|
DISCRIMINATION OF AN INVASIVE PLANT, Schinus terebinthifolius, FROM
AERIAL DIGITAL IMAGERY
By
LEONARD G. PEARLSTINE
A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL
OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT
OF THE REQUIREMENTS FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY
UNIVERSITY OF FLORIDA
2000
DISCRIMINATION OF AN INVASIVE PLANT, Schinus terebinthifolius, FROM
AERIAL DIGITAL IMAGERY
Copyright 2000
by
Leonard G. Pearlstine
ACKNOWLEDGMENTS
I owe a special debt to my parents for their caring support and the time they have
shared with my children. They were valuable partners in completing this dissertation. I
also thank my children, Timothy and Benjamin, for their patience with me during absent
evenings and weekends. We have some fun to catch up on.
The members of my committee--Scot Smith (chair), Bon Dewitt, Grenville
Barnes, Ken Portier, and Loukas Arvanitis--have been very supportive. Their comments
and edits contributed substantially to my research and to this document. I particularly
want to thank Ken Portier, who always has time and enthusiasm for the students who
come to his office, and to Scot Smith, for his constant encouragement.
Franklin Percival, my immediate supervisor and Unit Leader of the Florida
Cooperative Fish and Wildlife Research Unit, and Nat Fraiser, Chairman of the
Department of Wildlife Ecology and Conservation showed tremendous understanding in
allowing me to pursue this degree while continuing at work. Their encouragement is also
greatly appreciated.
TABLE OF CONTENTS
page
A C K N O W L E D G M E N T S ............................ ................. ..............................................iii
LIST OF TABLES ............. ....... ...... ........ ....... ......................... vii
LIST OF FIGURES .................................................. ........... ............ xi
AB STR A CT ........... .. ...................................................................................... .............. xv
CHAPTERS
1 IN TR O D U C T IO N ....................... ........................... ............ .............
Im pacts of B razilian Pepper................................................. ............................... 1
Texture Analysis ...................................................................... ........ 3
G general A approach .................................................................................... 8
2 IM A G E C O L L E C T IO N ....................................................................... ...................10
Intro du action ..................................................... 10
M eth o d s ..................................................... 10
Camera Setup ....................................... .............................10
Field Collection of Aerial Imagery and Ground Surveys ........................................ 11
R e su lts ....................................................................... 1 3
Discussion ........................ .............. ..... 18
3 SPECTRAL NEIGHBORHOOD TEXTURES .................................. ...............19
M eth o d s ................... ................................. 2 1
F irst O rd er Statistics .............. .... .. .. .. .. ........................................ ................ .. 2 2
Normalized Gray Level Dependence Matrices (NGLDM).................................... 24
U nivariate Classification Trials ........................................ .......................... 27
R esu lts .............. .. ... .. ... ......... .. .. .................................................... 3 0
Signature D evelopm ent...................... .. .... ......... ......................... .. .................. 30
Univariate Calibration Image Classifications .................................. .............. 31
Univariate Validation Image Classifications .................................................... 33
D iscu ssio n ............. ......... .. .. ......... .. .. ................................................... 3 3
iv
4 E D G E T E X T U R E S ......... ...... ........... ................. ........................... .........................64
Introduction....................... ............... ..... ............. 64
M ethods.............................. ............. ...... 65
R e su lts .............. ..... ............ ................. ....................................................... 6 7
Image Signatures.............................. ............... 67
Univariate Calibration Image Classifications ....................... .............. 70
Univariate Validation Image Classifications ...................................................... 72
D iscu ssion ................. ...................................................................................... 73
5 MULTIVARIATE DISCRIMINATION AND CLASSIFICATION ..........................117
Intro du action ..................................................... 1 17
M eth ods ....................................................................................... 117
Multi-texutral Classification of Schinus ................ ..................... 117
Multi-textural Classification for all Vegetation Classes....................................... 122
R results ........................................................................ ........ .. .......... 123
Multi-texutral Classification of Schinus .................................. 123
Multi-textural Classification for all Vegetation Classes ............... ....... ... 127
D discussion ........................................ ........129
6 C O N C L U SIO N S ................................................................ 145
Sum m ary ...................................... .................. .......... .. 145
O their M odeling A approaches ....................................................................... 148
Other Texture A approaches ..................................... ............... ............... 150
M anagem ent Considerations ........................................... ............... .............. 151
F u tu re S tu d ie s ...................................................................................... 1 5 2
APPENDIX C++ PROGRAMS DEVELOPED FOR THIS DISSERTATION ............154
Sum m ary of Program s ............ .. ..... ... .. ... .... .... .. ................... ............ ...... 154
N ote to U sers............................. ............. ...... 155
T ex tu reM ak er ......... .... ......... ............... ...................................... 162
E d g eT ex tu res .................................................................................................... 17 6
E rd2A sc............................. .............. ..... 19 1
E rd2A scStratified ................. ............................... .............................................. 195
B atchE x tractC lass ........................................................................................ 2 02
E rd 2 S p lu s ....................................................................................................... 2 0 6
E rd 2 S A S ...................................................................... ..... 2 1 4
S A S R eco d e ........................................................................................... 2 19
SASMultiRecode ......................................... ................. 223
D ensityT threshold ......... ...................................................... .. ..... .. .......... .. 23 1
Uclassify ..................................... ............. .............. 234
A c c A ssE rd a s ............................................................................................ 2 4 2
ErrorM atrix .................. ................................... ... .... .. .... ........ ...... 246
LogRegressClassify ....................................................... .. 250
v
L ogRE rrorM atrix .................. ....................................... ............. .. 256
M ultiProbClassify ..................................... ................................. .. 261
ErdHistogram ............................. .............. 266
E rdasB andV iew er .. ........... .................................................... ................ .. .. 270
ErdasEnhance.................................... .............. 275
L IST O F R E FE R E N C E S ........................................................................ ...................283
BIOGRAPHICAL SKETCH .............. ......................... .................... ...............288
LIST OF TABLES
Table Page
2-1. Frequency of occurrence of classes in the calibration images.................. ...............13
2-2. Frequency of occurrence of classes in the validation images...............................13
3-1. Decision vector for the signatures shown in Figure 3-3. .........................................28
3-2. Contingency table for wwl085 Mean textures classification .....................................51
3-3. Contingency table for ww1093 Mean textures classification .....................................51
3-4. Contingency table for ww1099 Mean textures classification. ........................................52
3-5. Contingency table for ww2042 Mean textures classification .....................................52
3-6. Contingency table for ww2046 Mean textures classification .....................................53
3-7. Contingency table for wwl085 Standard Deviation textures classification. .................55
3-8. Contingency table for ww1093 Standard Deviation textures classification. .................55
