels were any pixels that had the same classification in the masked image and image processed

image. The total number of correct pixels was then divided by the total number of pixels,

307200, to find the percent correct. This is presented in Table 4-2, with a fruit pixel count from

the image processing and the masked images. Notice, being able to predict the pixels

classification may not the best measure of identification. This will be discussed in more depth in

a little bit.

The image processing citrus pixel counts and manually masked citrus pixel counts are also

presented in Figure 4-17. Each show the complete number of fruit pixels counted in the results.

This count and be interpreted as a type of yield data set for each image. While the sizes of the

fruits in each image are not always the same and the distances of the camera to the targets are









variable, the total number of citrus fruit pixels does provide some information. The ability of the

image processing algorithm to detect this can be studied using a t-test and solving an R2 for the

two data vectors.

Table 4-2. Validation image citrus fruit pixel results.
Validation image Correct pixel Image processing Manually
number classification (%) fruit pixel count masked fruit
pixel count
1 94.8 1524 16624
2 96.9 9753 18716
3 77.5 35756 33861
4 96.1 1699 13084
5 81.9 23507 63607
6 95.1 5936 17109
7 82.3 4029 36512
8 92.9 11065 31841
9 89.5 11001 38565
10 97.9 1207 2733
11 88.1 12743 36706
12 90.2 9550 31901
average 90.3 9812.3 27538.3

Using a two sided t-test to assess the significance of the error percent when comparing

image processed fruit counts with manually masked fruit counts, in this case a highly significant

difference was found. The p-value was solved to be 4.295x10-4 making it far too small, which

shows the "null hypothesis" to be invalid for this data set. In effect, the t-test results are saying

the image processing method was unsuccessful, but only if it is assumed that the manual masking

method is defined as a correct reference. Again, the validity and limited usefulness of the

manually masked images for checking will be discussed in more depth later.

The two sided t-test method also tested for the equality of variances in the two data sets.

The probability was solved as being greater than 0.10, meaning the t-test was valid for the "null









hypothesis" of equal variances. The calculated variance for the image processing vector is

1.0190x108 while being 2.5317x108 for the manually masked images. In addition, the t-test

revealed the t-statistic is 4.9590 on 11 degrees of freedom.

Using the image processing techniques and variables discussed previously, the image

processing results underestimated the citrus pixel counts for all validation cases. This makes

sense when observing the image results in Figures 4-19, 4-20, and 4-21, but can be proven by

looking at the data presented in Figure 4-16. A list of four classes based on image processing

pixel classification and manually masked pixel classification is given for all 12 validation image.

The image processing pixel classification (vertical location) and the manually masked pixel

classification (horizontal location) define which corer the pixel count adds to. This means the

top left corner are fruit pixels correctly classified as fruit pixels, bottom left corner are non-fruit

pixels incorrectly classified as fruit pixels, top right corner are fruit pixels incorrectly classified

as non-fruit pixels, and bottom left corner are non-fruit pixels correctly classified as non-fruit

pixels. A total for all the validation images is given at the bottom. There are two important

observations to pull from this figure.

First, the great numbers of pixels are classified by the image processing as non-citrus fruit

and are manually masked as non-citrus fruit. This is seen by how much larger the value in the

non-citrus/non-citrus box is in the bottom right corner. This facilitates for a high percent correct,

as in Table 4-2, but does not display true identification.

The second important observation in Figure 4-16, is the relative values of image

processing classified non-fruit that are masked as fruit (bottom left corner), versus the image

processing classified fruit that are masked as fruit (top left corner). The lower the bottom left

corner value relative to the top left corer, the more correct the identification without improper











classification. The best examples of this are validation image numbers one, two, four, seven,


nine, eleven and twelve. All these validation images are examples of the image processing


finding the citrus fruits but did not completely covering the masked area. For a good example of


this, refer to Figure 4-20. The shadowed edges and bright glare of the sun eliminated parts of the


fruit from being classified correctly. The image processing algorithm is conservative with the


selection of citrus fruit pixels as a way to safe guard against classifying non-citrus fruit pixels as


citrus fruit pixels.


Image 1
f n
i L -'D0 I -'- L 'T
I.
t .: '-,


Image 4

f n





f n
': -."t





Image 1
f n

Image 10
f n





Total of allvalidation images
f n
-.' -' r'^- i -
-I;_u JilJ ^


Image 2
f
i 1, 6.u




Image 5
f


66': 7
a

Image 8
f





Image 11
f


i0
B,


Image 3
f





Image 6
f





Image 9
f


Z I.,


Image 12
f
c "G


Figure 4-16. Pixel classification results for all validation images and the overall total. (Notice,
top left corer are fruit pixels correctly classified as fruit pixels, bottom left corner are
non-fruit pixels incorrectly classified as fruit pixels, top right corer are fruit pixels
incorrectly classified as non-fruit pixels, and bottom left corner are non-fruit pixels
correctly classified as non-fruit pixels.)


n


_ :::4::-l



n






n




:"6 4V 1


n.
"'" '.''* "'




n
I 4:








n
I-'J'I ":





n
:':'-l'.ii

..r'I. ,I.I.









The R2 of all the fruit pixel counts is 0.390 (Figure 4-17). However, it is evident from the

graph that the point (35756, 33861) is an outlier. This even more evident from the poor

performance of this validation image number three, which is presented in detail in Figure 4-21.

If this validation image is dropped, a fruit pixel to fruit pixel R2 of 0.765 was calculated (Figure

4-18). Also the narrowing of the 90% confidence prediction bounds displays this same

significance of the outliner removal. The R2 results would be higher if the calculations come

from non-citrus pixel counts as most of the images were of non-citrus pixels.

A major question that this research has shown is that of, what is identification? Is the best

measure of identification the correct number of pixel counts or the number of markers (or fruit)

counted? The problem with counting pixels that are correct or incorrect lies in an example where

no citrus fruit pixels are identified by the image processing system. If image processing result

predicted zero citrus pixels, the answer would still be over 50% correct for these images used as

no image in this study had citrus pixels covering more than half of the image area. But there are

also problem with the use of citrus fruit markers to determine the number of citrus fruit, which

has been the standard method in previous citrus identification schemes (Annamalai et al., 2004;

Chinchuluun and Lee, 2006). While justification of this counting method can be made due to the

distance of the camera from the canopy and the resulting number of citrus per image (10 to 50),

it should be noted that a citrus pixel marker may not be accurate for other conditions. Close

images of a citrus canopy generally lead to problems of over or under estimation of fruit. In the

first example, the view of one fruit can be obstructed by a branch. The image processing might

work properly and identify the two visible sides independently, thus two markers for one fruit.

The second example is a great number of fruits collected together and touching while on the tree.

In this case, only one marker is created for two or more citrus fruits. These are pertinent












concerns for a relatively close camera-to-canopy experiment when only a few fruits are visible in


each image.


S104
8 ----- i ------ i----- -----------


7


6
0

-g
o



-0
S4

E
- 3


C2


1


0 00


0


O
0
0 0.5 1 1.5 2 2.5 3 3.5 4
Image processing pixel count x lo4

Figure 4-17. Manually masked fruit pixel count versus image processing fruit pixel count results
with linear best fit using all pixel count data.


x 104


0 Pixel counts
Linear fit
(R-squared = 0.765)
Prediction bounds
(90% confidence)


0 0.5 1 1.5 2 2.5
Image processing pixel count x 104
Figure 4-18. Manually masked fruit pixel count versus image processing fruit pixel count results
with linear best fit with one outlier removed.


0 Pixel counts
Linear fit
(R-squared = 0.390)
Prediction bounds
(90% confidence)


7


I6


0

4
0)

- 3

7-









The results from the validation image fruit markers and human fruit counts are listed in

Table 4-3. Fruit marker results reveal a stronger correlation between the image processing fruit

markers and the actual human observed fruit count. This is primarily due to the second example

discussed above; affecting validation images number five and nine the most. There is one case

of example one in validation image number four. The image processing fruit marker counts do

follow the human observation, but this might only be a result of noisy outputs at coincidentally

correct image numbers. There are cases when the extra fruit markers harm the results, such as

validation images six and seven.

Table 4-3. Validation image fruit marker and manually masked marker results.
Validation image Image processing Manually Human observed
number fruit marker masked marker fruit count
count count
1 1 1 1
2 1 1 1
3 5 3 4
4 1 2 1
5 8 3 9
6 4 1 1
7 4 2 2
8 1 1 2
9 5 4 7
10 2 1 1
11 3 1 1
12 4 2 5

Three complete image processing and resulting classifications and misclassifications

images are given (Figures 4-19, 4-20, 4-21). These three image results represent typical results

in both the training and validation sets. Figure 4-19 shows a very complex grouping of citrus

fruit with only shadows in the background. The raw images of the 1064 nm and 1150 nm band

pass filters suffer from glare on the right hand side. This creates the incorrect fruit markers on









the right hand side. Six of the nine fruit are marked, with three unmarked, and two image

processing fruit markers being incorrect. It is difficult for multiple green citrus fruit to be

properly identified together as they never have the same light illumination. This creates unequal

ratios between the two key band pass filters, 1064 nm and 1572 nm, which always leave some

fruit out or too many non-citrus fruit inside the thresholding boundary.

Figure 4-20 shows a very common single fruit marker identification. In this case the

Marker Checking stage is not even necessary. What is interesting about this image is the bright

sun spot on the citrus fruit because the loss of data due to saturation at this point prohibits the

region from being properly classified as fruit. This "C" shape is quite common and can be

closed up if desired. Such an algorithm was not included in image processing because

occasionally, as was observed in the training data, a very large "C" can occur when multiple fruit

are touching. In that case, closing up the "C" is not desirable. Because of the commonality of

this type of image result, the average number of image processed fruit pixels is about one-third

that of the masked fruit pixels. This same pattern can be observed in Figure 4-16 with validation

image numbers one, two, four, seven, nine, eleven and twelve.

Finally, Figure 4-21 reveals a case of extreme incorrect fruit identification. This is not a

result of the image processing algorithm, but the poor quality of images acquired. In this case

the band pass at 1572 nm is not dark enough. This makes the raw images of the three band

passes too similar and no multispectral information can be obtained. Instead, the green citrus

indexing reveals a sight ringing effect which is the result of placing optical band pass filters in

front of the focal lens of the NIR camera.

Conclusions

This Chapter tested a ground based multispectral image acquisition and image processing

system for identifying green citrus fruit against a green citrus leaf canopy. The image acquisition
















A










B C










D E










F G
Figure 4-19. Complete image processing and results check of validation image number five. A)
Three band pass images 1064 nm, 1150 nm and 1572 nm, respectively. B) Green
citrus index, upper limit cut off and histogram stretch. C) Thresholding using Otsu's
method. D) Possible fruit markers and the pixel value intensities. E) Marker
Checking output binary image. F) Manually produced mask of citrus locations. G)
Pixel classifications and misclassifications. Notice, six of the eight fruit are marked,
with three unmarked, and two image processing fruit mark being wrong.



















A











B C











D E
Figure 4-20. Complete image processing and results check of validation image number two. A)
Three band pass images 1064 nm, 1150 nm and 1572 nm, respectively. B) Green
citrus index, upper limit cut off and histogram stretch. C) Thresholding using Otsu's
method. D) Marker Checking output binary image. G) Pixel classifications and
misclassifications.



























B C










D E











F G

Figure 4-21. Complete image processing and results check of validation image number three.
A) Three band pass images 1064 nm, 1150 nm and 1572 nm, respectively. B) Green
citrus index, upper limit cut off and histogram stretch. C) Thresholding using Otsu's
method. D) Possible fruit markers and the pixel value intensities. E) Marker
Checking output binary image. F) Manually produced mask of citrus locations. G)
Pixel classifications and misclassifications.









was completed in a Florida citrus grove. A highly sensitive NIR camera was used in conjunction

with three optical band pass filters (1064 nm, 1150 nm, and 1572 nm). The resulting images

were separated into 24 training images and 12 validation images. A complex image processing

scheme was designed involving indexing of the band pass filter images. Measurements of

correctness were based on fruit pixel identifications with respect to a manually produced fruit

pixel mask.

Quantitative results showed that correct pixel class identification of the 12 validation

images was 90.3% when using the image processing algorithm and variable described. When

comparing citrus pixel count from the image processing algorithm to the manually masked citrus

pixel count, an R2 of 0.764 was achieved with the removal of one outlier. There were a great

number of problems during image acquisition that hampered the research in later stages of data

analysis. These problems included target shifting, lighting conditions, cloud cover, wind, band

pass filter removal, band pass filter alignment, and user mistakes. All of these problems,

however, can be summarized as a result of an uncontrollable environment. Despite these

problems, the results do prove the effectiveness of the concepts presented in this study. Green

citrus fruit can be identified based on spectral differences from leaves by an NIR camera.

However, there are still many refinements that require further investigation.

The most important knowledge gathered in this study was a first hand account of technical

issues faced and how they might be overcome in future systems. The most import requirements

for improved results include the system being capable of acquiring multiple waveband images

simultaneously. This would resolve almost all the environmental issues mentioned previously.

A second problem researchers would need to solve is lighting conditions. The sun is not a









reliable and consistent light source. It moves through the day, changes intensity based on cloud

cover, and cannot penetrate deep into the citrus canopy.

This research goal of this Chapter was to design and test a non-destructive NIR camera

based computer vision system for identifying and counting green citrus yield in-field while still

on the tree. Results have proved this is possible, but still difficult due to environment conditions

and technical difficulties. The last Chapter summarizes this thesis research and discusses what it

means for NIR sensing systems in the future.









CHAPTER 5
CONCLUSIONS AND FUTURE WORK

Conclusions of Research Objectives

Near infrared (NIR) sensing technology is a promising, rapidly affordable, and available

tool for precision agriculture systems. The non-destructive nature of spectral-based sensing is

welcomed by growers who would rather not sacrifice their products. This research has shown

that NIR sensors can be used to separate green citrus fruit and green citrus leaves by means of

spectral reflectance in the NIR range. This research has proven through the use of training and

validation sample sets that perfect separation is feasible using only two wavelengths features

(881 nm and 1381 nm). These results were found through the use of the Fisher linear

discriminant analysis (FLDA) algorithm for multi-dimensionality breakdown.

This research has also proven that correct identification of citrus fruit is possible using NIR

optical equipment. In this research an NIR monochromatic camera was outfitted with three band

pass filters (1064 nm, 1150 nm, and 1572 nm). Image processing and multispectral indexing was

used to classify the image contents pixel by pixel. By using 24 training images an image

processing algorithm was designed and tested. Twelve validation image results showed a 90.3%

correct pixel classification (citrus fruit or non-citrus fruit) obtained. In addition, the image

processing scheme provided an R2 of 0.746 for fruit pixel counting versus a manually masked

fruit pixel count with one outlier data being disposed. Despite these positive accomplishments,

this research has shown that better image processing methodologies need to be explored,

additional feature NIR spaces should be considered, and the high cost of NIR camera and optical

equipment needs to decline for this technology to thrive. Also uncontrollable environmental

issues need to be thought through and planned for to improve upon these results.









Future Work

Highly accurate spectroscopic systems are not currently easy to transport or use in

demanding environmental conditions, such as a Florida citrus grove. When smaller and cheaper

in-field spectrophotometer systems become more available, the development of nutrient variation

maps based on both leaf and fruit samples will be possible. This increase of grove information

will provide growers more opportunities to improve their management techniques on grove site-

specific basis.

Multispectral NIR cameras can be used for the identification of green citrus fruit in a citrus

grove. It is not hard to predict that future research will be more accurate and capable of counting

fruit while on the move. Combinations of this system with GPS receivers could create on-the-go

citrus yield maps early in the season. This dream system is still a long way off in the future, but

this research opens the door to possible uses of remote NIR sensor systems in the Florida citrus

industry. It is conceivable that in the future, autonomous VIS-NIR remote sensor guided

vehicles could be used to find, count, map, and even test citrus fruit for health and nutrient

content, as some research suggests, while in the grove. Such a futuristic system is the ultimate

goal of precision agriculture research not only in the Florida citrus industry but all modem

agricultural industries.









LIST OF REFERENCES


Alchanatis, V., Cohen, Y., Cohen, S., Moller, M., Meron, M., Tsipris, J., Orlov, V., Naor,
A., Charit, Z., 2006. Fusion of IR and multispectral images in the visible range for empirical and
model based mapping of crop water status. ASABE Meeting Paper No. 061171. ASABE, St.
Joseph, MI.