3-9. Contingency table for ww1099 Standard Deviation textures classification. .................56
3-10. Contingency table for ww2042 Standard Deviation textures classification. ................56
3-11. Contingency table for ww2046 Standard Deviation textures classification. ................57
3-12. Contingency table for 0212-528 Mean textures classification.....................................59
3-13. Contingency table for ww2008 Mean textures classification .....................................59
3-14. Contingency table for ww2060 Mean textures classification .....................................60
3-15. Contingency table for ww2091 Mean textures classification. .................... ...........60
3-16. Contingency table for 0212-528 Standard Deviation textures classification ...............62
3-17. Contingency table for ww2008 Standard Deviation textures classification. ................62
3-18. Contingency table for ww2060 Standard Deviation textures classification. ................63
3-19. Contingency table for ww2091 Standard Deviation textures classification. ................63
4-1. Contigency table for ww1085 Laplacian Edge Density texture classification ...............86
4-2. Contigency table for ww1093 Laplacian Edge Density texture classification ...............86
4-3. Contigency table for ww1099 Laplacian Edge Density texture classification ...............87
4-4. Contigency table for ww2042 Laplacian Edge Density texture classification ...............87
4-5. Contigency table for ww2046 Laplacian Edge Density texture classification ...............88
4-6. Contingency table for ww1085 Sorbel Density texture classification...........................90
4-7. Contingency table for ww1093 Sorbel Density texture classification............................90
4-8. Contingency table for ww1099 Sorbel Density texture classification............................91
4-9. Contingency table for ww2042 Sorbel Density texture classification ...........................91
4-10. Contingency table for ww2046 Sorbel Density texture classification .........................92
4-11. Contingency table for ww1085 Sorbel Edge Magnitude texture classification. ............94
4-12. Contingency table for ww1093 Sorbel Edge Magnitude texture classification. ............94
4-13. Contingency table for ww1099 Sorbel Edge Magnitude texture classification. ............95
4-14. Contingency table for ww2042 Sorbel Edge Magnitude texture classification. ............95
4-15. Contingency table for ww2046 Sorbel Edge Magnitude texture classification. ............96
4-16. Contingency table for ww1085 Sorbel Edge Slope texture classification...................98
4-17. Contingency table for ww1093 Sorbel Edge Slope texture classification...................98
4-18. Contingency table for ww1099 Sorbel Edge Slope texture classification...................99
4-19. Contingency table for ww2042 Sorbel Edge Slope texture classification...................99
4-20. Contingency table for ww2046 Sorbel Edge Slope texture classification.....................100
4-21. Contingency table for ww1085 Sorbel Edge Standard Deviation texture
classification. ......................................................................102
4-22. Contingency table for ww1093 Sorbel Edge Standard Deviation texture
classification .................................................................... 102
4-23. Contingency table for wwl099 Sorbel Edge Standard Deviation texture
classification .................................................................... 103
4-24. Contingency table for ww2042 Sorbel Edge Standard Deviation texture
classification. ......................................................................103
4-25. Contingency table for ww2046 Sorbel Edge Standard Deviation texture
classification. ......................................................................104
4-26. Contingency table for 0212-528 Sorbel Edge Density texture classification...............106
4-27. Contingency table for ww2008 Sorbel Edge Density texture classification ................106
4-28. Contingency table for ww2060 Sorbel Edge Density texture classification ................107
4-29. Contingency table for ww2091 Sorbel Edge Density texture classification ................107
4-30. Contingency table for 0212-528 Sorbel Edge Magnitude texture classification. ...........109
4-31. Contingency table for ww2008 Sorbel Edge Magnitude texture classification. ............109
4-32. Contingency table for ww2060 Sorbel Edge Magnitude texture classification. ..........110
4-33. Contingency table for ww2091 Sorbel Edge Magnitude texture classification. ..........110
4-34. Contingency table for 0212-528 Sorbel Edge Slope texture classification .................12
4-35. Contingency table for ww2008 Sorbel Edge Slope texture classification.....................112
4-36. Contingency table for ww2060 Sorbel Edge Slope texture classification.....................113
4-37. Contingency table for ww2091 Sorbel Edge Slope texture classification.....................113
4-38. Contingency table for 0212-528 Sorbel Edge Standard Deviation texture
classification. ................................................................... ..........115
4-39. Contingency table for ww2008 Sorbel Edge Standard Deviation texture
classification. ................................................................... ..........115
4-40. Contingency table for ww2060 Sorbel Edge Standard Deviation texture
classification. ................................................................... ..........116
4-41. Contingency table for ww2091 Sorbel Edge Standard Deviation texture
classification. ................................................................ ... ..........116
5-1. Logistic regression coefficients for Schinus terebinthifolius derived from the
calibration im ages. ........................ ....... ...... ...... ...............124
5-2. Accuracy assessment for images with a classification threshold at 50% probability......124
5-3. Accuracy assessment for images with a classification threshold at 30% probability......125
5-4. Commission errors for Schinus terebinthifolius in images that do not contain
Schinus...................................................... ................... ... .... ... .... . 126
5-5. Logistic regression variables and coefficients for the additional vegetation cover
ty p e s. .......................................................................... . 13 6
5-6. Contingency table for ww1085 logistic regression classification.............................139
5-7. Contingency table for ww1093 logistic regression classification.............................139
5-8. Contingency table for ww1099 logistic regression classification.............................140
5-9. Contingency table for ww2042 logistic regression classification ................................140
5-10. Contingency table for ww2046 logistic regression classification .............................141
5-11. Contingency table for 0212-528 logistic regression classification..............................141
5-12. Contingency table for ww2008 logistic regression classification .............................142
5-13. Contingency table for ww2060 logistic regression classification .............................142
5-14. Contingency table for ww2091 logistic regression classification ............................143
LIST OF FIGURES
Figure Page
1-1. Comparison of spectral response in three similar wavelength bands for imagery from
Landsat Thematic Mapper satellite and a high resolution aerial digital camera...4
2-1. Images used in testing and calibration of textures and their corresponding
classification masks.................... .............. ..............14
2-2. Images used for validation of texture classifications and their corresponding
reference classifications. ........................................ ....................................... 16
3-2. Creating the Normalized Gray Level Dependence Matrix............................................25
3-3. Combined signatures for the M ean Texture. .......................................... ............... 28
3-4. Illustration of accuracy assessment measures ................................................29
3-5. Spectral signatures from the infrared image band................ .............. ............... 36
3-6. Spectral signatures from the red image band ....................................... ............... 37
3-7. Spectral signatures from the green image band.............. .............. ................38
3-8. ww2042 spectral signatures of the first-order textures with a window size = 7x7. ........39
3-9. ww2042 spectral signatures of the first-order textures with a window size = 15x15.....40
3-10. ww2042 Spectral signatures of the first-order textures with a window size = 3 1x31...41
3-11. ww2042 Spectral signatures of the first-order textures with a window size = 63x63...42
3-12. M ean texture signatures ........................................................................ ...................43
3-13. Standard Deviation texture signatures. ............................................... ............... 44
3-14. Skew texture signatures ........................................................................ ...................45
3-15. Entropy texture signatures. ................................................ ................................ 46
3-16. H om ogeneity texture signatures .................................. ............... ............... 47