Aleixos, N., Blasco, J., Navarron, F., Molto, E., 2002. Multispectral inspection of citrus in
real-time using machine vision and digital signal processors. Computers and Electronics in
Agriculture 33, 121-137.

Annamalai, P., Lee, W.S., 2004. Identification of green citrus fruits using spectral
characteristics. ASAE Meeting Paper No. FL041001. ASAE St. Joseph, MI.

Annamalai, P., Lee, W.S., Burks, T.F., 2004. Color vision system for estimating citrus
yield in real-time. ASAE/CSAE Meeting Paper No. 043054. ASAE St. Joseph, MI.

Apan, A., Held, A., Phinn, Markley, J., 2003. Formulation and assessment of narrow-band
vegetation indices from EO-1 Hyperion Imagery for discriminating sugarcane disease. Proc
Spatial Sciences Conf., 1-13.

Chinchuluun, R., Lee, W.S., 2006. Citrus yield mapping system in natural outdoor scenes
using the watershed transform. ASABE Paper No. 063010. ASABE, St. Joseph, MI.

Chinchuluun, R., Lee, W.S., Ehsani, R., 2007. Citrus yield mapping system on a canopy
shake and catch harvester. ASABE Meeting Paper No. 073050. ASABE, St. Joseph, MI.

Citrus Research and Education Center (CREC), 2007. Precision agriculture history in
Florida. Available at:
http://www.crec.ifas.ufl.edu/crecwebsites/precision_agriculture/history.htm. Accessed: March
25, 2007.

Ding, P., 2005. Use of Nondestructive Spectroscopy to Assess Chlorophyll and Nitrogen in
Fresh Leaves. Unpublished Ph.D. Dissertation. Oregon State University, OR.

Duda, R.O., Hart, P.E., Stork, D.G., 1988. Pattern Classification, 2nd ed. John Wiley and
Sons, New York.

Fisher, R.A., 1936. The use of multiple measurements in taxonomic problems. Annels of
Eugenics 7, 79-88.

Florida Agricultural Statistics Service (FASS), 2005. Citrus 2004-05 Summary. Available
at: http://www.nass.usda.gov/fl/rtocOci.htm. Accessed: May 3, 2006.

Fraser, D.G., Jordan, R.B., Kunnemeyer, R., McGlone, V. A., 2002. Light distribution
inside mandarin fruit during internal quality assessment by NIR spectroscopy. Postharvest
Biology and Technology 27, 185-196.









Gaffney, J.J. 1972. Reflectance properties of citrus fruits. Transaction of the ASAE. 15,
310-314.

Gilabert, M., Melia, J., 1993. Solar angle and sky effects on ground reflectance
measurements in a citrus canopy. Remote Sensing Environment 45, 281-293.

Gonzalez, R.C., Woods, R.E., Eddins, S.L., 2004. Digital Image Processing Using
MATLAB. Pearson Prentice Hall, Upper Saddle River, NJ.

Grift, T., Ehsani, R., Nishiwaki, K., Crespi, C., Min, M., 2006. Development of a yield
monitor for citrus fruits. ASABE Meeting Paper No. 061192. ASABE, St. Joseph, MI.

Gumz, M., Weller, S.C., 2005. Using remote sensing to differentiate weeds in mint. Top
Farmer Crop Workshop Newsletter, April.

Jimenez, A.R., Ceres, R., Pons, J.L., 2000. A survey of computer vision methods for
locating fruit on trees. Transaction of the ASAE 43 (6), 1911-1920.

Labsphere, 2006. A Guide to Integrating Sphere Theory and Applications. Labsphere Inc.,
North Sutton, NH. Available at:
http://www.labsphere.com/data/userFiles/A%20Guide%20to%20Integrating%20Sphere%20The
ory&Apps.pdf Accessed: Nov 5, 2006.

Lee, W.S., Slaughter, D.C., 2004. Recognition of partially occluded plant leaves using a
modified watershed algorithm. Transaction of the ASAE 47 (4), 1269-1280.

MacArthur, D.K., Schueller, J.K., Lee, W.S., Crane, C.D., MacArthur, E.Z., and Parson,
L.R., 2006. Remotely-piloted helicopter citrus yield map estimation. ASABE Meeting Paper No.
063096. ASABE, St. Joseph, MI.

MATLAB, 2006. Learning MATLAB 7 Release 14. MathWorks Inc., Natick, MA.

Merzlyak, M.N., Gitelson, A.A., Chivkunova, O.B., Rakitin, V.Y., 1999. Non-destructive
optical detection of pigment changes during leaf senescence and fruit ripening. Physiologia
Plantarum 106, 135-141.

Min, M. 2006. Spectral-Based Nitrogen Sensing for Citrus. Unpublished Ph.D.
Dissertation. University of Florida, FL.

Morgan, M., Ess, D., 2003. The Precision-Farming Guide for Agriculturists. John Deere
Publishing, Moline, IL.

Nagy, S., Shaw, P.E., Veldhuis, M.K., 1977. Citrus Science and Technology. Vol. 1. Avi
Publishing Co., CT.

Otsu, N., 1979. A threshold selection method from gray-level histograms. IEEE
Transactions on Systems, Man and Cybernetics 9 (1), 62-66.









Regunathan, M., Lee, W.S., Annamalai, P., 2005. Green citrus fruit identification using
spectral characteristics. 5ECPA-2ECPLF Book of Abstracts, 246-247.

SAS, 2006. SAS 9.1 Software Documentation and Examples. SAS Institute Inc., Cary, NC.
Available at: http://support.sas.com/rd/app/papers/. Accessed November 1, 2006

Schertz, C.E., Brown, G.K., 1968. Basic considerations in mechanizing citrus harvest.
Transaction of the ASAE., 343-346.

Sevier, B.J., Lee, W.S., 2003. Adoption trends and attitudes towards precision agriculture
in Florida citrus: preliminary results from a citrus producer survey. ASAE Meeting Paper No.
031100. ASAE, St. Joseph, MI.

Sevier, B.J., Lee, W.S., 2004. Precision agriculture in citrus: a probit model analysis for
technology adoption. ASAE Meeting Paper No. 041092. ASAE, St. Joseph, MI.

Stajnko, D., Lakota, M., Hocevar, M., 2004. Estimation of number and diameter of apple
fruits in an orchard during the growing season by thermal imaging. Computers and Electronics in
Agriculture 42, 31-42.

Stajnko, D., Cmelik, Z., 2005. Modelling of apple fruit growth by application of image
analysis. Agriculturae Conspectus Scientificus 70, 59-64.

Tabb, A.L., Peterson, D.L., Park, J., 2006. Segmentation of apple fruit from video via
background modeling. ASABE Meeting Paper No. 063060. ASABE, St. Joseph, MI.

Tian, L.F., Slaughter, D.C., 1998. Environmentally adaptive segmentation algorithm for
outdoor image segmentation. Computers and Electronics in Agriculture 21, 153-168.

Welte, H.F., 1990. Forecasting harvest fruit size during the growing season. Acta
Horticulturae 276, 275-282.

Williams, P., and Norris, K., 2001. Near-Infrared Technology: In Agriculture and Food
Industries, 2nd ed. American Association of Cereal Chemists, Inc., St. Paul, MN.

Winter, F., 1986. Modeling the biological and economic development of an apple orchard.
Acta Horticulturae 160, 353-360.

Whitney, J.D., Wheaton, T.A., Miller, W.M., Salyani, M., Schueller, J.K., 1998. Site-
specific yield mapping for Florida citrus. Proc. Florida State Horticulture Society 111, 148-150.

Whitney, J.D., Q. Ling, Wheaton, T.A., Miller, W.M., 1999. A citrus harvesting labor
tracking and yield monitoring system. ASAE Meeting Paper No. 993107. ASAE, St. Joseph, MI.

Whitney, J.D., Q. Ling, Wheaton, T.A., Miller, W.M., 2001. A citrus harvesting labor
tracking and yield monitoring system. Applied Engineering in Agriculture 17, 121-125.









BIOGRAPHICAL SKETCH

The author was born in 1981 in Mansfield, Ohio. Most of his childhood was in

Melbourne, FL where he graduated from Eau Gallie High School, 1999 and Brevard

Community College with an Associate of Arts degree in 2001. In May 2005 he

graduated with a Bachelor of Engineering degree in electrical engineering from the

University of Florida. Continuing his education, Kevin graduated from the University of

Florida in August 2007 with a Master of Engineering degree in agricultural and

biological engineering and another Master of Engineering in electrical and computer

engineering. Kevin Kane currently lives in Aiken, SC with his loving wife Dr. M. Kane

and his two dogs Bailey and Killian.





PAGE 1

1 IDENTIFICATION AND CLASSIFICATION OF GREEN CITRUS BY SPECTRAL CHARACTERISTICS FOR PRECISION AGRICULTURE By KEVIN EDWARD KANE A THESIS PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLOR IDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF MASTER OF ENGINEERING UNIVERSITY OF FLORIDA 2007

PAGE 2

2 2007 Kevin E Kane

PAGE 3

3 To my loving wife for all her support

PAGE 4

4 ACKNOWLEDGMENTS First I would like to thank my major advisor, Dr. Won Suk Daniel Lee, for his support and encouragement, both in my research and in my life. I would also like to give thanks to my other supervisory committee members, Dr. Thom as Burks and Dr. Arnold Schumann, for their advice and suggestions during my research. My family and friends were a great help and sustained through the good and bad times of this educational journey. Above all however, I would like to thank my wonderful and caring wife. I would surely not have pursued and finished graduate school without her by my side.

PAGE 5

5 TABLE OF CONTENTS page ACKNOWLEDGMENTS...............................................................................................................4 LIST OF TABLES................................................................................................................. ..........7 LIST OF FIGURES................................................................................................................ .........8 ABSTRACT....................................................................................................................... ............11 CHAPTER 1 INTRODUCTION..................................................................................................................13 1.1 Florida Citrus................................................................................................................. ..13 1.2 Precision Agriculture in Florida Citrus............................................................................13 1.2.1 Research history and trends...................................................................................14 1.2.2 Yield monitoring and mapping..............................................................................15 2 OBJECTIVES..................................................................................................................... ....21 2.1 Identification of Critical Wavelengths.............................................................................21 2.2 Green Fruit/Leaf Separation Using an NIR Camera........................................................21 3 CITRUS CLASSIFICATION BASE D ON DIFFUSE RE FLECTANCE.............................22 3.1 Introduction................................................................................................................... ...22 3.2 Literature Review............................................................................................................22 3.3 Materials and Methods....................................................................................................23 3.3.1 Green Citrus and Green Leaves.............................................................................23 3.3.2 Diffuse Reflectance...............................................................................................25 3.3.3 Discriminability.....................................................................................................26 3.3.4 Fisher Linear Discrimi nant Analysis (FLDA).......................................................28 3.3.5 Training and Validation.........................................................................................29 3.4 Results........................................................................................................................ ......30 3.4.1 Reflectance Characteristics of Green C itrus Fruit vs. Green Citrus Leaves.........30 3.4.2 Reflectance Characteristics of Citrus Varieties.....................................................32 3.4.3 Spectral Growth Patterns of Maturing Citrus Fruit...............................................35 3.4.4 Training and Validation Results............................................................................36 3.4.5 Discriminability Using All Samples......................................................................37 3.4.6 Fisher Linear Discriminant Analysis (FLDA) Results..........................................41 3.5 Conclusions.................................................................................................................... ..43 4 CITRUS INDENTIFICATION BY NIR CAMERA SYSTEM.............................................44

PAGE 6

6 4.1 Introduction................................................................................................................... ...44 4.2 Cameras Used in Citrus Yield Mapping..........................................................................44 4.3 Materials and Methods....................................................................................................45 4.3.1 NIR Camera, Optical Equipment, and Hardware..................................................45 4.3.2 Experimental Location and Environment..............................................................49 4.3.3 Image Processing...................................................................................................51 4.3.3.1 Image processings major stages.................................................................52 4.3.3.2 Image Indexing stage..................................................................................53 4.3.3.3 Marker Checking stage................................................................................57 4.3.3.4 Result analysis.............................................................................................62 4.4 Results and Discussion....................................................................................................64 4.4.1 Image Acquisition Problems.................................................................................64 4.4.2 Quantitative Identification.....................................................................................68 4.5 Conclusions.................................................................................................................... ..75 5 CONCLUSIONS AND FUTURE WORK.............................................................................81 5.1 Conclusions of Research Objectives...............................................................................81 5.2 Future Work.................................................................................................................... .82 REFERENCES..................................................................................................................... .........83 BIOGRAPHICAL SKETCH.........................................................................................................86

PAGE 7

7 LIST OF TABLES Table page 3-1 List of dates that variety sample sets were harvested........................................................24 3-2 Discriminability results of the training data set, Methods I and II....................................37 3-3 Discriminability results using all the data..........................................................................39 3-4 Projection vector equa tions found by Fisher Linear Discriminant analysis......................42 4-1 Optical band pass filter details...........................................................................................46 4-2 Validation image citrus fruit pixel results..........................................................................69 4-3 Validation image fruit marker a nd manually masked marker results................................74

PAGE 8

8 LIST OF FIGURES Figure page 3-1 Samples from August 9, 2005 of A) green c itrus fruit and B) gr een citrus leaf................24 3-2 Growth chart of the average fruit variet ies size, as measured by fruit diameter..............25 3-3 Green citrus fruit positioned for diffuse reflectance testing with the Cary 500 spectrophotometer with integrating sphere........................................................................26 3-4 Histograms of example fruit and leaf refl ectance. A) Weak discriminability at 1650 nm. B) Strong discriminability at 881 nm........................................................................27 3-5 Green citrus and leaf average spec tral reflectance thr ough the fall 2005 growing season......................................................................................................................... ........31 3-6 Spectral reflectance of layered citrus leaves and single Orlando Tangelo.....................32 3-7 How multiple leaves/surfaces create added diffuse reflectance........................................32 3-8 Citrus varieties average spectral reflectance......................................................................34 3-9 All citrus and citrus varieties standa rd deviation values for all wavelengths....................34 3-10 Citrus varieties average spectral absorbance.....................................................................35 3-11 Reflectance changes over the growing season of Hamlin.................................................36 3-12 Fruit & Leaf (Method I) with Fisher projec tion line (solid line) and classification line (dashed line).................................................................................................................. .....37 3-13 Fruit & Leaf (Method II) with Fisher proj ection line (solid li ne) and classification line (dashed line)............................................................................................................. ...38 3-14 Discriminability results of a ll the sample citrus vs. leaves................................................39 3-15 Tangelo & Hamlin with Fisher projecti on line (solid line) an d classification line (dashed line).................................................................................................................. .....40 3-16 Tangelo & Valencia with Fisher projecti on line (solid line) a nd classification line (dashed line).................................................................................................................. .....40 3-17 Hamlin & Valencia with Fisher projec tion line (solid line) and classification line (dashed line).................................................................................................................. .....41 3-18 Scatter plot showing all classes: Valencia, Hamlin, Tangelo, and Leaf............................42

PAGE 9

9 4-1 Spectral sensitivity range of InGasAs technology (Graph courtesy of FLIR Systems, Inc.).......................................................................................................................... ..........46 4-2 Three band pass (1150 nm, 1064 nm, and 1572 nm).........................................................46 4-3 Solar spectral irradiance reference from the ASTM G173-03 of direct circumsolar. This data was gathered by the American Society for Testing a nd Materials (ASTM) and government research laboratories...............................................................................48 4-4 Percent transmittance for each of the band pass filters......................................................49 4-5 In field experimental setup with gasolin e generator, personal computer, monitor, NIR camera on tripod, and Teflon disk (from foreground to background)...............................50 4-6 Representation of an image spectral block........................................................................52 4-7 Histogram stretching functio n; raw image and histogram.................................................54 4-8 Image Indexing stage of valid ation image number eight...................................................58 4-9 Image Indexing stage of training image number 11..........................................................59 4-10 Marker Checking stage of validation imag e number nine. Input binary image of possible fruit markers.........................................................................................................61 4-11 Complete image processing procedure..............................................................................62 4-12 Manually produced image mask showing th e difference between citrus fruit pixels and non-citrus fruit pixels..................................................................................................63 4-13 Location of the target fruit and some of the leaves has shifted between the two raw images. The Teflon sheet and some leaves remain in the same place..............................65 4-14 Lighting condition extremes (blackout and saturation) from inside and outside the citrus canopy.................................................................................................................. ....66 4-15 Changes in sunlight and th e effect on image histograms...................................................67 4-16 Pixel classification results for all va lidation images and the overall total.........................71 4-17 Manually masked fruit pixel count versus image processing fruit pixel count results with linear best fit usi ng all pixel count data.....................................................................73 4-18 Manually masked fruit pixel count versus image processing fruit pixel count results with linear best fit with one outlier removed.....................................................................73 4-19 Complete image processing and results check of validation im age number five..............76 4-20 Complete image processing and results check of validation image number two..............77

PAGE 10

10 4-21 Complete image processing and results check of validation image number three............78

PAGE 11

11 Abstract of Thesis Presen ted to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Master of Engineering IDENTIFICATION AND CLASSIFICATION OF CITRUS FRUIT BY SPECTRAL CHARACTERISTICS FOR PRECISION AGRICULTURE By Kevin Edward Kane August 2007 Chair: Won Suk Daniel Lee Major: Agricultural a nd Biological Engineering Citrus production is a multi-million dollar indus try in the state of Florida. The crops economic significance is the drivi ng force behind current developm ents in precision agriculture technologies assuring citrus grow ers can minimize their costs, protect the environment, and increase overall yield. A real-time citrus yiel d map while the citrus fruit are still maturing provides information to growers, giving them time to be proactive at improving the groves growth and planning ahead for the harvest. Recent research has shown cameras a image processing techniques have the ability to identify and count orange citrus fruit in the grove. However, there is a desire by the industry to have citrus yield maps earlier in the growing season, a time when citrus fruit are green. In this case, a traditional visible spectral camera can not accurately identify green citrus fr uit against their green tree canopy. The objective of this research was to use spect ral information from the near infrared (NIR) reflectance spectrum to identify citrus fruit while they are still green. To begin this work, 540 freshly harvested samples of green citrus fruit and leaves were gather ed and measured their diffuse reflectance using a spectrophotometer. Th e resulting spectral curves from 400 nm to 2500 nm were analyzed using discriminability ca lculations to find critical wavelengths for separation. Fisher linear disc riminant analysis showed the wavelengths of 881 nm and 1383 nm

PAGE 12

12 provided perfect green leaf and green fruit separation. This research provided a foundation for the design of an in-field NIR camera syst em for green citrus fruit identification. A highly sensitive NIR camera outfitted with three optical band pass filters (1064 nm, 1150 nm, and 1572 nm) were used for natural in-field image acquisitions. The 256 bit monochromatic images were studied using spec tral indexing and image processing schemes. Using training images, an indexing and image pr ocessing algorithm was developed and tested on validation images. A 90.3 % co rrect pixel classifi cation result was obtai ned, proving that NIR camera images can successfully be used in the iden tification of green citrus fruit in the grove. An R2 of 0.746 for fruit pixel counts verse a manua lly masked fruit pixel counts was achieved. Despite these positive accomplishments, the rese arch has also revealed problems that are prohibitive to this identification method. These problems included the high financial cost of an NIR imaging system and the difficulty of proper illumination and multiple image alignments when working in a normal Florida citrus grove environment.

PAGE 13

13 CHAPTER 1 INTRODUCTION Florida Citrus Florida was first introduced to citrus in the 16th century by European explorers; with the first commercial production of c itrus beginning in 1763. Over the next two centuries the citrus industry grew to a peak of over 350,000 ha of citr us in Florida (Sevier and Lee, 2003). In 2002, there were 322, 658 ha of commercial citrus gr oves spread throughout Fl orida, accounting for more than 74% of the total citrus grown in the United States. Among the citrus fruit varieties, the orange accounted for 81.4% of the producti on, followed by grapefruits at 13.2%, and other specialty fruits (tangerines, tangelos, lemons, limes, etc.) ma king up the last 5.4%. (Florida Agricultural Statistics Services [FASS], 2005) With an on-tree value of citrus estimated to be $1.1 billion in 2005 (Florida Agricultural Statistics Services [FASS], 2005), the citrus market has become extremely important to Floridas economy. Currently international trade is putting pressure on the Flor ida citrus market. Cheaper labor in other countries, such as Brazil, has made the cost of staying competitive in the United States difficult for many growers. The need to increase yield per acre while lowering overall production costs is vital to the survival of the Florida Citrus industry. It appears that through technology and smarter grove management techni ques like precision agriculture, the citrus industry will continue to thrive. Precision Agriculture in Florida Citrus Traditional whole field management approach es treat the entire crop production area in a uniform manner ignoring natural in-field variability. This can lead to under or over application of field inputs at different loca tions. Precision agriculture is a managerial technique of using high-tech equipment to monitor and then treat sma ller sections of a larger area on a site-specific

PAGE 14

14 basis. This entails managing input s (fertilizer, limestone, herbicide, insecticide, seed, etc.) to the farming land based on the variability inherent in the field. The goals of precision agriculture are to reduce waste, increase profits, and maintain the quality of the environment (Morgan and Ess, 2003). Due to the input-intensive nature of citrus production and Floridas volatile trend of cost per unit area, precision agriculture is becoming a vital technique to manage the cost of production (Sevier and Lee, 2003). The Florid a citrus production industry lends itself perfectly to the saving that can come from ma naging and controlling ex pensive inputs to the grove. Cost of production from input applicatio ns can be brought to manageable levels by precision agriculture. In addition to minimizing input costs, preci sion agriculture offers better managed groves, a protected environment, a nd increased crop yield and profits. Proper implementation of a precision agri cultural system requires knowledge of in-field variability; in the case of this research, in-f ield citrus yield variability. Research history and trends According to the Citrus Research and Educa tion Center (CREC) at the University of Florida, in early 1996 a meeting be tween citrus industry personal a nd the University of Floridas Institute of Food and Agricultural Sciences (IFAS ) was held to discuss "decision support systems for citrus". This meeting resu lted in the creation of the Deci sion Information Support System for Citrus (DISC) group at the CREC in Lake Alfre d, FL. The goal of this group was to pursue research in the area of precision agriculture, with an emphasis on: mapping citrus yield, canopy volume, variable rate application and uses of the Global Positioning System (GPS) and geographic information systems (GIS) in commercial groves. This groups research has lead to advancements in precision citrus grove techno logies and numerous published papers over the past decade.

PAGE 15

15 Sevier and Lee (2003 and 2004) investigated Florida citrus growers adoption of new technologies, including: sensor-b ased variable rate applicators, prescription map based variable rate applicators, pest scouting/mapping, GPS, soil variability mapping, water table monitoring, and yield monitoring and mapping. Their finding s showed the most commonly used precision agriculture technologies were sens or-based variable rate appli cators and soil variability mapping at usage rates of 18.6% and 18.0%, respectivel y (Sevier and Lee, 2003). The least used technology was remote sensing data from plane or sa tellite images, with a us age rate of just over four percent. Findings also showed that gr owers age had a negative correlation with their likelihood of incorporating the technologies i nvestigated (Sevier and Lee, 2004). That is younger growers were more likely to adopt the new technologies th an older growers. This can be understood in a social realm by the younger gr ower being quicker to learn and use the computer technologies they have grown up wit h. The study also showed growers with higher in-grove variability are more likely to adopt new technology wh en compared to growers with less variability in their field. In the Sevier and Lee survey (2004), responde nts were asked to provide a reason for not adopting all new technologies. The leading reas on for none adoption was that growers were satisfied with current practices. They also described themselves as being somewhat reluctant to adopt, and normally wait to see others suc cess before implementing new technologies. These trends show the Florida citrus industr y is moving slowly in the direction of new technologies, while continued research at the un iversity level is needed to show growers the success they desire before adoption. Yield monitoring and mapping Knowledge of variable yields within blocks of citrus groves may provide information that a grower can use to find the cause of the vari ability (Whitney et al., 1998). There are many

PAGE 16

16 elements within a grove or block that can lead to variability, including tr ee sizes, age, health, spacing, soil type, fertility, wate r availability, fertilizer applic ation, and more. With proper knowledge of the variability in th e field, grove managers may act not only in the right locations but also in a timely manner. Fruit yield monitoring techniques may offer be nefits for forecasting the number of fruit and quality at the time of harvest. Apple yield information together with ecological, cultivar, and price parameters can predict future yields, which in turn allows management to plan for incomes and calculations of profit (Welte, 1990). The Eur opean apple and pear industry only uses the Prognosfruit Forcasting Model for their estimati ons of yield quantity and quality; however, this is a long tedious process requiri ng counting measurements of required parameters. This also limits the ability to predict individual orchards future yields (Winter, 1986; Stajnko et al., 2004). Similarly, future citrus grove forecasting systems will also require yield quantity and quality parameters. Thus, the Florida citrus industry, like the European apple and pear industry will benefit from the use of fruit yield monitoring systems. There have been several yield monitoring and mapping concepts proposed but few have found their way to actual development and field rese arch. For the most part these systems can be grouped into three basic methods: 1) counting and mapping citrus tubs during manual harvesting 2) tracking citrus flow during mechanical harv esting, and 3) using automated computer vision systems before harvesting. Citrus grove yield maps can be created dur ing manual harvesting by marking the locations of filled tubs of fruit with a GPS unit in the grove (Whitney et al., 1998). This discrete data can be interpolated to create two dimensional yi eld maps. Using this method, yield maps can provide more meaning to growers by being presente d in a boxes/acre format. DISCs early work

PAGE 17

17 was with GeoFocus Inc, the only company at the time to offer a commerci ally available citrus yield monitoring system. Their Crop Harves t Tracking System (CHTS) required the truck operator to press a button marking his/her GPS location every time a tube was lifted and emptied into a goat truck. This seemingly easy system however had problems when operated with untrained or ill-informed drivers. Some times the operator would forget to push the button, or push the button multiple times just to be su re it was done. Other times the operator would remember to press the button later on, only after moving th e truck location which means improper markings. All these button pushing mistak es diluted the data and resulting yield maps accuracy. Advancements to this system include d automating the GPS system by way of a weight threshold switch on the trucks lifting arm and pr essure transducers mounted on the goat truck for weighing the citrus total (Whitney et al., 199 9; 2001). Currently GeoAg Solutions located in Lehigh Acres, Florida, is the only company to provide citrus yield and monitoring solutions. They have incorporated more value added in centives with the technol ogy such as monitoring worker progress, chemical a pplications and managing paymen ts based on harvest counts. GeoAg Solutions main system is called Citr iTrack with three a dditional sub-programs available: HaverstMap 2.0, HarvestPay, and HarvestWatch. The CitriTrack system can handle payroll, grove mapping, and tracks real-time ha rvesting progress. CitriTrack uses GPS and wireless technology to connect the grove manager di rectly to the grove. Up to date information is collected by a computer (also av ailable for purchase) that is att ached directly to the harvesting equipment. Similar to the GeoFocus system, the equipment operator logs each tub and its location and links it with the a ppropriate picker. HarvestMap 2.0 uses the yield data information collected from CitriTrack to produce informativ e full-color maps using GIS technology. GeoAg Solutions claim a grower can track the daily pr ogress in a block down to specific boxes, specify

PAGE 18

18 variations within a block for chemical applicat ions or even target and apply products based on return on investment potential. This data is collected in the fi eld can also be routed to the HarvestPay software. The program tracks the in formation and can provide reports about worker productivity and hourly wages. The final s ub-program, HarvestWatch, can allow the grove manager stay connected the crews with up-to -the-minute reports no matter their location. A second method currently being explored is th e use of fruit flow tracking technologies on mechanical citrus tree canopy shaker harvesting equipment (Grift et al., 2006; Chinchuluun et al., 2007). Currently a canopy s haker dislodges a citrus fru it from the canopy by violently shaking the citrus canopy. Freed citrus fruit can either fall to the ground and be harvested later by hand or the fruit can be caught by a conveyor mechanism, lifting the citrus up to a ramp where the citrus rolls down into a collection bin. Grift et al (2006) have expanded previous research in estimating particle flow rates, more specifically fertilizer, by observing and measuring the space and time between clumps of particles. In this case, the clumps and spacings is measure by a laser and the particles are much larger and slower moving. Another more precise, but ultimately more complex solu tion, is counting the individual citrus by high speed cameras and using image processing t echniques (Chinchuluun et al., 2007). Although current research is focused on fruit catching equi pment working in conjunction with tree shakers, it should be noted that other forms of mechan ized citrus harvesting equipment would produce even finer grade citrus grove yield maps, such as robotic citrus harves ting equipment. Such future systems could provide not only an exact c ount of fruit per tree, but also a size and quality of each fruit. The final method explored for citrus yield map creation is the use of cameras with automated image processing and assessment technique s to count the number of citrus on tree.

PAGE 19

19 By corresponding the images citrus counts with their acquisition location in the grove by a GPS receiver, a highly detailed citrus yield map can be formulated. What separates this method from the pack is the availability of the yield map befo re the citrus is harvested. This provides time to grove managers to prepare more precisely the logi stics of that harvested season and a head start at improving conditions for the following growi ng season. The research discussed throughout the remainder of this thesis is with respect to this on tree fruit coun ting citrus yield mapping method. Agriculturally based vision systems have been used for the identification of everything from apples to weeds (Stajkno and Cmelik, 2005; Lee and Slaughter, 2004). On tree fruit identification using vision systems is not a ne w concept with original concepts dating back nearly forty years (Schertz and Brown, 1968). Research in fruit identification systems using machine vision have had mixed results with wide variations of camera systems, experimental tests, and image processing algorith ms (Jimenez et al., 2000). Resent research at the University of Florida in the area of on tree citrus identification incl udes Annamalai et al. (2004), Chinchuluun and Lee (2006), and M acArthur et al. (2006). In thes e cases, a digital camera took visible light images of trees and then used image proces sing techniques to separate yellow/orange colors from the green canopies. In Chinchuluun and Lees (2006) research, multiple cameras were used and the system wa s pulled behind a truck while passing through the grove rows. This allowed the images to be acqui red from the side of the trees in high detail, which permitted morphological image processing techniques to work due to the closeness of the targets. MacAuthur et al. (2006) investigated the same concep t but by using a remotely-piloted helicopter. This gave an advantage of flexible mobility throughout the grove, but additional skill was needed to gather high quality images. In both cases, citrus occlusion, shadows and the

PAGE 20

20 separation of touching fruit were the most difficult problems face d. Research preformed on trees with different water treatment leve ls provided an interesting observ ation. The health ier the trees canopy, the more difficult it was to count the fruit. A tree with thin or little leaf coverage made counting fruit easier and increa sing the yield estimation, while a tree with healt hy thick leaf coverage made the observation off all fruit impos able, thus low yield estimation (MacAuthur et al., 2006). It has been suggested that a wei ghted multiplier based on canopy fullness could be used to adjust these citrus yiel d estimations; however, no such work has been started at the time. However, nearly every fruit finding vision system has used the visible light spectrum as the only means to decipher fruit from the surroundi ng leaf canopy. One resent exception was the promising work of Stajnko et al. (2004; 2005) us ing thermal imaging of apples, but there is no published work with citrus imaging outside the vi sible spectrum. A majo r consideration for the imaging of citrus is how the vi sible spectrum imaging research ha s only been preformed late in the growing season only after citrus has changed from a green to an orange color. A gap in citrus research is early season identification of citrus fr uit, a time when the fruit is the same dark green color as the surrounding leaf canopy. Solving these problems can allow precision agricultural techniques to be used earlier in the growi ng season, thus compounding the benefits of early information on yield, healt h, and in-field variability.

PAGE 21

21 CHAPTER 2 OBJECTIVES Identification of Critical Wavelengths To properly identify and separate green citr us fruit from surrounding green citrus leaves using only near-infrared (NIR) spectral information, an investigation into the reflective characteristics was preformed. This involved th e use of highly sensitive optical equipment to determine the spectral res ponse of green citrus fruit and green citr us leaves in the visible to nearinfrared range (VIS-NIR). Use of a statisti cally significant number of samples through the growing season provided large amount s of data to reliably analyze. It was the first objective of this research to identify the significant waveleng ths needed to separate green citrus fruit from green citrus leaves in a controlled laboratory environment (Chapter 3). Green Fruit/Leaf Separation Using an NIR Camera For future developments of early season citrus yield mapping systems, a test of in-field green citrus fruit/leaf separati on was conducted in a Florida citrus grove using a highly sensitive NIR camera outfitted with optical band pass filters. Selection of the band pass filters was determined by the results of th e first objective. The NIR cam eras resulting monochromatic green citrus fruit and green leaf canopy imag es were post processed using image processing techniques. Differences in the spectral character istics were the only image processing tools used for citrus identification. The s econd objective of this research wa s to determine if separation of green citrus fruit from green citrus leaves wa s possible in-field using current NIR camera and image processing technologies (Chapter 4).