3-17. Simpson's Diversity texture signatures. ..................................... .....................48
3-18. Normalized Gray Level Difference Matrix contour plots for vegetation covers in
im age w w 2042. ....................... ...................... ........................ 49
3-19. Classification of calibration images from the Mean texture. .......................................50
3-20. Classification of calibration images from the Standard Deviation texture...................54
3-21. Classification of validation images from the Mean texture.........................................58
3-22. Classification of validation images from the Standard Deviation texture....................61
4-1. V alues used in 3x3 edge filters................................................. .................................65
4-2. Exam ple Sorbel edge im ages ................................................ .............................. 76
4-3. Exam ple Laplacian edge im ages. ............................................ ............................ 77
4-4. Sorbel and Laplacian image histograms from edge images ofwwl085........................78
4-5. Laplacian Edge Density texture signatures. ........................................ ............... 79
4-6. Sorbel Edge Density texture signatures ........................................ ........ ............... 80
4-7. Sorbel Edge Magnitude texture signatures....................... ... .......................... 81
4-8. Sorbel Edge Slope texture signatures. ........................................ ......................... 82
4-9. Sorbel Edge Standard Deviation texture signatures......................................................83
4-10. Image Standard Deviation texture signatures ..............................................84
4-11. Classification of calibration images from the Laplacian Edge Density texture .............85
4-12. Classification of calibration images from the Sorbel Edge Density texture...................89
4-13. Classification of calibration images from the Sorbel Edge Magnitude texture..............93
4-14. Classification of calibration images from the Sorbel Edge Slope texture ...................97
4-15. Classification of calibration images from the Sorbel Edge Standard Deviation
texture ............................................................... .... ..... ......... 101
4-16. Classification of validation images from the Sorbel Edge Density texture..................105
4-17. Classification of validation images from the Sorbel Edge Magnitude texture .............108
4-18. Classification of validation images from the Sorbel Edge Slope texture.....................111
4-19. Classification of validation images from the Sorbel Edge Standard Deviation
texture ............................................................... ..... ..... ......... 114
5-1. Logistic regression classifications for Schinus terebinthifolius in the calibration
im ag e s...................................................................... . 1 3 2
5-2. Logistic regression classifications for Schinus terebinthifolius in the validation
im ag e s...................................................................... . 1 3 4
5-3. Classification of calibration images from logistic regression of each vegetation
ty p e .............................................................................. 1 3 7
5-4. Classification of validation images from logistic regression of each vegetation type. ....138
5-5. Logistic regression probabilities for hardwood in image ww1085..............................143
5-6. Examples of logistic regression classification of each vegetation type when
hardw oods w ere excluded .............................................................................. . ...144
A-1. Screen shot of TextureM aker program .............. .................................... ............... 162
A-2. Screen shot of EdgeTextures program ...................................................... ...............176
A-3. Screen shot of Erd2Asc program ........................................... ............................ 191
A-4. Screen shot of Erd2AscStratified program. ....................................... ............... 195
A-6. Screen shot of BatchExtractClass program .......................................... ............... 202
A-7. Screen shot of Erd2Splus program ...........................................................................206
A-8. Screen shot of Erd2SA S program ............................................................................ 214
A-9. Screen shot of SA SRecode program ........................................ .......................... 219
A-10. Screen shot of SASMultiRecode program. ...................................... ............... 223
A-11. Screen shot of DensityThreshold program. ........................... .................. ...............231
A-12. Screen shot of Uclassify program ...........................................................................234
A-13. Screen shot of AccAssErdas program ............ ..................................... .................242
A-14. Screen shot of ErrorM atrix program ............................................................. .......... 246
A-15. Screen shot of LogRegressClassify program. .................................... .................250
A-16. Screen shot of LogRErrorMatrix program. ....................................... ............... 256
A-17. Screen shot of MultiProbClassify program................. ...........................................261
A-18. Screen shot of ErdHistogram program ................................... ........................ ......... 266
A-19. Screen shot of ErdasBandViewer program....................................... ............... 270
A-20. Screen shot of ErdasEnhance program. ........................... ..................... 275
Abstract of Dissertation Presented to the Graduate School
of the University of Florida in Partial Fulfillment of the
Requirements for the Degree of Doctor of Philosophy
DISCRIMINATION OF AN INVASIVE PLANT, Schinus terebinthifolius, FROM
AERIAL DIGITAL IMAGERY
By
Leonard G. Pearlstine
December 2000
Chairman: Dr. Scot E. Smith
Major Department: Department of Civil and Coastal Engineering
Schinus terebinthifolius, known as Brazilian pepper, is an exotic, invasive plant
species in Florida that displaces native plant species to the detriment of beneficial native
plant and animal habitat in aquatic, wetland, and terrestrial environments. Rapidly and
accurately identifying invasive species is critical to successful monitoring and
management of ecosystem change. Aerial surveys typically are used for this purpose, but
observers or photointerpretation may be inconsistent and costly. Image processing for the
classification of land cover types that relies on by-pixel spectral response of the different
vegetation in an image to discriminate the cover types is problematic. In high-resolution
imagery, adjacent pixels of the same cover type are likely to represent different parts of
the plant with very different spectral responses. Measuring spatial patterns of texture in
neighborhoods of pixels can be an effective alternative.
Textures derived from first-order statistics, co-occurrence matrices, and
components of edge in high-resolution digital color infrared images were tested for their
application to the discrimination of Schinus terebinthifolius. Logistic regression on the
texture variables found the combination of the mean, Sobel edge density, Sobel edge
magnitude, and edge standard deviation to be the best discriminators of Schinus in these
images with a median user's accuracy of 87.3 percent and producer's accuracy of 44.6
percent. The high user's accuracy and lower producer's accuracy reflect the model's
consistent identification of core areas of Schinus, but not the edges of Schinus patches.
Misidentification of other cover types as Schinus was extremely low except where Sabal
palmetto was present in the images. Contextual discrimination between Schinus and
Sabal requires more exploration. Logistic regressions also were used for combined
classifications of all cover types in an image. User accuracies for Schinus generally
remained high, but omission errors increased in cases where Schinus was misidentified as
one of the other cover types.
CHAPTER 1
INTRODUCTION
The objective of this dissertation was to develop multivariate image textural
feature classification approaches for rapid, unsupervised identification of Brazilian
Pepper, Schinus terebinthifolius, from aerial low altitude, color infrared imagery. Rapidly
and accurately identifying land covers is critical to successful monitoring of ecosystem
change. The identification of invasive and nonnative species is particularly important.
Invasive and nonnative species may severely disrupt native ecosystem functions and
displace wildlife habitats. Thirty percent of Florida's flora, or approximately 1000
species, is exotic. Sixty five of those species are listed as Category I by the Florida Exotic
Pest Plant Council (1999). These Category I species are known to disrupt native plant
communities in Florida. They displace native plant species to the detriment of beneficial
native plant and animal habitat in aquatic, wetland, and terrestrial environments. Another
59 plants are Category II species: potentially disruptive, but ecological damage is not yet
demonstrated. Successful monitoring, management, and eradication of invasive plants is
critical to maintaining native habitats in Florida.