PAGE 22

22 CHAPTER 3 CITRUS CLASSIFICATION BASE D ON DIFFUSE REFLECTANCE Introduction Before designing a remote sensing camera system for the identification of green citrus in a green citrus leaf canopy, an in -depth study of the diffuse refl ectance characteristics of green citrus and green citrus leaves was required. Wh ile the spectral responses of citrus leaves and citrus fruit had been studied many times in the past, the use of this information to find and identify a robust separation scheme in the near in frared (NIR) range had never been published. It was vital to identify critical wavelengths for separation, before proceeding with in-field camera research. It is only through the spectral characteristics that prop er judgments could be made for the design and execution of the in-field NIR camera research. Literature Review Research in the field of near infrared (NIR ) sensing technologies on agriculture has been around for over a century. In most early research, the interactive nature between light and leaves was studied (Williams and Norris, 2001). It has been a more recent adaptation to incorporate NIR remote sensing equipment into the actual agri cultural process, both pre and post harvest. This can be observed in the ne w technologies to track and m onitor crop products and personnel (Whitney et al., 2001; Aleixos et al., 2002). Citrus should be considered a late entry in to the field of NIR sensing, despite the first major study on NIR diffuse reflectance of citrus fruit being three decades old (Gaffney, 1972). Since that research no real follow up work was conducted. Most likely this gap was due to two issues: 1) lack of valuable uses of the NIR technologies in the industry. Agriculture is among the slowest industries to research and adapt new techno logies, as managers trad itionally held a, if it works dont fix it mentality (Sev ier and Lee, 2003), and 2) the high cost of performing NIR

PAGE 23

23 research in the past. Lower prices for better equipment and technologies are quickly eliminating these concerns for both the researchers and the even tual end users, the grove managers. Florida, and all U.S. citrus growers, will need to apply new managing solutions to compete with the growing international markets (Florida Agricult ural Statistics Services [FASS], 2005). This means high-tech solutions are desired for preharvesting and post harvesting, resulting in the increase of NIR sensing rese arch over the last decade. Currently the most commonly used NIR technol ogy in the agricultural fields of the world is multiand/or hyper-spectral imagery from planes and satellites. This has lead to a lot of research in spectral vegetation indices (SVI) for finding a great number of important field information, such as disease, crop moisture, a nd weeds (Apan et al., 200 3; Alchanatis et al., 2006; Gumz and Weller, 2005). Despite this, there has been little research preformed with these types of cameras and sensors on the ground. Th is research is a preliminary study of the reflectance characteristics of gr een citrus and green citrus leaves for the conceptual method design of a simple multi-spectral ground system, fo r separation of green citrus from green leave canopy. Materials and Methods Green Citrus and Green Leaves During the fall 2005 citrus growing season, J une 2005 to January 2006, samples consisting of one green citrus leaf and one green citrus fruit located next to each were acquired from the University of Floridas citrus re search grove. A sample set consisted of 10 samples (10 fruits and 10 leaves) gathered on the same day and of the same variety. Sample sets were obtained weekly from one of the three citrus varieties: Hamlin ( Citrus sinensis ), Orlando Tangelo ( Citrus X tangelo ), and Valencia ( Citrus sinensis ). There were a total of 27 sample sets

PAGE 24

24 collected: for a total of 270 indivi dual fruits and 270 i ndividual leaves. Th e dates and varieties of each harvesting sample set are listed in Table 3-1. Table 3-1. List of dates that va riety sample sets were harvested. Tangelo Hamlin Valencia 7/6/05 7/12/05 10/18/05 7/19/05 7/26/5 11/1/05 8/2/05 8/23/05 11/22/05 8/9/05 9/7/05 12/6/05 8/16/05 9/20/05 12/20/05 8/30/05 10/5/05 1/3/06 9/14/05 10/19/05 1/11/06 9/27/05 11/2/05 10/11/05 11/29/05 10/25/05 12/13/05 Figure 3-1 is an example of the green citrus and green leaves that were harvested during the growing season. Each sample was weighed to the nearest 1/100th of a gram by a digital scale (Adventurer, Ohaus, Inc., Pine Brook, NJ), and diameters (horizontally at the widest cross section with the stem pointing upwards) were m easured using a sliding caliper to the nearest 1/100th of a millimeter on the same day as being harvested. A B Figure 3-1. Samples from August 9, 2005 of A) gr een citrus fruit and B) green citrus leaf. Figure 3-2 shows the citrus varieties growth during the fall season as measured by average width. Only ten citrus fruit samples were harv ested every other week, or more, which allowed

PAGE 25

25 the variance in average size to swing up and dow n. Close observation of Figure 3-2 even shows instances where the average size of the ten samp les declined from the proceeding ten samples, such as Hamlin week two to week four. This was due to no other reason than the random nature of the fruit sizes harvested. The overall growth pattern, fruit sizes and seasonal developments are clear. Growth results, as measured by weight showed similar results but have been excluded due to its irrelevance to the remainder of this thesis. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 45 50 55 60 65 70 75 Week Number (week 1 is June 6th, 2005)Size (mm) Tangelo Hamlin Valencia Figure 3-2. Growth chart of the average fruit va rieties size, as measured by fruit diameter. Diffuse Reflectance Diffuse reflectance of the leaf and fruit sa mples was measured by a spectrophotometer (Cary 500, Varian, Inc., Palo A lto, CA) with an integrating sphere (DRA-CA-5500, Labsphere, Inc., Brossard, Qc, Canada) on the same day as harvesting (Figure 3-3). The samples were not cleaned prior to measurements as natural on tr ee and in-field charact eristics were desired. The spectrophotometer measured percent diffuse reflectance of the samples between 200 and 2500 nm in one nanometer increments. When the spectrophotometer was used, the two lamps (deuterium and tungsten) were allowed to warm up for one hour prior to te sting to stabilize the

PAGE 26

26 light sources. A diffuse reflectance base line was measured using a 50 mm diameter polytetrafluoroethylene (PTFE) sample disk, which was used to obtain the optical reference standard for the system each day. Figure 3-3. Green citrus fruit positioned for diffuse reflectanc e testing with the Cary 500 spectrophotometer with integrating sphere. Discriminability Two types of classification were desired from the data set. The most important was a distinction between citrus leaves and any citrus variety. This recognition would allow citrus growers to identify all immature green fruit regardle ss of variety. This information is critical as it offers transferable functionality to the resear ch not only to local ora nges and tangelos but also to limes, grapefruits and other members of the ci trus family. The sec ond classification studied the separation of citrus fruit into their variet y classes: Tangelo, Vale ncia, and Hamlin. There were normal minor spectral reflectance changes over the course of the si x month growing season due to chemical and physical alte rations within the fruits peel, based on the research of Nagy et citrus integrating sphere Car y 500

PAGE 27

27 al. (1977). However, for the purpose of this research a normal Gau ssian distribution was assumed at each wavelength for the reflectance of a ll sample fruits and leaves. Histograms of all fruits and leaves for wavelengths of 1650 a nd 881 nm are shown in Figures 3-4A and B, respectively. These example wa velengths serve as a visual co mparison of strong and weak discriminability. A 5 10 15 20 25 30 35 40 45 50 0 10 20 30 40 50 60 Reflectance (%) at 1650 nmSample count Fruit Leaf 45 50 55 60 65 70 75 80 85 90 0 10 20 30 40 50 60 Reflectance (%) at 881 nmSample count Fruit Leaf B Figure 3-4. Histograms of exampl e fruit and leaf reflectance. A) Weak discriminability at 1650 nm. B) Strong discriminability at 881 nm. Discriminability comprises two mathematical measures: the distance between means and the standard deviation of two proba bility density functions (PDFs). If we image the graphs from Figure 3-4 to be normalized and have an area un der the curve of one unit, then Figure 3-4(A) shows a small overlap between the two PDFs, while Figure 3-4( B) shows a gap between the PDFs. This difference is a result of th e discriminability of the two wavelengths. The mathematical strength of the discrimi nability between two PDFs with the same standard distributions is defined by Duda et al. (1988) as: 1 2' d (3-1) Where, d = discriminability = standard deviation

PAGE 28

28 1 ,2 = means of class 1 and 2 In general, a higher discriminability value is de sired as it shows a greater separation exists between classes. In the case of this data set, the standard deviations are different between the two classes, thus the formula a bove can not be used. For this reason the following alteration to equation 3.1 was made: 2 / ) ( '2 1 1 2 d (3-2) Where, d = discriminability 2 1, = standard deviations of class 1 and 2 1 ,2 = means of class 1 and 2 Averaging the standard deviations of the tw o classes and replacing this for the standard deviation of equation 3.1 allows the discriminabili ty to scale with the standard deviation change of both PDFs. Discriminability is a simple and reliable way to determine which wavelengths have the greatest possible separation; howev er, this should not be the only consideration. Reflectance characteristics at one wavelengt h share common properties with t hose wavelengths near it. To maximize the quality of multiple feature extraction, a limit on the minimal distance between two wavelength features was set. This threshold lim it will be defined in the Training and Validation section that follows. Fisher Linear Discriminant Analysis (FLDA) Dimensionality reduction is an issue with many recognition systems using more than one variable. The most well known and commonly used dimension reduction t echnique is principal component analysis (PCA). This method search es for feature spaces in the multi-dimensional

PAGE 29

29 data that contains the greatest separation between classes with rega rd to variance. By projecting multi-dimensional data points onto a new calcu lated space, a smaller number of dimensions maybe used for classification. In the case of two dimensional data, a one dimensional direction is found. Fisher linear discriminant analysis (F LDA) is a common type of PCA first seen in the famous paper by Fisher (1936). In our research the analysis tec hnique was used to calculate the direction, w, for projection. Duda et al. (1988) defines w as the linear function yielding the maximum ratio of between-class scatter to with in-class scatter. A positive byproduct of this analysis technique is a reduction of noisy directions. A onedimensional projection direction was calculated, reducing data from a two dimensional space, defined by two feature wavelengths, to a one dimensional space. This methodology was chosen for this research as it adapts well to future NIR computer vision and im age processing research. By using this method a stack of digital images at different spectra l bands, know as a spectral image cube, can be dimensionally reduced to a single two dimensional image. Each pixel of this image would have its own uniquely calcul ated class likelihood. Training and Validation Prior to using the pattern r ecognition techniques discusse d previously, a performance evaluation was conducted using two-thirds of th e samples as training data and one-third as validation data. The 180 fruit and leaf training samples were selected at random. Using only the training samples, wavelengths for separation were selected, by discriminability as was discussed previously. FLDA was then used to find the best projection vector (dir ection). The remaining 90 fruit and leaf validation samples were cl assified after being projected onto the one dimensional space. Only the discrimination between fruit and leaf was verified using these techniques, as it was the prim ary purpose of the research.

PAGE 30

30 For fruit vs. leaf classificati on, two methods (Method I and Me thod II) were investigated. Method I included the use of the feature wavelengths with the tw o highest discriminability ( d ) calculations that also met a th reshold distance of 100 nm apart. This 100 nm threshold was chosen arbitrarily to remove multicolinearity. Thus, the two selected wavelengths were allowed anywhere in the spectral range as long as they remained 100 nm apart. Method II required one wavelength feature having a fruit reflectance greater than that of the leaf, while the second wavelength feature having the fruit reflectance less than that of th e leaf. The reader might be best served to glance over Figure 3-5 of the ne xt page to better unders tand Method II. This second method was tested for fruit and leaf cl assification as it offe red wavelengths with contrasting reflectance magnitudes. Results Reflectance Characteristics of Green Citrus Fruit vs. Green Citrus Leaves Figure 3-5 presents the average reflectance spectra of 270 citr us samples with respect to the 270 leaf samples. From the spectral reflectan ce characteristic curves, two traits should be observed. The large increase of reflectance be tween 690 and 720 nm is referred to as the red edge. This is a result of plants having highe r absorption rates in the visible range due to photosynthetic pigments. Absorption rates of th e red edge region by chlorophyll pigments are lower resulting in higher reflect ance (Ding, 2005). The second tra it is water absorption bands at 970, 1450, and 1940 nm. These valleys in reflecta nce magnitude are a result of light absorption by water within the samples. Differences between the fruit and leaf reflect ance include a large magnitude change from 720 to 1120 nm. A simple experiment suggests this to be a result of the thickness of the citrus fruit compared to the thinness of the leaves. This was verified by stacking multiple leaves and testing the reflectance as the leaf count increased. The results show ed that an increase in number

PAGE 31

31 of leaves lead to an increased diffuse refl ectance, most significantly between 720 and 1400 nm, as seen in Figure 3-6. This is explained as the first leafs transmitted light may become a second leafs reflected light. Thus, the more leaves the more opportunities there are to increase the reflectance (Figure 3-7). This theory is in agreement with Williams and Norris (2001) which preformed similar tests on thin potato slices There were no significant increases in the reflectance due to leaf layering, outside this 720 to 1400 nm range. These results are supported by Fraser et al. (2002) which show ed light penetration depths in apples was larger in the 700 to 900 nm range than 1400 to1600 nm. They claimed this to be a result of the absorption profile of water. A second important spectr al difference between fruit and l eaf in Figure 3-5 is the curve crossover occurring at near 1150 nm All citrus varie ties showed lower reflectance than the citrus leaves at higher wavelengths. 400 800 1200 1600 2000 2400 0 10 20 30 40 50 60 70 80 90 Wavelength (nm) Reflectance (%) Fruit Leaf Figure 3-5. Green citrus and leaf average spectral reflectance th rough the fall 2005 growing season.

PAGE 32

32 400 800 1200 1600 2000 2400 0 10 20 30 40 50 60 70 80 90 Wavelength (nm)Reflectance (%) Single fruit Five leaves Three leaves Single leaf Figure 3-6. Spectral reflectance of layered citrus leaves and si ngle Orlando Tangelo. Figure 3-7. How multiple leaves/surf aces create added diffuse reflectance. Reflectance Characteristics of Citrus Varieties All three of the tested citr us varieties are highly corre lated throughout the ultravioletvisible (UV-VIS) to NIR range, as shown in Fi gure 3-8. It should be explained that the difference in the visible light ra nge average, 400 to 750 nm, was a result of the fruits maturing during the experiment and the colo r brightness differences of those mature fruits. A difference in the chlorophyll to carotenoid conve rsion is the main cause of this color brightness (Merzlyak et light source increased diffuse reflectance transmission

PAGE 33

33 al., 1999). A consistent separation of citrus variety is seen in the diffuse reflectance magnitudes reflectance of the higher NIR spectral waves: Ha mlin had the highest, followed by the Valencia and finally the Tangelo showing the lowest (Figur e 3-9). Another way to view the separate of varieties is based on their individu al standard deviations as compar ed to the standard deviation of all the citrus samples, presented in Figure 3-9. It shows the greatest standa rd deviation of all the varieties to be much higher in the spectral range of 1400 to 1880 nm with an additional standard deviation range maximum around 2200 nm. Mean while, the standard deviations of the individual varieties remain low at those same wavelength ranges. This suggests that they are prime wavelengths for variety classifications information, a statement which will be mathematically proved later (Table 3-3). Note th at the highest standard deviations are at 660 nm due to the color conversion mentioned before. B ecause the standard deviation of the individual citrus varieties remains high here as well, this would not be a good location to gather classification information. The average reflectance of leaf varieties s howed a maximum less than 3% separation. Those separations are insignificant when compared with standard deviat ions on the order of 3.5%, thus the leaf variety classifica tion was not possible with our results. The reflectance curves were converted to ab sorbance spectra by use of the Beer-Lambert law, Equation 3-3 (Williams and Norris, 2001) (F igure 3-10). Although analysis for light absorbance compared to absorber is not discus sed in depth in this th esis, the reader should understand that light absorbance by materials is completely dependent on the molecular structures of the said material. More specifically the Beer-Lambert Law st ates, the concentration of an absorber is directly proportional to the sample absorbance (Williams and Norris, 2001). This fact is used throughout th e agricultural and food industry in non-destructive examinations of

PAGE 34

34 food. It should also be realized by the reader th at the different absorbance curves of Figure 3-10 mean different absorber components and quant ities exist among the citrus varieties. R A 1 log (3-3) Where, A = Absorbance and R = Reflectance 400 800 1200 1600 2000 2400 0 10 20 30 40 50 60 70 80 90 Wavelength (nm)Reflectance (%) Tangelo Hamlin Valencia Figure 3-8. Citrus varieties average spectral reflectance. 400 800 1200 1600 2000 2400 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 Wavelength (nm)Standard Deviation Valencia Hamlin Tangelo All Varieties 660 1400 1880 2200 Figure 3-9. All citrus and citr us varieties standard deviati on values for all wavelengths.

PAGE 35

35 400 800 1200 1600 2000 2400 0 0.2 0.4 0.6 0.8 1 1.2 1.4 Wavelength (nm)Absorbance (Abs) Tangelo Hamlin Valencia 660 Water absorption bands 970 1450 1940 1200 Figure 3-10. Citrus varietie s average spectral absorbance. Spectral Growth Patterns of Maturing Citrus Fruit To verify that the chemical and biological changes occurring inside the maturing green citrus fruit would not significantl y harm our identification methods, a brief survey of the spectral growth pattern was conducted. Previous re search (Merzlyak et al., 1999) conducted on the maturing process of citrus focused mostly on id entifying wavelengths signaling growth. In this research, sampling ranges that show growth are to be avoided, to prevent maturing from harming the accuracy of our identification. A solution that can identify both early season and end of season citrus offers more robustn ess and/or could be implemented other future citrus harvesting systems. Changes to reflectance are dramatically il lustrated in Figure 3-11, where early season Hamlin sets harvested on July 12, 2005 and Sept ember, 20 2005 are compared with later sets on November 2, 2005 and December 12, 2006. By mid December, 2005 Hamlin samples turned a bright orange. To design a machine vision system that identifies green or orange citrus fruit

PAGE 36

36 from green citrus leaves, the 500 to 750 nm range needs to be avoided due to these dramatic spectral reflectance changes. 400 600 800 1000 1200 1400 0 10 20 30 40 50 60 70 80 90 Wavelength (nm)Reflectance (%) 12-13-05 11-2-05 9-20-05 7-12-05 Figure 3-11. Reflectance changes ov er the growing season of Hamlin. Training and Validation Results Discriminability results of the training data set, Methods I a nd II are shown in Table 3-2. The training data shows separation between green ci trus fruit and leaves to be the strongest at 863 nm, with a discriminability value of 7.84. The strongest separati on wavelength based on discriminability that met the 100 nm distance th reshold was 763 nm. The strongest wavelength for separation above 1150 nm, where leaf reflecta nce is greater than fruit reflectance, was 1389 nm. These second and third feature wavelength s, 763 and 1389 nm, respectively, show weaker separation but are still relatively strong for classification purposes. Scatter plots of the validation data using Methods I and II ar e shown in Figures 3-12 and 313, respectively. Using a classification line pa ssing through the centroid of the training data and perpendicular to the proje ction line, all samples but one validation fruit were correctly

PAGE 37

37 Table 3-2. Discriminability results of the training da ta set, Methods I and II. identified using Method I wavelengths, yielding R2 = 0.994 (Figure 3-12). Using Method II wavelengths, all of the validation data was correctly identified, yielding R2 = 1.000 (Figure 313). It can be observed that Me thod I of the fruit to leaf se paration possesses the strongest discriminability values; however, the scatter plot of Method II displayed stronger separation in the two dimensional space, as shown in Figures 3-12 and 3-13. 50 60 70 80 90 50 60 70 80 90 Reflectance (%) at 863 nmReflectance (%) at 763 nm Leaf Fruit projection line classification line Figure 3-12. Fruit & Leaf (Method I) with Fisher projection line (solid line) and classification line (dashed line). Discriminability Using All Samples The discriminability results us ing all the samples data is s hown in Figure 3-14. It is important to notice the general trend of this graph and what regions show the strongest discriminability, 740 to 940, 1060, 1380, and 1570 to 1830 nm, and what regions have the Feature 1 Feature 2 wavelength (nm) discriminability, d wavelength (nm) discriminability, d Method I 863 7.84 763 6.72 Method II 863 7.84 1389 4.87

PAGE 38

38 40 50 60 70 80 90 10 20 30 40 50 Reflectance (%) at 863 nmReflectance (%) at 1389 nm Leaf Fruit Figure 3-13. Fruit & Leaf (Method II) with Fisher projection line (solid line) and classification line (dashed line). weakest discriminability, 1140, 1930, and 2500 nm. Although the remainder of the mathematical analysis in this Chapter is de pendent on the randomly selected training and validation sets, this figure will be referenced back to in Chapter 4 in the selection of optical equipment. The results of discriminability calculations usin g all samples are similar to the training data results, with the strongest disc riminability at 881 nm. Using Method I gives the second feature wavelength at 781 nm while Method II gives 1383 nm, as shown in Table 3-3. Notice that weak discriminability in wavelengths with high mean separations can be an artifact of the multi-modal nature inherent in the total fruit PDF, for example 1650 nm seen in Figure 3-4A. The difference between fruit variety reflectance increases the standard deviations at these wavelengths. The three peaks seen in the fru it histogram of Figure 3-4A serves as visual evidence of the three fruit varieties multi-model effect.

PAGE 39

39 400 800 1200 1600 2000 2400 0 1 2 3 4 5 6 7 8 Wavelength (nm)Discriminability Figure 3-14. Discriminability results of all the sample citrus vs. leaves. Table 3-3. Discriminability results using all the data. Feature 1 Feature 2 wavelength (nm) discriminability, d wavelength (nm) discriminability, d Fruit & Leaf (Method I) 881 7.44 781 4.98 Fruit & Leaf (Method II) 881 7.44 1383 4.92 Tangelo & Hamlin 1712 5.52 1392 5.44 Tangelo & Valencia 1417 2.96 1882 2.83 Hamlin & Valencia 1711 3.20 1813 2.89 Figures 3-15, 3-16, and 3-17 show scatter pl ots created by the feature discriminability results in Table 3-3. When observing the fruit variety plots, most signi ficantly Figure 3-15, the difficulty in using features with dependencies on each other are seen. The result is a class cluster of cigar shapes with high variances in the same direction. Incase of Fi gure 3-15, the variance is between the bottom left corner and the top right co rner of the graph. This can create difficulty Fruit and leave cross over 881 1383 781

PAGE 40

40 with class separations by mean s of PCA, due to the dependences of one feature space on the next. 12 16 20 24 28 32 10 14 18 22 26 30 Reflectance (%) at 1712 nmReflectance (%) at 1392 nm Hamlin Tangelo Figure 3-15. Tangelo & Hamlin with Fisher pr ojection line (solid line) and classification line (dashed line). 15 19 23 27 31 13 17 21 25 29 Reflectance (%) at 1711 nmReflectance (%) at 1813 nm Valencia Hamlin Figure 3-16. Tangelo & Valencia with Fisher pr ojection line (solid line) and classification line (dashed line).

PAGE 41

41 8 10 12 14 16 6 8 10 12 14 Reflectance (%) at 1417 nmReflectance (%) at 1882 nm Valencia Tangelo Figure 3-17. Hamlin & Valencia with Fisher pr ojection line (solid line) and classification line (dashed line). Fisher Linear Discriminant Analysis (FLDA) Results The solid lines in Figures 3-12, 3-13, 315, 3-16 and 3-17, are the projection lines calculated by FLDA. While several of the calculated projection lines, w, are intuitively correct like Figure 3-13, others are counter intuitive such as Figure 3-16. The pa rallel nature of the two cigar shaped class clusters is the cause of this odd looking pr ojection line. The reasoning is that the class means of the projected data on to the one-dimensional space is close in distance; however, the between-class scatter to within-c lass scatter (Duda et al., 1988) is the best for the most accurate classification, because of a lowe r variance. A look back at Figure 3-12 is an example of the same tendency for the projection line to be perpendicular to cigar shaped clusters. Mathematical definitions of the calculated pr ojection vectors for the two class systems are shown in Table 3-4. Notice the projection space is one dimensional and not defined by a line but

PAGE 42

42 rather a direction; a line was included in the fi gures only as a visual re ference. A complete scatter plot including all 270 data samples of every class is in Fi gure 3-18. The x-axis uses a feature wavelength of 881 nm as it displayed the strongest discri minability for fruit and leaf identification, while the y-axis uses a feature wa velength of 1713 nm as it displayed the strongest average discriminability between fruit varieties. Table 3-4. Projection vector equations found by Fisher Lin ear Discriminant analysis Feature 1 (x-axis) Feature 2 (y-axis) Projection vector, w Fruit & Leaf (Method I) 881 nm 781 nm w = ( -0.1309, 0.0439 ) Fruit & Leaf (Method II) 881 nm 1383 nm w = ( -0.0612, 0.0380 ) Tangelo & Hamlin 1712 nm 1882 nm w = ( -0.5119, -0.1092 ) Tangelo & Valencia 1417 nm 1813 nm w = ( -0.9040, 0.7114 ) Hamlin & Valencia 1711 nm 1381 nm w = ( -0.1751, -0.0312 ) 50 60 70 80 90 10 15 20 25 30 35 40 Reflectance (%) at 881 nmReflectance (%) at 1713 nm Valencia Hamlin Tangelo Leaf Figure 3-18. Scatter plot showing all cla sses: Valencia, Hamlin, Tangelo, and Leaf.

PAGE 43

43 Conclusions Samples of green citrus varieties and citrus leaves were harvested during the late 2005 growing season. Diffuse reflectance over th e UV-Vis and NIR range (200 to 2500 nm) was measured with a spectrophotometer. Average sp ectral reflectance curves show the opportunity to use NIR sensors and/or cameras to identify citrus fruit from citrus leaves, but also to classify different citrus varieties. Wavelength features for classifications were chosen by discriminability calculations. The selected feature spaces were used to create two dimensional scatter plots, which were then used to calculate the best proj ection line by Fisher linear discriminant analysis (FLDA). Using two-thirds of data as trai ning and one-third data as validation, an R2 of 1.0 was possible using these pattern rec ognition techniques. As expected separating green leaves from green citrus fruit proved to be more accurate than distinguishing among different citrus varieties. It has been shown in this Chapter that while using only two NI R feature wavelengths, extremely accurate green citrus fruit from green ci trus leaf identification is possible. A scatter plot of feature wavelengths 881 nm (x-axis) and 1383 nm (y-axis) projected onto a one dimensional feature space defined by the dir ection (-0.0612, 0.0380) proved to be the best mathematical method for separation when using the data gathered in the laboratory. The transferability of these critical wavelengths and mathematical methods to in-field systems have some interesting difficulties to be discussed later. This Chapters research was designed to test NIR data validity for the use in a computer vision system for counting green citrus yield in-fie ld, while still on the tree. The in-field testing of such a NIR camera based system is in the fo llowing chapter, Green Citrus Identification by a NIR Camera System (Chapter 4).

PAGE 44

44 CHAPTER 4 CITRUS INDENTIFICATION BY NIR CAMERA SYSTEM Introduction After completion of the spectral analysis of green citrus fruit and green citrus leaves, it was important to use that knowledge in the selection of optical camera equipment. The selection of wavelengths is of significance during image processing because an understanding of the expected spectral responses helps in spectral index designing and testing. Cameras Used in Citrus Yield Mapping Fruit identification using machine vision sy stems was proposed nearly forty years ago (Schertz and Brown, 1968). However, technology has been only recently advanced enough to allow researchers to investigate their usefulness more fully. Major fruit identification studies have focused mostly on apples and citrus fr uit, the most common a pplication being robot harvesting (Jimenez et al., 2000). Traditionally the visible spectrum has been used for fruit identification, which lends itself very convenien tly to non-green colore d fruits, such as red apples and orange colored citrus. The biggest issues with these camera systems have been occlusion and grouped fruit segm entation (Jimenez et al., 2000). Citrus harvesting system research has used cameras with different forms of traditional machine vision for many years with mixed results and a wide variation of algorithms (Jimenez et al., 2000). Recent research in this field of study at the University of Florida includes Annamalai et al. (2004), Chinchuluun and L ee (2006) and MacArthur et al. ( 2006). However, all of these systems utilized only the visible light spectrum as a means to decipher fruit from the surrounding green leaf canopy. This leads to problems with early season citr us identification, a time when citrus are a dark green color, the same color as the leaves. By solving this problem, precision

PAGE 45

45 agricultural techniques ca n be used earlier in the growing se ason, compounding their benefits of early information on citrus fruit yiel d, health, and in-field variability. Annamalai and Lee (2004) proposed a method to decipher green citrus fruits from leaves by their spectral differences in the near infrar ed (NIR) region. This work was extended in Chapter 3 by using a spectrophotome ter to identify critical wavele ngths that could be used to separate green citrus fruit from green citrus leaves. The objective was to test a simple nondestructive computer vision system for the identification of green citrus fruit while on-tree and during normal in-field growing c onditions utilizing the previously obtained results of Chapter 3 Materials and Methods NIR Camera, Optical Equipment, and Hardware A Merlin NIR InGaAs camera (FLIR Systems, Inc., Indigo Operations; Wilsonville, OR) was used for all image acquisitions. The spectra l range of this NIR camera is 900 to 1700 nm with the light sensitivity dropping o ff to zero at each side of this range and a peak sensitivity at approximately 1600 nm. This sensitivity curve is true of all InGaAs technology and is displayed in Figure 4-1, courtesy of FLIR, Inc. Th e Merlin NIR camera recorded 640 x 480 pixel monochromatic images (307200 pixels), saved in a two-dimensional grayscale TIFF format with 256 possible pixel values (0 to 255). Each imag e was taken with one of three optical band pass filters (1064, 1150, and 1572 nm) positioned in front of the camera lens, permitting only a thin spectral band of light to pass. Figure 4-2 show s the three optical band pass filters used during the image acquisitions of this study. Additional technical details about ea ch band pass filter are provided in Table 4-1. The two most critical wavelengths discove red and discussed in Chapter 3, for the separation of all green citrus vari eties from green citrus leaves were those of Method II (i.e., 881 nm and 1383 nm). However, the spectral range of an NIR InGaAs camera starts at 900 nm, and

PAGE 46

46 Figure 4-1. Spectral sensitivity range of InGasAs technology (Graph courtesy of FLIR Systems, Inc.). Figure 4-2. Three band pass (1150 nm, 1064 nm, and 1572 nm). Table 4-1. Optical band pass filter details. Supplier/Manufacture Model CWL (nm) FWHM (nm) Transmittance ThorLabs FL1064-10 1064.0 10.0 70% ThorLabs FB1150-10 1150.0 10.0 45% Andover 1572.0 / 20.0 275391572.0 15.0 75%

PAGE 47

47 even at 900 nm the spectral sensitivity is very low, less than 20% of the maximum sensitivity range. Referring back to Chapter 3, the waveleng th selection to best mimic the wavelength of strongest citrus fruit and leav e separation is around 1060 nm. Th is was the reasoning for the selection of the FL1064-10 model with a central wavelength (C WL) of 1064 nm. The second band pass filter, model FB1150-10 has a CWL of 1150 nm and was selected because the wavelength had a low discriminability between citrus and leaf, making it ideal for normalizing images. The final band pass filter was the 1572.0 / 20. 0 with a CWL of 1572 nm. A filter selection of 1383 nm was not made fore two important reasons. One, no filter around 1383 nm is commercially available while maintaining a transmittance of over 50% and a full width half maximum (FWHM) value of less than 50 nm. Tw o, and far the most importantly, the spectral solar intensity between 1450 and 1350 nm is low, close to zero. It is desirable to use wavelengths of higher solar intens ity to present the best lighting conditions possible (Figure 4-3). The range 1550-1590 nm provides the stronger solar in tensity that is desired. Also referring to Figure 4-3 the spectral irradiance of 1064 nm is a strong 0.65 W m-2 nm-1, while the 1150 nm remains lower but still functional with a spectral irradiance at about 0.3 W m-2 nm-1. Figure 4-4 shows spectrophotometer (Cary 500, Varian, Inc., Palo Alto, CA) transmittance results of the three band pass filters. The observation of most interest is the transmittable light bands outside each specific band pass filter design range (i.e., the 1064 nm band pass filter transmittance spikes at 800 nm, 1600 nm, and above; the 1572 nm band pass filter transmittance spikes at 1950 nm and above). Most of thes e extra band passes are limited by the NIR camera range of 900 to 1700 nm; however, the exception is the 1600 nm spike of the 1064 nm band pass

PAGE 48

48 filter. While this was of some concern, it should be noted that the solar intensity at 1064 nm is greater than that of 1572 nm. As a result, th e transmittance of the 1600 nm spike are overcome 0 0.25 0.5 0.75 1 1.25 1.5 30050070090011001300150017001900210023002500Wavelength (nm)Spectral Irradiance (W m-2 nm-1) Figure 4-3. Solar spectral irradi ance reference from the ASTM G 173-03 of direct circumsolar. This data was gathered by the American Society for Testing a nd Materials (ASTM) and government research laboratories. by the stronger solar light at 1064 nm, having only a small effect on the re sulting images. In contrast, the same transmittance levels for th e 1572 nm band pass filter remain acceptable, because the NIR camera mechanical iris can be opened allowing enough of the 1570 nm solar light in. In fact, during image acquisition the iris was always opened more for the 1572 band pass filter imaging then for the other two filters. The band pass filters were all bi-directional with the same transmittances in both dire ctions. Although this was verified by the spectrophotometer, the results shown in Figur e 4-4 are from the same direction as was consistently used throughout th e image acquisition process.