Impacts of Brazilian Pepper
Brazilian pepper, a Category I plant species, was introduced at the turn of the
century (Morton 1978) or possibly earlier (Ferriter 1997) as an ornamental shrub. It
earned the name 'Florida holly' because of its bright red fruits and shiny leaves. In its
native South and Central American range, Brazilian pepper occurs as scattered
individuals in habitats from sea level to over 700 m elevation (Ewel et al. 1982). In south
and central Florida, Brazilian pepper can dominate the landscape as monotypic stands in
a broad range of moist to mesic habitats (Ewel 1986). The surveys of the South Florida
Water Management District indicate that Brazilian pepper occupies more than 280,000 ha
(700,000 ac) in Florida and it is the most widespread exotic in the state (Ferriter 1997).
Characteristics that make Brazilian pepper a successful opportunistic species
include fast growth, prolific seed production, near continuous shoot extension and leaf
renewal, vigorous resprouting, tolerance of a wide range of growing conditions, seed
distribution by forging birds, and shade tolerant seedlings (Ewel 1979). The latter
characteristic is particularly unusual in weedy species. These characteristics result in an
aggressive plant species that has invaded many of Florida's native plant communities and
disturbed sites such as highway and canal right-of-ways, fallow fields, and drained
cypress stands (Woodall 1982). Wildlife and plant diversity is threatened when Brazilian
pepper colonizes large areas of the state, displaces native species, and eliminates many
indigenous sources of food for wildlife (Morton 1978). For example, Curnutt (1989) has
documented declines in avian diversity and population density. Direct contact with the
sap can cause severe and persistent skin irritation. Brazilian pepper is a relative of poison
ivy. It is a hazard for rangelands where consumption of foliage by horses and cattle can
cause hemorrhages, intestinal compaction, and fatal colic (Morton 1978).
An approach to surveying and monitoring exotic species such as Schinus
terebinthifolius that 1) uses unmodified small fixed wing aircraft, 2) is ready for
evaluation immediately after the flights, 3) provides archival documentation of the exotic
species' existence and location, and 4) provides automated computer discrimination of
the target species with statistically determined confidence levels would benefit natural
resource managers and researchers attempting to document and control exotic plant
expansion. This study focused on the last item: automated computer discrimination of a
target species, Schinus terebinthifolius, from low altitude aerial digital imagery.
Texture Analysis
Image spectral tone (grey level) differences among vegetation and land covers are
often used to create signatures of the cover types in imagery of the earth's surface. These
spectral signatures may be compared to identify the likely cover type of each pixel in the
image. This pixel-by-pixel classification approach becomes more problematic as image
resolution increases because more individual features are discriminated and the variance
in spectral values observed within a cover type increases. For example, at spatial
resolution below 1 meter, imagery of a forest or shrub type would return many pixels of
leaf surfaces from different angles and different degrees of shading, stems, and possibly
various ground covers under the trees or shrubs. Individual pixels representing one of
these features may vary considerably in spectral reflectance from a pixel representing one
of the other features of the same cover type. This makes it impossible to classify a pixel
to a cover type based on its spectral tone. Spectral tone from the various features of a
cover type is averaged as images decrease in spatial resolution. Lower resolution images,
such as Landsat and SPOT satellite imagery, are often successfully classified from pixel-
by-pixel identification of a spectral signature (Lillesand and Kiefer 2000, Jensen 1986)
because the reflectance within a cover type averaged over a multi-meter square pixel is
generally more homogeneous across the pixels in the imagery. Figure 1-1 illustrates this
4
problem with sample images of Schinus terebinthifolius. A Landsat image of Schinus
shows low variance and sharp peaks in the spectral response of Schinus, making it a good
candidate for spectral tone classification. A high-resolution image of Schinus, however,
has many pure pixels of light glint off the tops of leaves and other pixels of deep shadow
below the leaves. Its tonal response extends over the entire spectral range of the image.
Figure 1-1. Comparison of spectral response in three similar wavelength bands for
imagery from Landsat Thematic Mapper satellite and a high resolution aerial digital
camera. The Landsat image has a spatial resolution of 28.5 m vs 5 cm for the digital
camera image. The bands are blue line: blue wavelengths (TM 2), magenta line: red
wavelengths (TM 3), and yellow line: near infrared wavelengths (TM 4).
Photo-interpreters rely on many clues for classification in addition to spectral
tone. Texture, shape, and context of features in the image are elements that go beyond the
individual pixel to the neighborhood of features surrounding the pixel. When a pixel is
viewed in context with its neighboring pixels, a pattern may emerge that is consistent
with a cover type and different from other cover types. This sort of contextual
Landsat TM Schinus Signature
1.20 -
5 0 ..0 . ......... ........... ..- ........ .... ........ ........ ..... ........
01.0 .................... .. ..................................... .:'.I.Z.I.I-I.I. .I-I. .I. .Z.I. .: .*.:..-...-......-.:.
S . . ........................................
m 0 .20 -
. ............. ., .. ...O .. .. .
Refledance Value
ii
Li- SD
6 :oc n""""""""""
"' Al____ __ """""""""
a020" " " "" " "" "
om" "" " "" "
ci :,tc Co ct CO 81
Reflectance Viue ""
Digital Camera Schinus Signature
1.320
S1.00 -
g- 0.80 --
a I
--: 0.60 ,--'*.--..- ,-- .. .....-- -- -
0.03
e Om f 'C D -nCO C, o-M
0 i M CO c W t N C C
Reflectance Value
identification can improve the accuracy of satellite image classification (Kushwaha et al.
1994, Irons and Petersen 1981) and is essential for classification of high-resolution
images.
Haralick et al. (1973) defined texture as the statistical properties of the spatial
distribution of gray tones (varying shades of gray in the resolution cells) in an image.
Texture is composed of patterns of these gray tones that are smaller than the area where
they are repeated (Sun and Wee 1983). Texture analyses to improve classification of
satellite imagery and aerial photography have been applied to forested landscapes
(Holopainen and Wang 1998, Wulder et al. 1998, Riou and Seyler 1997, Kushwaha et al.
1994), wetlands (Yamagata and Yasuoka 1993), landuse (Emerson et al. 1999, Irons and
Petersen 1981), and geomorphology (Shih and Schowengerdt 1983). Texture analyses
have also been applied to sets of close-up "natural" images such as the Brodatz texture
images (Brodatz 1967) of sand, pebbles, cow hide, fieldstones, and other regular textures
(Potlapalli and Luo 1998, He and Wang 1991), and even images of patterns on animal
and plant surfaces (Kiltie et al. 1995).
No single texture or contextual approach has emerged in the last 3 decades of
computerized texture analysis as superior in its capability to improve classification
accuracy. Two of the most common approaches are first- and second-order statistical
textures. First-order statistics such as the mean, standard deviation, and skew quantify the
distribution properties of spectral tone in the image. Second-order statistics describe the
frequency with which one gray tone appears in specific relationship to another gray-tone
on the image (Haralick et al. 1973) through the construction of relative frequency
distribution matrices. Second-order statistical textures also are referred to as spatial co-
occurrence textures. Yamagata and Yasuoka (1993) obtained enhanced classification
performance in SAR microwave imagery of wetlands using the co-occurance matrix
textures. Microwave images characteristically contain speckle noises that confuse
conventional classification methods. Kushwasha et al. (1994) reported modest
improvements in IRS satellite classification of forests using Haralick textures, but Dikshit
(1996) obtained better results with first-order statistics (airborne Thematic Mapper
images). Both authors observed better accuracies with a combination of tone and texture
layers rather than the spectral tone layers alone. Irons and Petersen (1981) as well as Hsu
(1978) had earlier results suggesting that higher order statistics do not improve the
texture classification for Landsat MSS or aerial photography (high and low altitude),
respectively. Bijlsma (1993) examined first- and second-order texture measures with
respect to herbaceous vegetation structure (i.e., species richness and structural diversity)
rather than cover type. For this application, Bijlsma found textures derived from second-
order statistics had no meaning, and, in fact, the simplest model, containing the mean
texture only, yielded the best results.