PAGE 49

49 400 800 1200 1600 2000 2400 0 10 20 30 40 50 60 70 80 90 100 Wavelength (nm)Transmittance (%) 1150 nm band pass filter 1064 nm band pass filter 1572 nm band pass filter Figure 4-4. Percent transmittance for each of the band pass filters. An in-field experimental setup used a ga soline generator to su pply a standard North American wall outlet power s upply of 120 Volts, peak-to-peak The generator powered a computer monitor, the NIR camera, and a personal computer (Pentium 4, 2.4 GHz). The NIR camera was connected by coaxial cable to the PC, which was running image acquisition software (Intellicam, Matrox Electronic Systems Ltd., Dorval Canada). The in-field experimental setup is shown in Figure 4-5. The us e of a light weight table and wagon was to facilitate the easy moving of equipment from one image location to another. Experimental Location and Environment In order to acquire multispectral images, the NIR camera was positioned in front of citrus fruit and leaves and an optical band pass filter was manually locked into place in front of the lens. Images were taken before switching to the next optical filter without moving the NIR camera. Each time a filter was positioned, seve ral images (three to seven) were obtained with the same filter. A commercial digital camera (Canon Digital Elph S300) was used to record a

PAGE 50

50 visible spectrum image of the NIR camera field of view to be used later for target scene referencing. For the purpose of th is thesis, an NIR camera field of view showing citrus fruit and leaves will be referred to as a target; a group of spectral images acquired at a target will be referred to as an image, and a raw imag e will refer to a single monochromatic image. Figure 4-5. In field experiment al setup with gasoline generato r, personal computer, monitor, NIR camera on tripod, and Teflon di sk (from foreground to background). Preliminary test images were acquired outside in a controlled environment before taking the system to the grove. This was to test th e system and handling of the optical equipment before being isolated in a citrus grove. Thirty-six images were acquired at the University of Florida Citrus Research Grove in Gainesville, FL for in-field testing of the multi-spectral imaging system. Target citrus trees were all Ham lin variety. Both the test images and in-field images were acquired in November 2006 under s unny weather conditions. A total of 552 raw images were obtained. PC Generator Monitor NIR Camera Teflon Disk

PAGE 51

51 The predetermined experimental plan included a Teflon disk or sheet to be used for raw image light intensity normalization. However, the brightness of the Fl orida sun saturated the Teflon material in the raw monochromatic images (refer to raw images of Figures 4-9A, 4-10A, or 4-14). Efforts to limit the light to the camer a by closing the NIR camera iris and/or shading the Teflon or target area limited the visibility of th e fruit and leaf targets. Raw images with the Teflon were used in the study; however, the Teflon was ignored. Image Processing Results from in-field image acquisitions were tr eated as a series of three-dimension matrix values. The xand y-axes formed the two dimens ional image while the z-axis was the number of images at the target. This z-axis is also referred to as an image block depth. The z-axis length varied among images depending on the total number of raw images gathered from the target. It was essential for this research that each image bl ock included at least three raw images: one at each of the band pass filters (1064 nm, 1150 nm, a nd 1572 nm). In every image block studied one raw image for each band was selected fo r image processing based on image clarity, alignment, and brightness. When only images of each spectral band are us ed to create an image block it is known as a spectral image block (G onzalez, et al., 2004). Figure 4-6 shows an example of a spectral image bl ock used in this study. Image processing was conducted using MatLab 7.0 software with the Image Processing Toolbox (MATLAB, 2006). Raw images were tr eated as 480 x 640 two dimensional matrices consisting of 8-bit unsigned values. All im age processing techniques and mathematical calculations preformed between raw images were computed on a matrix point by point (pixel by pixel) format.

PAGE 52

52 Figure 4-6. Representation of an image spectral block. Image processings major stages The image processing steps used in this study consisted of two major stages, each having sub-steps. The two major stages will be referred to as Image Indexing and Marker Checking. In the Image Indexing stage a likelihood image wa s calculated of each pixel being one of two classes, either citrus fruit or noncitrus fruit. It should be noted that not every pixel is either citrus fruit or citrus leaf as images include branches, sky, ground, Teflon, metal pole or other non-fruit objects; all these possible non-fruit objects are referred to as non-fruit, as it is the desire of this image processing to not count them as c itrus fruit. Likelihood images were created based on pixel values of each band pass image. Using Otsus method, a threshold value was found and each pixel was classified as either citrus fruit or non-citrus fruit. In the Marker Checking stage, a calculation from the original raw images was used to check the validity of the citrus fruit markers, or pixel groups, identified as possible ci trus fruit pixels. This calculation was based on x-axis z-axis y-axis

PAGE 53

53 the regional variances which describe the comple xity of the image marker. A more complex, thus higher variance region is more likely to be leaves or branches intertwined while a citrus fruit will be less complex. The stages wi ll now be discussed in more detail. Image Indexing stage The first image processing step of the Im age Indexing stage is a simple histogram stretching in order to maximize the reflective light intensities of the monochromatic images. This assures that 1) the maximum amount of data can be extracted from each image and 2) images to be indexed together are at about th e same image brightness level. In a histogram stretching process the maximum ( imax) and minimum ( imin) pixel value of each image is found. Then each pixel value is multiplied by an enhancement multiplier, M defined in equation 4-1. The result is rounded to the nearest whole number while remaining in the range of 0 to 255. min max255 i i M (4-1) Where, imax, imin = maximum and minimum pi xel value, respectively. This histogram stretching allows overly dark or bright images to use a wider spread of pixel values enhancing the quality of the image as displayed in Figure 4-7. This same process is used again at a later step of the Image Indexing stage. The second step of the Image Indexing stage wa s the use of an image smoothing filter. Smoothing filters, also known as image smoothi ng functions, can be very complex and powerful or quite simple. The resulting smoothed image requirements decide what type of smoothing filter should be used (Gonzalez et al., 2004). Becau se clear grove images are desirable in this research, a small 3 x 3 pixel smoothing filter using an averaging function was used. The filter is defined mathematically by equation 4-2. 24 / 8 3 3 3 3, 1 1 1 1 1 1 1 1 1 1 1 1 j i j i j i j i j i j i j i j i j i j ip p p p p p p p p P (4-2)

PAGE 54

54 Where, pi j = raw image pixel (i, j) Pi,j = resulting image pixel (i, j) Raw Image 0 50 100 150 200 250 0 500 1000 1500 2000 2500 3000 3500 4000 HistogramA Raw Image 0 50 100 150 200 250 0 500 1000 1500 2000 2500 3000 3500 4000 HistogramB Figure 4-7. Histogram stretc hing function; raw image and hi stogram A) Before histogram stretching. B) After histogram stretching. The smoothing filter performed three tasks for improving the image to image calculations: 1) removed random noise in the images; 2) allo wed image target edges to smooth the transition between the background and foreground, helping re solve situations where targets did not align perfectly between raw images; 3) extracted inform ation from saturated pixels (i.e., a pixel of 255 located next to pixels of less th an 255 was not considered as bright/saturated as a pixel totally surrounded by other saturated pixels of value 255). The critical image processing step of the Imag e Indexing stage was the actual index itself. To identify the citrus fruit, an index calculati on using all three band pass images was designed.

PAGE 55

55 Indices in remote sensing applications are the mathematical calculations of different spectral band images. Most often in agricultural applica tions these calculations are preformed with high spatial resolution images acquired by aerial or sa tellite based systems. The most widely used index is the NDVI or Normalized Difference Vegetation Index (Morgan and Ess, 2003). The use of diffuse reflectance information from Chapter 3 was used as a guide, but image investigations revealed that in-field lighting and target orient ation conditions created ve ry different reflectance values than had been expected. Many index calcul ations were attempted w ithout quality outputs. Examples include ratio indices (such as th e 1064 band pass divided by the 1572 band pass; both normalized and un-normalized), indexing by way of finding class likelihoods using Fisher linear discriminant analysis (FLDA) as described in Chapter 3, and even more complex hybrid indices (using multiple indices while solv ing for the average). When studying the common and reliable spectral indices used in agricultural application th e most widely used indices appear very simple (Apan et al., 2003). Further study of the literature inspired the design of the index used in this thesis research. The equation for th e index is shown in equation 4-3. ) / ( ) / (1150 1572 1150 1064b b b b B (4-3) Where, b1064 = images with 1064 nm band pass filter b1150 = images with 1150 nm band pass filter b1572 = images with 1572 nm band pass filter This green citrus index functions by solvi ng the differences between the reflectance of leaves and citrus fruit at 1064 nm and 1572 nm, but first the images are normalized by the 1150 nm band where the leaves and citrus fruit reflectance are very similar. This normalization eliminates the effects of dark images of one ba nd being indexed with light images of the other band. This index does not represent the ideal segmentation system that was calculated in

PAGE 56

56 Chapter 3 but still uses an interpretation of it. The higher reflectance of the citrus fruit versus citrus leaves around 800-1100 nm is directly compared against the lower reflectance of the citrus fruit versus citrus leaves around 1500-1800 nm. Two critical band pass filt ers are used, as was desired and a third band pass filter normalizes the image shadows and bright reflections. The resulting two-dimensional image from the green citrus index can be understood as a pixel likelihood for each class, where high values ar e citrus fruit and low values are non-citrus fruit. One problem remaining in this likelihood image is that misaligned images can result in pixel calculations being too high; therefore an upper value cut-off was set to 165. This value selection was based on training imag e histograms. The pixels that exceeded this upper value of 165 were by default set to zero, therefore eliminati ng them from being classi fied as citrus fruit pixels. A fine example of this can be seen in the index images (Figure 4-8B and Figure 4-9B), where the brightest spots are not the fruit but the edges of objects which are not properly aligned. The histogram stretching function was used again after this upper value cut off to maximize the pixel value separation for the next thre shold step utilizing Otsus method. Otsus method automatically chooses a thres hold based on the histog ram of a grayscale image (Otsu, 1979). The method is based on discrimi nant analysis, as discussed in Chapter 3. The objective of such an analysis is to sepa rate each pixel into one of two classes, C0 or C1, based on a threshold value, t Assume C0 represents lower pixel values {0, 1, 2, 3, t } and C1 represents higher pixel values, {t+1, t +2, t +3, L-1}, where L is the number of grayscale values. Let 2 b and 2 d be the between-class variance and w ithin-class variance, respectively. The ideal separation threshold can then be obtained by maximizing the separation of the histogram. The objective is to maximize the ra tio of the between-class to within-class with respect to the threshold level, t This is found by stepping thr ough all intensity levels of the

PAGE 57

57 grayscale image and at each level calculate 2 b/ 2 d. The higher this ratio is, the stronger the pixel class separation. MATLAB 7.0 has an Otsu function built into the image processing toolbox (MATLAB, 2006). The function returns only th e recommended threshold value and has no effect on the original grayscale image. Otsu s value can be multiplied by a scaling factor lowering or raising the threshol d and provided an opportunity to fine tune the method for this citrus image application. The training imag es showed better classification results with a multiplier value of 1.5. This is due to the citr us images needing a threshold for only the highest valued pixels, as most pixels were non-citrus fruit pixels. The final step of the Image Indexing stage was to improve the binary image, removing holes and noise using a ten pixel si zed 'disk' for erosion then dila tion, also referred to as opening (Gonzalez et al., 2004). The selec tion size of ten pixels was a choice based on the training images. It is feasible to use a larger disk shape and this would improve some results by clearing out medium sized errors, however some images would have all the possible fruit pixels eliminated as well. Again, the MATLAB im age processing toolbox has a function for this morphological process (MATLAB, 2006). The comp lete step by step process of the Image Indexing stage is shown in Figures 4-8 and 4-9. Marker Checking stage Examination of the final images reveal severa l extra large and small fruit markers, which was defined as groups of possible ci trus fruit pixels in the images. If each marker signals a citrus fruit location in the image, then most images have too many such markers. The Marker Check stage is a means of going back to the 1150 nm band pass image and checking if the marker in the final image makes sense. This is accomplished by treating each marker as a separate feature and checking it individually based on th e variance of the pixels in the marker with respect to the number of pixels making up the fruit marker The 1150 nm band pass image is not the raw

PAGE 58

58 A B C D E Figure 4-8. Image Indexing stage of validation image number eigh t. A) Three band pass images 1064 nm, 1150 nm and 1572 nm, respectively. B) After smoothing, histogram stretch and the index calculation steps. C) After upper limit cut off and histogram stretching. D) After thresholding using Ot su's method. E) Final binary image after the use of a 10 pixel erosion and dilation disk.

PAGE 59

59 A B C D E Figure 4-9. Image Indexing st age of training image number 11. A) Three band pass images 1064 nm, 1150 nm and 1572 nm, respectively. B) After smoothing, histogram stretch and the index calculation steps. C) After upper limit cut off and histogram stretching. D) After thresholding using Ot su's method. E) Final binary image after the use of a 10 pixel erosion and dilation disk.

PAGE 60

60 image, but saved after being histogram stretc hed and smoothed in the Image Indexing stage before being used again in this stage. This pr events noise from negatively affecting the variance calculation results. The first step of the Marker Checking stage consisted of a simple separation of the markers based on connectivity. If a nonbreaking line could be drawn ar ound the marker and connected to the starting point of the line without crossing over another marker or being inside another marker, the enclosed marker would be identified. The next step calculated the distance between the highest and lowest pixel valu es of the marker and divided it by the total number of pixels, n in the marker. The result is defined as the mark er value for that marker, defined in equation 4-4. n p p MV / ) (min max (4-4) Where, MV = marker value pmax, pmin = maximum and minimum pixel values, respectively n = number of marker pixels Markers with an MV value greater than 0.2 were assu med to be non-citrus fruit markers and were eliminated from consideration. The MV value used in this step was chosen based on the trial and error and best results from the traini ng images. It can be interpreted that for every 100 pixels of the marker, a 20 pixel level ch ange was allowed wit hout the marker being removed. The MV used was for image processing can be considered conservative, as it permitted some small non-fruit marker errors to be kept rath er than risk removing a la rge correct citrus fruit marker. This major stage also filled any holes in the citrus fruit markers, such as the one seen in Figure 4-10. This MV concept was based on the major assumption that leaves, branches, Teflon sheet edges, and other non-citr us fruit objects are more likely to have edges appear in the markers. These edges will have a higher variance in pixel values ranging from the shadow to a

PAGE 61

61 bright glare within a short range. Citrus fruit objects on the othe r hand tended to have a smoother and less severe rapid change in pixel va lue. Additionally, because of the method in A B C D Figure 4-10. Marker Checking stage of validatio n image number nine. Input binary image of possible fruit markers (A). The 1150 nm ba nd pass image (B). Markers and the pixel value intensities (C). The new binary image output after removal of most non-fruit markers and filling of marker hole (D). Of the five output fruit markers four are correctly identified as citrus fruit. which the MV is calculated, only smaller markers are at real risk of being eliminated. This means correctly identified citrus markers that in clude both the dark shadowy edge and the bright sunny spot of the citrus fruit would not be rem oved as long as the marker size is sufficiently large. On the other hand, small markers are at great risk of being removed and sometimes even correct markers were removed, as in the Figure 4-10. The small marker on the bottom left is incorrectly removed by the Marker Checking sc heme, however, an incorrect marker was

PAGE 62

62 Marker Checking stage removed as well. This type of scheme balances the pros and cons to make images improvements, and for most of the training images the positives outweigh the negatives. The complete image processing routine involved two major stages, ten sub-steps, and three raw image inputs to produce one final output image. This included variables that were fine tuned by the training image data set. Variables included: smoothing filter type, Otsus method multiplier of 1.5, upper value cut off of 165, dilatation and erosion disk size, and MV limit of 0.2. Slight changes to any one these variables would result in different binary image results. Figure 4-11 is a flow diagram of the complete image processing major stages and sub-steps. Figure 4-11. Complete image processing procedure. Result analysis Of the 36 in-field images, two-thirds were randomly selected and us ed as training data, while one-third was used as validation (i.e., 24 training and 12 valida tion images). Training images were used in this study to permit the testing of image processing techniques on some images while not being the same images final re sults would be taken from. This prohibits the marker cleanup and removal Image Indexing stage histogram stretch marker value calculation ( MV ) histogram stretching marker identification erosion and dilation threshold (Otsus method) upper value cut off green citrus index smoothing filter

PAGE 63

63 image processing steps from being overly designed and conditioned to the validation images. At the conclusion of the image pr ocessing steps each validation image was checked for three measures of accuracy; percent of correct pixels, fruit pixel co unts and number of markers. Percent of correct pixels and the fruit pixel counts were measured against manually produced image masks of each validation image showing the citrus pixels (Figure 4-12). The measure of fruit markers is compared to the number of fruit seen by human observation in the images and the number of fruit markers in the manually produced image masks. A B Figure 4-12. Manually produced image mask showi ng the difference between citrus fruit pixels and non-citrus fruit pixels. Percent of correct pixels were measured on how many pixels of each image were correctly classified and how many were in correctly classified. The resu lting classifications were judged using Bayes error rate, define d in equation 4-5 (Duda et al ., 1988). There are two possible misclassifications: either a pixel x falls in R2 (non-citrus fruit) while its true state of nature is w1 (citrus fruit) or the pixel x falls in R1 (citrus fruit) while its true state of nature is w2 (non-citrus fruit). Otherwise, the classi fication is defined as correct. )] ( ) ( [ 1 ) ( 1 ) (2 1 1 2w R x P w R x P error P correct P (4-5) Where, R1, R2 = regions 1 (citrus fruit) and 2 (non-citrus fruit) w1, w2 = classes 1 (citrus) and 2 (non-citrus fruit)

PAGE 64

64 The measure from Bayes error rate (equation 45) is out of a probability factor, meaning a perfect 100% rate would be 1.0, wh ile a 50% rate is valued 0.5. In the results section, these unit rates will be discussed in per cent correct and percent error. Image processed fruit pixel counts were meas ured with respect to the manually masked fruit pixel counts. For each of th e 12 validation images used in th is study a count of citrus fruit pixels was made from the resulting processed image and the manually masked image. An R2 value was calculated for a linear approximation relating the two measures. This is a very important measure as it provides a quantitative value of the ability to predict the density of green citrus in each image using this method. This is a methodology future researchers could use in determining green citrus yield. Lastly, a survey was conducted regarding th e number of citrus fruit markers found, manually counted number of citrus fruit, and the masked image citrus fruit markers. This is important as it has been the measuring tool of previous research for on-tree citrus fruit identification (Annamalai et al., 2004; Chinchuluun and Lee, 2006). The results conclude with a discussion about the value of such marker count information. Results and Discussion Image Acquisition Problems There were many problem that appeared ea rly during the in-field image acquisition process, but there was very little that could be done at the time. These problems and their effects on the results will be discussed before the qualitativ e results section. The most important issues faced with this experimental design were targ et shifting, light/shadow changes, raw image saturation and multiple leaf reflections incr easing the expected diffuse reflectance. During the time it took to change out one optical band pass filter for another (30-45 seconds) slight changes could occur in the uncon trolled environment. The two most prevalent

PAGE 65

65 changes were what is referred to as target shifting and light/shadow alterations. While one image may be taken under ideal conditions, subsequent nonideal conditions could follow closely. This includes the biggest in-field prob lem of strong breezes shifting th e leaves and/or branches and even swinging the target fruit(s) between image acquisitions. Without the fruit being perfectly aligned from one band pass image to another, incorrect edge classification will occur. Figure 413 is an example of a target fruit and surrounding branches and leaves shif ting as a result of wind turbulence. The two images were taken with th e same band pass filter only two or three seconds apart, however, the location of th e fruit and leaves shifted dramatically. This problem can not be resolved by a simple shifting of the image to align the fruit, because not all the objects are shifted by the same amount. For example, the leaves ju st above the Teflon sheet in Figure 4-13 are at approximately the same location in both images. These means a dynamic shifting would need to be used, stretching and compressing different parts of the images to better align them. Such a complex process would become increasingl y difficult on such a complex background. Figure 4-13. Location of the target fruit and some of the leaves has shifted between the two raw images. The Teflon sheet and some leaves remain in the same place. In addition to wind moving the target a nd the resulting shadows, shadow and sun illumination changed during image acquisitions du e to inconsistent cloud cover and the slow shifting of the sun. These shifts in the suns lo cation in the sky creates more than a change in

PAGE 66

66 shadow location. The shift cha nges the solar angle which may aff ect the magnitude of resulting fruit and leaf reflectance (Gilabert and Melia, 199 3). Because raw images at each target location were taken within a short period of time thes e solar angle effects were not considered in individual images blocks, however it may be a c oncern if future systems require images to be acquired all day. The only solution used for this research during image acquisition was to take many images quickly and select the images that facilitated better image to image comparisons based on target fruit and/or leaf alignment. Another major problem during image acquisition was raw image saturation. Many of the images became too dark or too bright. This was most prevalent on the outer and inner layers of the citrus canopy where lighting conditions could vary dramatically. Figure 4-14 shows one example of the light saturation among the leaves and part of the citrus fruit. This was a critical issue as information at these pixels was lost a nd comparisons to the refl ective behaviors of the other wavebands would not be correct. The Te flon sheet in several of the images became the most susceptible to this problem. This is why the use of the Teflon sheet was suspended when it became apparent that the material was too reflec tive to be used as a normalization material in sunlight. Figure 4-14. Lighting condition extremes (blackout and saturation) from inside and outside the citrus canopy.

PAGE 67

67 The effects of the saturated pi xels are dramatically observed in the histograms; Figure 4-15 shows two images of the same targets using the same band pass filter. They were taken moments apart when the strength of sunlight changed dramatically due to sudden cloud cover. The histogram on the left has around 30000 pixels satu rated with a value of 255. The histogram on the right has only about 700 pixels saturated. Comparing these va lues to the to tal number of pixels per image (307200 pixels), a loss of almost 10% of the pixels is experienced by the image on the left as apposed to less than one quarter of one percent on the right. During the image acquisition process the mechanical iris of the NIR camera was not precis e enough to correct all lighting problems. Figure 4-15. Changes in sunlight and the effect on image histograms. 0 50 100 150 200 250 0 500 1000 1500 2000 2500 3000 3500 0 50 100 150 200 250 0 0.5 1 1.5 2 2.5 3 3.5 4 x 104 30000 pixels 700 pixels

PAGE 68

68 The last issue faced in the experimental setup was foreshadowed in an experiment discussed in Chapter 3. The diffuse reflectan ce of multiple leaves stacked together were measured with a spectrophotometer (Cary 500, Vari an, Inc., Palo Alto, CA) and revealed as the number of leaves increased so did the percent of diffuse reflectance. It was theorized that additional leaves became a second, or third, refl ective surface for light wa ves that passed through the first leaf. After reflecting off the second (or third) leaf, the light w ould transmit back through the first leaf again, and be detected as light reflected from the first leaf. The acquired NIR camera images show a similar phenomenon as the 1064 nm band images showed only a minor intensity difference between fruit and leaves. In addition, leaves in the NIR camera images were not facing flat but had more variations in direct ion, which changed the reflectance among leaves and altered the expected reflectance from Chapter 3. Quantitative Identification Using the image processing methods and variable values described previously, the average correct pixel classification was 90.3% with resp ect to the manually masked images. Correct pixels were any pixels that ha d the same classification in the masked image and image processed image. The total number of correct pixels was then divided by the total number of pixels, 307200, to find the percent correct. This is presen ted in Table 4-2, with a fruit pixel count from the image processing and the masked images. Notice, being able to predict the pixels classification may not the best measure of identification. This will be discussed in more depth in a little bit. The image processing citrus pixel counts and ma nually masked citrus pixel counts are also presented in Figure 4-17. Each show the complete number of fruit pixels c ounted in the results. This count and be interpreted as a type of yield data set for each image. While the sizes of the fruits in each image are not always the same a nd the distances of the cam era to the targets are

PAGE 69

69 variable, the total number of citrus fruit pixels does provide some information. The ability of the image processing algorithm to detect this can be studied using a t-te st and solving an R2 for the two data vectors. Table 4-2. Validation image citrus fruit pixel results. Validation image number Correct pixel classification (%) Image processing fruit pixel count Manually masked fruit pixel count 1 94.8 1524 16624 2 96.9 9753 18716 3 77.5 35756 33861 4 96.1 1699 13084 5 81.9 23507 63607 6 95.1 5936 17109 7 82.3 4029 36512 8 92.9 11065 31841 9 89.5 11001 38565 10 97.9 1207 2733 11 88.1 12743 36706 12 90.2 9550 31901 average 90.3 9812.3 27538.3 Using a two sided t-test to assess the signi ficance of the error percent when comparing image processed fruit counts with manually masked fruit counts, in this case a highly significant difference was found. The p-value was solved to be 4.295x10-4 making it far too small, which shows the null hypothesis to be in valid for this data set. In ef fect, the t-test results are saying the image processing method was unsuccessful, but only if it is assumed that the manual masking method is defined as a correct reference. Ag ain, the validity and limited usefulness of the manually masked images for checking will be discussed in more depth later. The two sided t-test method also tested for the equality of variances in the two data sets. The probability was solved as being greater than 0.10, meaning the t-test was valid for the null

PAGE 70

70 hypothesis of equal variances. The calculated variance for the image processing vector is 1.0190x108 while being 2.5317x108 for the manually masked images. In addition, the t-test revealed the t-statistic is 4.9590 on 11 degrees of freedom. Using the image processing techniques and variables discussed previously, the image processing results underestimated the citrus pixe l counts for all validation cases. This makes sense when observing the image results in Figur es 4-19, 4-20, and 4-21, but can be proven by looking at the data presented in Figure 4-16. A list of four classes based on image processing pixel classification and manually ma sked pixel classification is give n for all 12 validation image. The image processing pixel classification (vertic al location) and the manually masked pixel classification (horizontal location) define which corner the pi xel count adds to. This means the top left corner are fruit pixels correctly classified as fruit pixe ls, bottom left corner are non-fruit pixels incorrectly classified as fruit pixels, top ri ght corner are fruit pixels incorrectly classified as non-fruit pixels, and bottom left corner are non -fruit pixels correctly cl assified as non-fruit pixels. A total for all the validation images is given at the bottom. There are two important observations to pull from this figure. First, the great numbers of pixels are classifi ed by the image processing as non-citrus fruit and are manually masked as non-citrus fruit. This is seen by how much larger the value in the non-citrus/non-citrus bo x is in the bottom right corner. This facilitates for a high percent correct, as in Table 4-2, but does not display true identification. The second important observation in Figure 4-16, is the relative values of image processing classified non-fruit that are masked as fruit (bottom left corner), versus the image processing classified fruit that are masked as fr uit (top left corner). The lower the bottom left corner value relative to the top left corner, th e more correct the identification without improper

PAGE 71

71 classification. The best examples of this are validation image numbers one, two, four, seven, nine, eleven and twelve. All these validation images are examples of the image processing finding the citrus fruits but did not completely covering the mask ed area. For a good example of this, refer to Figure 4-20. The shadowed edges and bright glare of the sun eliminated parts of the fruit from being classified correctly. The imag e processing algorithm is conservative with the selection of citrus fruit pixels as a way to safe guard against classi fying non-citrus fruit pixels as citrus fruit pixels. Figure 4-16. Pixel classification results for all va lidation images and the overall total. (Notice, top left corner are fruit pixels correctly cla ssified as fruit pixels, bottom left corner are non-fruit pixels incorrectly cla ssified as fruit pixels, top ri ght corner are fruit pixels incorrectly classified as non-fruit pixels, and bottom left corner are non-fruit pixels correctly classified as non-fruit pixels.)

PAGE 72

72 The R2 of all the fruit pixel counts is 0.390 (Figur e 4-17). However, it is evident from the graph that the point (35756, 33861) is an outlier. This even more evident from the poor performance of this validation image number three, which is presented in detail in Figure 4-21. If this validation image is dropped, a fruit pixel to fruit pixel R2 of 0.765 was calculated (Figure 4-18). Also the narrowing of the 90% conf idence prediction bounds displays this same significance of the outliner removal. The R2 results would be higher if the calculations come from non-citrus pixel counts as most of the images were of non-citrus pixels. A major question that this research has shown is that of, what is identification? Is the best measure of identification the corr ect number of pixel counts or the number of markers (or fruit) counted? The problem with countin g pixels that are correct or in correct lies in an example where no citrus fruit pixels are identified by the imag e processing system. If image processing result predicted zero citrus pixels, the answer would s till be over 50% correct for these images used as no image in this study had citrus pixels covering mo re than half of the image area. But there are also problem with the use of citrus fruit marker s to determine the number of citrus fruit, which has been the standard method in previous citr us identification scheme s (Annamalai et al., 2004; Chinchuluun and Lee, 2006). While justification of this counting method can be made due to the distance of the camera from the canopy and the resulting number of citrus per image (10 to 50), it should be noted that a citrus pixel marker may not be accura te for other conditions. Close images of a citrus canopy generally lead to problems of over or unde r estimation of fruit. In the first example, the view of one fruit can be obstructed by a branch. The image processing might work properly and identify the two visible sides independently, thus two markers for one fruit. The second example is a great number of fruits co llected together and touching while on the tree. In this case, only one marker is created for tw o or more citrus fruits. These are pertinent

PAGE 73

73 concerns for a relatively close ca mera-to-canopy experiment when onl y a few fruits are visible in each image. 0 0.5 1 1.5 2 2.5 3 3.5 4 x 104 0 1 2 3 4 5 6 7 8 x 104 Image processing pixel countManually masked pixel count Pixel counts Linear fit (R-squared = 0.390) Prediction bounds (90% confidence) Figure 4-17. Manually masked frui t pixel count versus image proces sing fruit pixel count results with linear best fit using all pixel count data. 0 0.5 1 1.5 2 2.5 x 104 0 1 2 3 4 5 6 7 8 x 104 Image processing pixel countManually masked pixel count Pixel counts Linear fit (R-squared = 0.765) Prediction bounds (90% confidence) Figure 4-18. Manually masked frui t pixel count versus image proces sing fruit pixel count results with linear best fit with one outlier removed.