Some other texture analysis approaches include experimentation with Fourier
transformations (Riou and Seyler 1997, Coggins and Jain 1985, Matsuyama et al. 1980)
and their more recent superset, wavelets (Kiltie et al. 1995, Unser 1995). He and Wang
(1991) observed that traditional Fourier transformation is better in characterizing absolute
intensity relationships and spatial relationships of pixels in a large scale, whereas texture
analysis is interested in relative intensity relationships among pixels in a small
neighborhood. They advocated a textural spectrum method for analysis.
Semivariograms measure spatial variability and spatial continuity in an image.
Spatial continuity (i.e., data that are close together are more likely to have similar values
than data that are far apart) exists in most earth science data sets (Isaaks and Srivastava
1989). Carr and De Miranda (1998) reported that textures based on the semivariogram
result in better classification accuracies for microwave imagery but are less effective than
the second-order Haralick textures for optical images. Wulder et al. (1998), however,
reported significant increases in accurate classification of aerial imagery when the
semivariogram was combined with second-order statistical textures vs. the use of either
texture alone.
Fractals are another relatively new approach to texture analysis that can provide a
measure of surface roughness. Potlapalli and Luo (1998) and Dennis and Dessipris
(1989) have demonstrated its applicability with Brodatz textures. Dennis and Dessipris
also presented a computationally efficient approach to generating the fractal information
layer. Emerson et al. (1999) demonstrated fractal textures computed from the NDVI of
high-resolution aerial imagery. When images were degraded from 10 m to 80 m spatial
resolutions, agricultural land cover increased in fractal dimension, forested land cover
grew smoother, and urban areas remained roughly self-similar.
As Ryherd and Woodcock (1996) observed, texture is only a useful addition when
classes in the images exhibit differences in spatial tone patterns. The authors also
observed that for the range of remote sensed images they tested, results never were
degraded when textural information layers where combined with the spectral bands.
The literature is replete with comparisons of single new texture approaches to
other, more established approaches such as the Haralick co-occurrence matrix textures.
When textures are used in combination with other bands of information, it is most
commonly with the original spectral bands. Rarely have multiple textures been tested for
the best combination to discriminate land covers. A premise of this study was that
multivariate discriminate function analysis yields results superior to those obtained using
only a single texture. Several features are believed to be desirable for textures selected in
this study. Textures that are invariant to rotation are necessary because vegetation does
not always grow in the same orientation and planes do not always fly in one direction.
Textures that are invariant to gray level transformations also are desirable because
lighting conditions and camera parameters can vary between images. Finally, simple
algorithms for computational efficiency and ease of transfer to other applications are
most suited to the objective of this study.
General Approach
Low altitude, digital aerial photography was collected over areas containing
Schinus terebinthifolius and other vegetation likely to occur in the vicinity of Schinus.
First-order, second-order, and edge-texture analyses were conducted with these images,
and texture signatures generated for each of the vegetation types were observed. Texture
values are computed for each pixel in an image by examining the neighborhood
surrounding that pixel, i.e., a moving window. Optimal window size is one of the issues
that must be resolved for a successful class identification. A small neighborhood is
computationally faster and may capture higher frequency changes. A larger neighborhood
may better represent the pattern of texture in a class. However, the larger neighborhood
typically smoothes the resulting image, thus reducing the ability of detecting accurate
class boundaries. Single- and multiple-texture classifications were conducted to test
classification accuracies of each texture and the best combination of textures to maximize
discrimination of Schinus from the other image cover types. Non-parametric methods
were used for classification. Single-texture classifications proceeded by comparison of
the density functions of texture signatures for the cover types and selection of the highest
density class for the texture value at each location. Multivariate class discrimination was
accomplished with a "best" combination of texture variables by best subset selection with
logistic regression and backward elimination stepwise logistic regression. Classification
success was evaluated with the now-standard method of reporting contingency tables and
their associated producer and user accuracies as presented in Story and Congalton (1986)
and Congalton and Green (1999).
CHAPTER 2
IMAGE COLLECTION
Introduction
The Florida Cooperative Fish and Wildlife Research Unit acquired high-
resolution digital aerial imagery to assist land cover mapping. Because individual plant
species often can be visually identified on the imagery, there appeared to be a potential to
use the imagery to establish the identification of invasive plant species. Images from
these flights were selected, and the cover types in them were mapped to test the potential
for computer recognition of the cover types.
Methods
Camera Setup
A Kodak DCS420 CIR 24-bit digital color infrared digital camera with a 1012 x
1524 fixed plane array was mounted on a Cessna 172 aircraft. An Accuphoto aerial
photography control system took input from an on-board Garmin 12-channel GPS and a
Watson Industries Attitude and Heading Reference System (AHRS-BA303). This device
utilizes three solid state attitude gyros to provide two attitude angles and a third heading
angle. An accuracy of 0.3 degrees in attitude angles determination can be achieved using
this instrument (Watson Industries, 1995). The GPS received real-time differential
correction signals from a DCI FM receiver. The control system used input from the GPS
to trigger the digital camera and at the same time record the GPS and attitude
information. This system was used to calculate the geo-location of each frame of the
aerial imagery. Geolocations were not required for the image processing objective of this
study. However, geolocations would be critical for any monitoring applications using the
results of this research.
Field Collection of Aerial Imagery and Ground Surveys
Aerial images from this setup were collected for testing and calibration of texture
signatures. The images were selected to provide a variety of vegetation covers that are
likely to occur adjacent to Schinus terebinthifolius. They were also selected to provide
images of Schinus as large clusters as well as individual plants. Five images were
selected from a series of flights conducted in October 1997 over Cape Canaveral National
Seashore, on the east coast of Florida (- 80 37'W, 280 40'N). The vegetation
composition of the images is as follows:
1. ww2042 Brazilian pepper, cattail, wax mrytle, and open water
2. ww2046 hardwoods, sabal palm
3. ww1093 Brazilian pepper, black mangrove, herbaceous, open water
4. ww1099 Brazilian pepper, shrub, herbaceous
5. ww1085 sabal palm, herbaceous
where Brazilian pepper is Schinus terebinthifolius, cattail is Typha spp., wax mrytle is
Myrica cerifera, black mangrove is Avicennia germinans and sabal palm is Sabal
palmetto.
All of the images were flown at approximately 366 m above terrain with a 135
mm lens yielding a 37 x 25 m ground footprint and a spatial resolution of 2.4 cm. All
images were cropped to an area of interest to reduce processing times and storage
requirements.