PAGE 74

74 The results from the validation image fruit ma rkers and human fruit counts are listed in Table 4-3. Fruit marker results reveal a str onger correlation between th e image processing fruit markers and the actual human observed fruit count. This is primarily due to the second example discussed above; affecting validation images number five and nine the most. There is one case of example one in validation image number four The image processing fruit marker counts do follow the human observation, but this might only be a result of noisy outputs at coincidentally correct image numbers. There are cases when the extra fruit markers harm the results, such as validation images six and seven. Table 4-3. Validation image fruit mark er and manually masked marker results. Validation image number Image processing fruit marker count Manually masked marker count Human observed fruit count 1 1 1 1 2 1 1 1 3 5 3 4 4 1 2 1 5 8 3 9 6 4 1 1 7 4 2 2 8 1 1 2 9 5 4 7 10 2 1 1 11 3 1 1 12 4 2 5 Three complete image processing and resul ting classifications and misclassifications images are given (Figures 4-19, 4-20, 4-21). Th ese three image results re present typical results in both the training and validation sets. Figur e 4-19 shows a very complex grouping of citrus fruit with only shadows in the background. The raw images of the 1064 nm and 1150 nm band pass filters suffer from glare on the right hand side This creates the incorrect fruit markers on

PAGE 75

75 the right hand side. Six of the nine fruit are marked, with three unmarked, and two image processing fruit markers being incorrect. It is difficult for multiple green citrus fruit to be properly identified together as they never have the same light illumination. This creates unequal ratios between the two key band pass filters, 1064 nm and 1572 nm, which always leave some fruit out or too many non-citrus fr uit inside the thresholding boundary. Figure 4-20 shows a very common single fruit ma rker identification. In this case the Marker Checking stage is not even necessary. What is interesting about this image is the bright sun spot on the citrus fruit because the loss of da ta due to saturation at this point prohibits the region from being properly classi fied as fruit. This C shape is quite common and can be closed up if desired. Such an algorithm was not included in image processing because occasionally, as was observed in the training data, a very large C can occur when multiple fruit are touching. In that case, closing up the C is not desirable. Because of the commonality of this type of image result, the average number of image processed fruit pi xels is about one-third that of the masked fruit pixels. This same pa ttern can be observed in Figure 4-16 with validation image numbers one, two, four, se ven, nine, eleven and twelve. Finally, Figure 4-21 reveals a case of extreme in correct fruit identification. This is not a result of the image processing algorithm, but the poor quality of images acquired. In this case the band pass at 1572 nm is not dark enough. This makes the raw images of the three band passes too similar and no multispectral information can be obtained. Instead, the green citrus indexing reveals a sight ri nging effect which is the result of placing optical ba nd pass filters in front of the focal lens of the NIR camera. Conclusions This Chapter tested a ground based multispect ral image acquisition and image processing system for identifying green citrus fruit against a green citrus leaf canopy. The image acquisition

PAGE 76

76 A B C D E F G Figure 4-19. Complete image processing and results check of validation im age number five. A) Three band pass images 1064 nm, 1150 nm and 1572 nm, respectively. B) Green citrus index, upper limit cut off and histogram stretch. C) Thresholding using Otsus method. D) Possible fruit markers and th e pixel value intensit ies. E) Marker Checking output binary image. F) Manually produced mask of citrus locations. G) Pixel classifications and misc lassifications. Notice, six of the eight fruit are marked, with three unmarked, and two image processing fruit mark being wrong.

PAGE 77

77 A B C D E Figure 4-20. Complete image processing and results check of validation image number two. A) Three band pass images 1064 nm, 1150 nm and 1572 nm, respectively. B) Green citrus index, upper limit cut off and histogram stretch. C) Thresholding using Otsus method. D) Marker Checking output binary image. G) Pixel classifications and misclassifications.

PAGE 78

78 A B C D E F G Figure 4-21. Complete image processing and resu lts check of validation image number three. A) Three band pass images 1064 nm, 1150 nm and 1572 nm, respectively. B) Green citrus index, upper limit cut off and histogram stretch. C) Thresholding using Otsus method. D) Possible fruit markers and th e pixel value intensit ies. E) Marker Checking output binary image. F) Manually produced mask of citrus locations. G) Pixel classifications and misclassifications.

PAGE 79

79 was completed in a Florida citrus grove. A hi ghly sensitive NIR camera was used in conjunction with three optical band pass f ilters (1064 nm, 1150 nm, and 1572 nm). The resulting images were separated into 24 training images and 12 validation images. A complex image processing scheme was designed involving indexing of the band pass filter images. Measurements of correctness were based on fruit pixel identifications with respect to a manually produced fruit pixel mask. Quantitative results showed that correct pixe l class identification of the 12 validation images was 90.3% when using the image processi ng algorithm and variable described. When comparing citrus pixel count from the image pro cessing algorithm to the manually masked citrus pixel count, an R2 of 0.764 was achieved with the removal of one outlier. There were a great number of problems during image acquisition that ha mpered the research in later stages of data analysis. These problems included target shif ting, lighting conditions, cloud cover, wind, band pass filter removal, band pass f ilter alignment, and user mistakes. All of these problems, however, can be summarized as a result of an uncontrollable environment. Despite these problems, the results do prove the effectiveness of the concepts presented in this study. Green citrus fruit can be identified based on spectra l differences from leaves by an NIR camera. However, there are still many refinement s that require further investigation. The most important knowledge gathered in th is study was a first ha nd account of technical issues faced and how they might be overcome in future systems. The most import requirements for improved results include the system being capable of acquiring multiple waveband images simultaneously. This would resolve almost all th e environmental issues mentioned previously. A second problem researchers would need to so lve is lighting conditions. The sun is not a

PAGE 80

80 reliable and consistent light s ource. It moves through the day, changes intensity based on cloud cover, and cannot penetrate d eep into the citrus canopy. This research goal of this Chapter was to design and test a non-destructive NIR camera based computer vision system for identifying and c ounting green citrus yield in-field while still on the tree. Results have proved this is possible, but still difficult due to environment conditions and technical difficulties. The last Chapter summarizes this thesis research and discusses what it means for NIR sensing systems in the future.

PAGE 81

81 CHAPTER 5 CONCLUSIONS AND FUTURE WORK Conclusions of Research Objectives Near infrared (NIR) sensing t echnology is a promising, rapidly affordable, and available tool for precision agriculture syst ems. The non-destructive nature of spectral-based sensing is welcomed by growers who would rather not sacr ifice their products. This research has shown that NIR sensors can be used to separate green citrus fruit and green citrus leaves by means of spectral reflectance in the NIR range. This res earch has proven through the use of training and validation sample sets that perfect separation is feasible using only tw o wavelengths features (881 nm and 1381 nm). These results were found through the use of the Fisher linear discriminant analysis (FLDA) algorit hm for multi-dimensionality breakdown. This research has also proven that correct iden tification of citrus fru it is possible using NIR optical equipment. In this research an NIR monochromatic camera was outfitted with three band pass filters (1064 nm, 1150 nm, and 1572 nm). Im age processing and multispectral indexing was used to classify the image contents pixel by pixel. By using 24 training images an image processing algorithm was designed and tested. Tw elve validation image results showed a 90.3% correct pixel classification (citrus fruit or non-ci trus fruit) obtained. In addition, the image processing scheme provided an R2 of 0.746 for fruit pixel counting versus a manually masked fruit pixel count with one outlier data being disposed. Despite these positive accomplishments, this research has shown that better image processing methodologies need to be explored, additional feature NIR spaces shoul d be considered, and the high cost of NIR camera and optical equipment needs to decline for this technology to thrive. Also uncont rollable environmental issues need to be thought through and planned for to improve upon these results.

PAGE 82

82 Future Work Highly accurate spectroscopic systems are not currently easy to tr ansport or use in demanding environmental conditions, such as a Flor ida citrus grove. When smaller and cheaper in-field spectrophotometer systems become more av ailable, the developmen t of nutrient variation maps based on both leaf and fruit samples will be possible. This increase of grove information will provide growers more opportunities to impr ove their management techniques on grove sitespecific basis. Multispectral NIR cameras can be used for the id entification of green citrus fruit in a citrus grove. It is not hard to predic t that future research will be more accurate and capable of counting fruit while on the move. Combinations of this system with GPS receivers could create on-the-go citrus yield maps early in the season. This drea m system is still a long wa y off in the future, but this research opens the door to possible uses of remote NIR sensor systems in the Florida citrus industry. It is conceivable that in the future, autonomous VIS-NIR remote sensor guided vehicles could be used to find, count, map, and even test citrus fruit for health and nutrient content, as some research suggest s, while in the grove. Such a fu turistic system is the ultimate goal of precision agriculture research not only in the Florida citrus i ndustry but all modern agricultural industries.

PAGE 83

83 LIST OF REFERENCES Alchanatis, V., Cohen, Y., Cohe n, S., Moller, M., Meron, M., Tsipris, J., Orlov, V., Naor, A., Charit, Z., 2006. Fusion of IR and multispectral images in the visible range for empirical and model based mapping of crop water status. ASABE Meeting Paper No. 061171. ASABE, St. Joseph, MI. Aleixos, N., Blasco, J., Navarron, F., Molto, E., 2002. Multispectral insp ection of citrus in real-time using machine vision and digital signa l processors. Computer s and Electronics in Agriculture 33, 121-137. Annamalai, P., Lee, W.S., 2004. Identification of green c itrus fruits using spectral characteristics. ASAE Meeting Pape r No. FL041001. ASAE St. Joseph, MI. Annamalai, P., Lee, W.S., Burks, T.F., 2004. Color vision system for estimating citrus yield in real-time. ASAE/CSAE Meeti ng Paper No. 043054. ASAE St. Joseph, MI. Apan, A., Held, A., Phinn, Markley, J., 2003. Formulation and assessment of narrow-band vegetation indices from EO-1 Hyperion Imagery for discriminating sugarcane disease. Proc Spatial Sciences Conf., 1-13. Chinchuluun, R., Lee, W.S., 2006. Citrus yi eld mapping system in natural outdoor scenes using the watershed transform. ASABE Paper No. 063010. ASABE, St. Joseph, MI. Chinchuluun, R., Lee, W.S., Ehsani, R., 2007. Citrus yield mapping system on a canopy shake and catch harvester. ASABE Meeting Paper No. 073050. ASABE, St. Joseph, MI. Citrus Research and Education Center ( CREC), 2007. Precision agri culture history in Florida. Available at: http://www.crec.ifas.ufl.edu/crec_websites/precisi on_agriculture/history.h tm. Accessed: March 25, 2007. Ding, P., 2005. Use of Nondestructive Spectrosc opy to Assess Chlorophyll and Nitrogen in Fresh Leaves. Unpublished Ph.D. Dissert ation. Oregon State University, OR. Duda, R.O., Hart, P.E., Stork, D. G., 1988. Pattern Classification, 2nd ed. John Wiley and Sons, New York. Fisher, R.A., 1936. The use of multiple measur ements in taxonomic problems. Annels of Eugenics 7, 79-88. Florida Agricultural Statisti cs Service (FASS), 2005. Citrus 2004-05 Summary. Available at: http://www.nass.usda.gov/fl/rto c0ci.htm. Accessed: May 3, 2006. Fraser, D.G., Jordan, R.B., Kunnemeyer, R ., McGlone, V. A., 2002. Light distribution inside mandarin fruit during internal quality assessment by NIR spectroscopy. Postharvest Biology and Technology 27, 185-196.

PAGE 84

84 Gaffney, J.J. 1972. Reflectance properties of ci trus fruits. Transaction of the ASAE. 15, 310-314. Gilabert, M., Melia, J., 1993. Solar a ngle and sky effects on ground reflectance measurements in a citrus canopy. Re mote Sensing Environment 45, 281-293. Gonzalez, R.C., Woods, R.E., Eddins, S.L ., 2004. Digital Image Processing Using MATLAB. Pearson Prentice Hall, Upper Saddle River, NJ. Grift, T., Ehsani, R., Nishiwaki, K., Crespi C., Min, M., 2006. Development of a yield monitor for citrus fruits. ASABE Meeti ng Paper No. 061192. ASABE, St. Joseph, MI. Gumz, M., Weller, S.C., 2005. Using remote se nsing to differentiate weeds in mint. Top Farmer Crop Workshop Newsletter, April. Jimenez, A.R., Ceres, R., Pons, J.L., 2000. A survey of computer vision methods for locating fruit on trees. Transacti on of the ASAE 43 (6), 1911-1920. Labsphere, 2006. A Guide to Integrating Sphere Theory and Applications. Labsphere Inc., North Sutton, NH. Available at: http://www.labsphere.com/data/userFiles/ A%20Guide%20to%20Integrating%20Sphere%20The ory&Apps.pdf. Accessed: Nov 5, 2006. Lee, W.S., Slaughter, D.C., 2004. Recognition of partially occluded plant leaves using a modified watershed algorithm. Trans action of the ASAE 47 (4), 1269-1280. MacArthur, D.K., Schueller, J.K ., Lee, W.S., Crane, C.D., MacArthur, E.Z., and Parson, L.R., 2006. Remotely-piloted helicopter citrus yield map estimation. ASABE Meeting Paper No. 063096. ASABE, St. Joseph, MI. MATLAB, 2006. Learning MATLAB 7 Release 14. MathWorks Inc., Natick, MA. Merzlyak, M.N., Gitelson, A.A., Chivkunova, O.B., Rakitin, V.Y., 1999. Non-destructive optical detection of pigment ch anges during leaf senescence a nd fruit ripening. Physiologia Plantarum 106, 135-141. Min, M. 2006. Spectral-Based Nitrogen Sensing for Citrus. Unpublished Ph.D. Dissertation. Universi ty of Florida, FL. Morgan, M., Ess, D., 2003. The Precision-Farming Guide for Agriculturists. John Deere Publishing, Moline, IL. Nagy, S., Shaw, P.E., Veldhuis, M.K., 1977. C itrus Science and Technology. Vol. 1. Avi Publishing Co., CT. Otsu, N., 1979. A threshold selection me thod from gray-level histograms. IEEE Transactions on Systems, Man and Cybernetics 9 (1), 62-66.

PAGE 85

85 Regunathan, M., Lee, W.S., Annamalai, P., 2005. Green citrus fruit identification using spectral characteristics. 5ECPA-2E CPLF Book of Abstracts, 246-247. SAS, 2006. SAS 9.1 Software Documentation a nd Examples. SAS Institute Inc., Cary, NC. Available at: http:/ /support.sas.com/rnd/app/papers/. Accessed November 1, 2006 Schertz, C.E., Brown, G.K., 1968. Basic consid erations in mechaniz ing citrus harvest. Transaction of the ASAE., 343-346. Sevier, B.J., Lee, W.S., 2003. Adoption trends and attitudes towards precision agriculture in Florida citrus: preliminary results from a citrus producer survey. ASAE Meeting Paper No. 031100. ASAE, St. Joseph, MI. Sevier, B.J., Lee, W.S., 2004. Precision agricu lture in citrus: a probit model analysis for technology adoption. ASAE Meeting Pa per No. 041092. ASAE, St. Joseph, MI. Stajnko, D., Lakota, M., Hocevar, M., 2004. Esti mation of number and diameter of apple fruits in an orchard during the growing season by thermal imaging. Computers and Electronics in Agriculture 42, 31-42. Stajnko, D., Cmelik, Z., 2005. Modelling of a pple fruit growth by application of image analysis. Agriculturae Conspect us Scientificus 70, 59-64. Tabb, A.L., Peterson, D.L., Park, J., 2006. Se gmentation of apple fruit from video via background modeling. ASABE Meeting Pape r No. 063060. ASABE, St. Joseph, MI. Tian, L.F., Slaughter, D.C., 1998. Environm entally adaptive segmentation algorithm for outdoor image segmentation. Computers and Electronics in Agriculture 21, 153-168. Welte, H.F., 1990. Forecasting harvest fru it size during the growing season. Acta Horticulturae 276, 275-282. Williams, P., and Norris, K., 2001. Near-Infra red Technology: In Agriculture and Food Industries, 2nd ed. American Association of Cere al Chemists, Inc., St. Paul, MN. Winter, F., 1986. Modeling the biological and economic development of an apple orchard. Acta Horticulturae 160, 353-360. Whitney, J.D., Wheaton, T.A., Miller, W.M., Salyani, M., Schueller, J.K., 1998. Sitespecific yield mapping for Florida citrus. Proc. Florida State Ho rticulture Society 111, 148-150. Whitney, J.D., Q. Ling, Wheaton, T.A., Mille r, W.M., 1999. A citrus harvesting labor tracking and yield monitoring system. ASAE M eeting Paper No. 993107. ASAE, St. Joseph, MI. Whitney, J.D., Q. Ling, Wheaton, T.A., Mille r, W.M., 2001. A citrus harvesting labor tracking and yield monitoring system. Applie d Engineering in Agriculture 17, 121-125.

PAGE 86

86 BIOGRAPHICAL SKETCH The author was born in 1981 in Mansfield, Ohio. Most of his childhood was in Melbourne, FL where he graduated from Eau Gallie High School, 1999 and Brevard Community College with an Associate of Arts degree in 2001. In May 2005 he graduated with a Bachelor of Engineering de gree in electrical engineering from the University of Florida. Continuing his educat ion, Kevin graduated from the University of Florida in August 2007 with a Master of E ngineering degree in agricultural and biological engineering and anot her Master of Engineering in electrical and computer engineering. Kevin Kane currently lives in Aiken, SC with his loving wife Dr. M. Kane and his two dogs Ba iley and Killian.