Visits to Cape Canaveral included the sites of the ww2042 and ww1093 images.
These site were visited to confirm the vegetation types interpreted to occur on the images.
Accuracy assessments were conducted using three additional images from the
same date over Cape Canaveral and one image from February 2000 containing Schinus
over an area south of Tampa, near Ruskin on the west coast of Florida (- 82o 24'W, 27o
40'N), 60 miles south of the Cape Canaveral site. The southern flightlines were
overexposed and lost much of the image detail. One image (image 0212-528) was
retained to test the limits of the sensitivity of texture analyses. These images are referred
to as "validation images" in the remaining chapters. The vegetation composition of the
validation images is as follows:
1. 0212-528 Brazilian pepper and Brazilian pepper stressed by herbicide
2. ww2008 Brazilian pepper and herbaceous ground cover
3. ww2060 Brazilian pepper, marsh, and water
4. ww2091 sabal palm, hardwoods, and water
To compare the texture results to specific vegetation covers, a raster mask file was
created that classifies the original images into their known vegetation types for both the
calibration images and the validation images. The masks were created by supervised
classification in Imagine image processing software (ERDAS, Inc.) followed by
extensive on-screen digitizing to correct and clean up the final classification. These
classifications are referred to in the remaining chapters as "reference classifications" or
"reference masks."
Results
The images selected for testing and calibration of textures are shown in Figure 2-
1. To the right of each image is the reference classification. Table 2-1 presents the
frequency of occurrence of each of the cover types in the calibration images. The
validation images and their corresponding reference classifications are shown in Figure
2-2. Table 2-2 includes the frequency of occurrence of each cover type in the validation
images.
Table 2-1. Frequency of occurrence of classes in the calibration images.
CLASS
ww1085 ww1093 ww1099 ww2042 ww2046
Table 2-2. Frequency of occurrence of classes in the validation images.
ICLASS [2012-528 ww2008 ww2060 ww2091
Unclassified
Schinus terebinthifolius
Sabal palmetto
Open Water
Avicennia germinans
Herbaceous
Shrub
Typha spp./ Marsh
Myrica cerifera
Hardwoods
22.9% 2.4%
77.1% 48.2%
15.1%
31.9%
25.9%
49.4%
18.9%
27.1%
Unclassified 7.2% 27.1% 3.9% 5.2% 49.4%
Schinus terebinthifolius 32.0% 30.0% 34.1%
Sabal palmetto 3.7% 0.9% 48.3%
Open Water 15.0% 20.5%
Avicennia germinans 8.7%
Herbaceous 17.2% 51.4% 2.3%
Shrub 14.7%
Typha spp. 23.3%
Myrica cerifera 16.0%
Hardwoods 89.1%
29.7%
8.8%
42.6%
WWZU4Z
WW1UY.i
Figure 2-1. Images used in testing and calibration of textures and their corresponding
classification masks. The reference classifications show a 31 pixel buffer around each
class. Masks with 63 pixel buffers were also created for each image.
;rfLWW9Q
a -~
r
Legend
Schinus terebinthifolius
Myrica cerfera
Typha spp.
Sabal palmetto
Water
D Herbaceous
SScrub
D Avicennia germinans
Mixed Hardwoods
SUnclassified
Figure 2-1 continued.
wwl Q099
wwZU46
U
U
m
U
U
r.,:C ~XI
d.
9;
WWZUUS
ww2060
Figure 2-2. Images used for validation of texture classifications and their corresponding
reference classifications.
Legend
0 Schinus terebinthifolius Herbaceous
D Stressed Schinus t. Scrub
Typha spp. Mixed Hardwoods
Sabal palmetto Unclassified
SWater
Figure 2-2 continued.
Discussion
The calibration and validation images depict a diverse set of vegetation cover
types from forested to herbaceous that is likely to occur in images with Schinus
terebinthifolius. The images also portray Schinus in densities ranging from individual
bushes to large clumps. These images provide a good range of conditions for testing
texture analysis for classification.
CHAPTER 3
SPECTRAL NEIGHBORHOOD TEXTURES
Introduction
First-order statistical textures and second-order or co-occurrence statistical
textures are explored in this chapter with regard to their potential to discriminate Schinus
terebinthifolius and other cover types. First-order statistics are simple continuous
measures of variability such as the mean, standard deviation, and skew.
Second-order statistical textures are used to measure variability of values in a
neighborhood in relationship to the other values in a neighborhood. This approach
requires that a co-occurrence matrix be generated from which the textures are derived.
The most cited example of second-order, co-occurrence textures are the Haralick
textures. Haralick et al. (1973) computed "angular nearest-neighbor gray-tone spatial-
dependence matrices" as the relative frequency, P,,, with which two neighboring cells,
one with gray tone, i, and the other with gray tone, j, occurred on the image. Separate
matrices were computed for the horizontal, vertical and two diagonal directions. Texture
features extracted from the matrices measured image homogeneity, contrast, coarseness,
and other spatial properties.
The Haralick textures require that images being compared are normalized to one
another so that gray tones of like features in the different images produce the same tones.
Haralick et al. (1973) suggested an involved equal-probability quantizing procedure
before texture analysis. Another drawback of many of the Haralick textures measures is
their angular dependence. Because textures must be angularly independent before they
can be used for classification, Haralick et al. suggested using the average and range of the
textures taken in the four directions. In 1983, Sun and Wee presented an alternative
neighboring gray level dependence matrix (NGLDM) for generating the co-occurrence
matrices and texture features that considers the relationship between a cell and all of its
neighboring cells at one time instead of one direction at a time. In addition to being
angularly independent, Sun and Wee stated that their approach is invariant under linear
gray level transformations, eliminating the need for image normalizations. The Sun and
Wee procedures are more computationally efficient because the neighborhood
relationships are computed once rather than 4 times and because the generated individual
matrix requires fewer operations than that of a single directional Haralick matrix.
Texture distributions within a particular cover type are referred to as "signatures"
for that cover type throughout this dissertation. The texture distributions or signatures
generated in this chapter and in Chapter 4 are transformed to kernel density estimates for
comparison and classification applications. The kernel estimator approximates the
distribution of the data by the summation of smooth, symmetrical functions placed at
each observation (Silverman 1986). Kernel density estimation is used rather than the
original data distribution because it produces a distribution with an area equal to 1 (i.e.,
integral of K(x)dx = 1) and is, therefore, a probability function necessary for modeling.
The kernel function also smoothes through the usual spikiness of the original data
distribution making a model more suitable for generalization.
Methods
First-order and NGLDM textures were computed for each of the three bands in
each of the five calibration images. The texture values are computed in a rectangular
moving window or neighborhood. The center cell in a neighborhood is assigned a value
that is the computed texture for that neighborhood. The window then is moved over one
pixel, and a new value is computed for that neighborhood. The window continues to
move across the image assigning texture values to the center cell of the window until the
entire image has been scanned. The texture programs described below assign values to all
the cells in the image by "mirroring" the edges. Because analysis windows centered on
cells near the edge of an image extend beyond the image, values are assigned to the
empty cells in the window by assuming that values just outside the image are the same as
values just inside the image and replicating those values to the empty cells (Figure 3-1).
The size of the moving window was varied from 7 pixels to 63 pixels in the ww2042
image to find the smallest analysis window in which texture discrimination remained
effective (if the texture is ever effective). The selected window size was applied to
analysis of the other images.
Min 25 7 |Figure 3-1. Mirroring image values
Moving Window .
Io in a moving window. This example
S -shows a 5 x 5 pixel moving window
-centered at the upper left corer of
I I the image. Textures computed from
I the 25 values in this window will be
5 3 3 2 recorded to the center pixel (grayed
cell).
Unless otherwise noted, the programs used in this chapter were written by the
author in C++ using the Borland Builder programming environment. The source code is
listed in the Appendix.
Output from each of the procedures below is a multiple band image file (Erdas 7.x
lan format) with each band being one of the texture images. The texture results were
compared to specific vegetation covers using the reference masks that classify the
original images into their known vegetation types (Chapter 2). The program Erd2Splus
produces an ASCII text file of randomly sampled texture values in a format for import to
the S-Plus statistical package (MathSoft, Inc.) for density estimation and plotting.
Erd2Splus uses the mask file to stratify the output by vegetation type. The ERDAS
Imagine Search routine and ModelMaker tools (ERDAS, Inc.) were used to create
versions of the mask that differentiate between vegetation near edges to other vegetation
types and vegetation away from edges (32 or 64 pixels inside the edge between classes).
First Order Statistics
The author-created program TextureMaker was used to generate first order
statistical textures for the mean, standard deviation, skew, entropy, and Simpson's
diversity. The algorithms are presented here.
From Tomaita and Tsuji (1990):
n
Mean = ip,
1=1
Standard Deviation = I(i p)2p,
n
Third Moment (Skewness) = (i P)3p,
1=1
n
Entropy= p, log p,
1=1
where i = the observed spectral intensity
n = the number of observations
p, = the probability of i occurring,
i.e., p, = number of points at intensity i / total number of points
S= mean of the intensities
The third moment measures the skew of the histogram of the intensities. When the
third moment is zero, the histogram is symmetric. The value is negative when the
histogram is skewed to the left and positive when the histogram skews to the right
(Tomaita and Tsuji, 1990). Entropy is a measure of the distribution of the intensities. Its
value approaches zero as the distribution becomes more clumped and its values increase
as the distribution becomes more uniform.
The Entropy metric is also known as Shannon's Diversity Index (McGarigal and
Marks, 1995). Shannon's Index uses the natural log and returns a value based on
information theory that represents the amount of "information" in the image (Shannon
and Weaver, 1949). The absolute magnitude of value is less useful than the relative
magnitude used for comparison (McGarigal and Marks, 1995).
Another diversity index is Simpson's Diversity Index:
n
5. Simpson's = 1.0- p,2
1=1
Simpson's Diversity Index represents the probability that intensities selected at
random would be different. The probabilities can be interpreted in both absolute and
relative terms. Simpson's Diversity Index is more intuitive than Shannon's and less
sensitive to rare values (McGarigal and Marks, 1995).
Normalized Gray Level Dependence Matrices (NGLDM)
There are two steps to creating an NGLDM for computing textures (Sun and Wee
1983). The first step is to find the NGLDM numbers, s, themselves. Within a moving
window neighborhood, s is equal to the number of values that are the same as the center
value. A large s suggests a smooth texture and a small s suggests a rough texture around
the center value. The output of this step is referred to as the table P. Next, the NGLDM,
Q, is constructed where Q(k,s) is the total number of entries in P that have gray level k
and NGLDM number s. The Q matrix is the frequency histogram of grayness variation in
the image. If a texture is coarse, it is likely that a pixel and its neighboring pixel have
similar gray level values so the occurring frequencies of large NGLDM numbers will be
higher than the occurring frequency of small NGLDM numbers. For a fine texture, the
NGLDM number is likely to be small because a pixel is likely to have a different gray
level value from its neighboring pixels. Therefore, the more large numbers are
concentrated on the right-hand side of the Q matrix, the coarser the image. The higher
the frequency counts on the left-hand side of the Q matrix, the finer the image texture.
Figure 3-2 illustrates the creation of an NGLDM.
In the TextureMaker program, the P table is generated for the entire image by
recording the NGLDM number, s, in a small window about each pixel. The Q matrices
are then generated by creating a frequency count of NGLDM numbers in a new, larger
4 3 6 5 4 1 (4 P,1) R .2) ,1) 4.t (1)
3 5 6 3 3 3 (j) (55. (6.3 P, p, P.2)
5 5 6 1 3 1 (2) 5,2) (6,) (1.1) .3) (1,2
3 3 8 1 1 t1 P,1) E. .- 3 3) (185] 81,3
2 0 8 1 8 1 1o0I.f) Qi, M .Q.1) .12)
a. b. C.
NOLDO NuImbs, s
0123 6B7 B
0100000000
0 1 0 0 0 42 0 D D
210 000 D 0 D 0 O
3043200000
502300000
6011300000
70000 D 000 0 D
.0 1. 2 1 0 0 0 00
Figure 3-2. Creating the Normalized Gray Level Dependence Matrix. (a.) Gray values in
original image. (b.) P table generated from a. showing (k,s) values (note that no edge
mirroring was used) in a 3x3 pixel window about each pixel. (c.) Q matrix from b.
showing frequency count of P(k,s) values. Adapted from Sun and Wee (1983).
moving window centered on each pixel in the image. Additionally, whether neighboring
pixels appear different or not can be affected by inning of the original data. The original
image has a possible 256 gray levels. If, for example, the image data are binned into 64
possible values, then gray levels that were originally different but close together would
often be assigned the same gray level. Because the window size used for the P table, the
window size of the Q matrix, and the inning range can all influence the resulting
textures, sensitivity tests were run on the ww2042 image to find suitable values for
application to the other images.
Sun and Wee (1983) discuss five textures generated from the NGLDM Q matrix
which are analogous to the Herelick et al. (1973) textures. These textures are replicated in
the program TextureMaker.
K S K S
Small Number Emphasis: N1 = I [Q(k,s)/s2] 1 Q(k,s)
k=l s=l k s=l
K S K S
Large Number Emphasis: N2 = _[s2Q(ks)] Q(k,s)
k=l s=l k=l s=l
S K 2 K S
Number Non-Uniformity: N3 = I YQ[k, s) Q(k, s)
s=l k=l k s=l
K S K S
NGLDM Second Moment: N4= [Q(k, s)]2 Q(k,s)
k=l s=l k=l s=l
K S IK S
NGLDM Entropy: N5= I ,Q(k,s)log[Q(k,s)] Q(k,s)
k=l s=l kl s=l
Small Number Emphasis is a measure of the fineness of the image. In a fine grain
image, the largest frequency of values will be in the small NGLDM numbers making
Q(k,s)/s2 larger. The reverse is true when the matrix element is multiplied by s2 rather
than divided. In this case, the Large Number Emphasis is a measure of the coarseness of
the image. Second Moment is a measure of homogeneity. In a homogeneous image, the Q
matrix will mostly be filled with small values, and, as a result, the NGLDM Second
Moment, which is the sum of squares of the entries, will be smaller. Number Non-
Uniformity and NGLDM Entropy are measures of the coarseness of an image determined
from the homogeneity of the number distribution in the Q matrix.
Univariate Classification Trials
Decision rules were created for each of the textures to test texture signatures
against the calibration images and validation images. These rules were used to classify
each of the images. To create the decision rules, kernel density signatures (Silverman
1986) were generated in S-Plus for each vegetation type combining data from all the
calibration images that contain a particular vegetation type. The number of output points
in the kernel density was set equal to the range of data so that ay value (density) would
be present for each integer x value. That output objective matches the integer values of
the texture images so that the probability that a particular image value belongs to a
particular vegetation class can be determined. The program DensityThreshold was used
to remove the tails of the kernel distributions from each vegetation class signature.
Because the x values output from S-plus denotes real values, the program also rounds to
the nearest integer. Rounding results in some missing integer values which the program,
extrapolating between neighboring values for a smooth density function, fills in. Finally,
all of the vegetation class signatures are combined in the Uclassify program to create a
decision vector. This vector is a simple list of the vegetation class that will be assigned to
a pixel when it has a given x value. The decision vector is created for each x value in the
range of observed x values by finding the highest density value (y value) among the
vegetation signatures and assigning that class to that x location. Figure 3-3 and Table 3-1
present an example of the procedure using the mean texture signatures. Once a decision
vector is created for a particular texture, it can be used repeatedly in the Uclassify
program for rapid classification of new texture images.
Mean Texture Signatures
0.25
0.2
S1. Schinus
--2. Palm
0.15 -- 3. Water
ft -*X--4. Mangrove
L -5. Herb
6. Shrub
0.1.- -j--- 7. Typha
S- 8. Myrica
9. Hardwood
0.05
DN
Figure 3-3. Combined signatures for the Mean Texture.
Table 3-1. Decision vector for the signatures shown in Figure 3-3.
Texture 85- 93- 108- 113- 122- 128- 135- 147- 156- 160- 182-
Value <85 92 107 112 121 126 127 134 145 146 155 159 181 191 >191
Class 0 3 0 9 7 6 7 6 8 5 4 5 1 2 0
Reference images (see Chapter 2) for each of the calibration and validation
images are used in the ErrorMatrix program to generate accuracy assessments of the
texture classifications. ErrorMatrix creates a contingency table (Congalton and Green
1999) of the distribution of classified pixels in relationship to the class of those pixels in
the reference image. The contingency table also reports user and producer accuracies by
class and overall accuracies. Producer accuracies are defined as the percent of area of a
specific class on the ground that is identified as that class on the map. User accuracies are
defined as the percent of areas called a specific class on the map that are that class on the
ground. The corollaries to user and producer accuracies are the omission and commission
errors. Omission error is when an area is excluded from the class to which it belongs.
Omission error is equal to 1 minus the producer's accuracy. Commission error, equal to 1
s accuracy, is when an area is included in an incorrect class. Figure 3-4
illustrates these definitions.
Ground Class A Classified as Class A
SI
S Agreement J '
Commission
Error
Ommission n1 .,. ii
Error
Figure 3-4. Illustration of accuracy assessment measures. In this example producer's
accuracy is high because most of the class has been correctly mapped and omission error
is low. User's accuracy is low because the mapped class includes a large area that is
misidentified and, therefore, commission error is high.
Results
Signature Development
Figures 3-5 through 3-7 show the distributions of the original gray level values for
each vegetation cover type in the infrared, red, and green bands. While the distribution of
Schinus gray levels values is generally higher than for the other species, the extreme
overlap between species distributions makes it impossible to distinguish one species from
another based solely on spectral values. The infrared band stretches the signatures apart
somewhat better than the red or green bands; however, the individual species remain
indistinguishable.
First-order statistical texture signatures for vegetation classes in the infrared of
image ww2042 are graphed for moving window neighborhoods ranging from 7 to 63
pixels in Figures 3-8 through 3-11. Texture signatures for green and red bands at various
window sizes were also graphed but have less separation than for the infrared and so are
not presented here. Separation of species resulting from sharper peaks in the species'
distributions, at least for the mean texture, improves with increasing window size. First-
order statistical textures of all the calibration image textures are graphed using the
infrared band at a moving window size of 63x63 pixels in Figures 3-12 through 3-17.
From all the 1st order textures in all the images, only the mean in the infrared band
and, to lesser extent, the standard deviation in the infrared band, showed any
discrimination of the vegetation types. In the three images containing Schinus, its range
for the mean texture was a consistent 150 to 200 (see Figure 3-12). Separation from the
range of values representing Schinus was excellent for Myrica, Typha, hardwoods, shrub
and, of course, water. Sabal and herbaceous separate in image ww2046 and both occur at
higher mean texture values than the Schinus found in other images. However, the Sabal
signature in ww1085 and the herbaceous and Avicennia signatures in ww1099 and
ww1093 are all within the same range as the Schinus signatures and will confuse with
Schinus in a classification. The standard deviation texture separates the herbaceous class
from Schinus with little overlap in both ww1093 and ww1099. Shrub also separates from
Schinus in ww1099. The other cover types do not separate at all. The skew texture does
not separate any of the cover types tested. Similarly, the entropy, homogeneity, and
Simpson's diversity textures do not separate any of the cover types except water.
Figure 3-18 graphs the Sun and Wee (1983) 2nd order or co-occurrence texture Q
matrix for vegetation classes in image ww2042 at moving window neighborhoods for the
P table of 3x3 and 7x7 and a inning of the data values to a maximum range of 64. There
is no effective discrimination between classes. The vegetation classes all fall in the lower
left-hand side of the Q matrix suggesting generally fine-textured classes with
approximately equal gray levels. Because the NGLDM textures are derived from the Q
matrix, they also show no discrimination of vegetation classes and are not presented here.
Univariate Calibration Image Classifications
The mean and the standard deviation textures were selected for classification
trials because they were the only textures to exhibit any discrimination between the
classes. Core areas of Schinus were discriminated when the mean texture was used as the
classifier for images ww2042 and ww1093 and core areas of Myrica discriminate in
ww2042 (Figure 3-19). In both images the classifications fail at the edges of the
vegetation types. Classification of the other calibration images with the mean texture was
completely unsuccessful. The accuracy assessments for the calibration images generally
show much higher user's accuracies than producer's accuracies (Tables 3-2 through 3-6).
This result is from the classifications correctly identifying vegetation types, but with
large omission errors. Image ww1085 is a good example. When hardwoods and Sabal are
identified, it is with 100% and 73% accuracy, respectively (user's accuracy); however,
hardwoods and Sabal are rarely identified, so the producer's accuracy is only 7% and
11% (Table 3-2). Image ww1093 has user accuracies of 73% and 100% for Schinus and
water and producer's accuracies of 54% for both. Avicennia and herbaceous cover types
both have low user and producer's accuracies, however (Table 3-3). Image ww1099 has
50% user's and producer's accuracy for Schinus and higher user's accuracies but low
producer's accuracies for herbaceous and shrub (Table 3-4). Schinus and water in image
ww2042 perform well (88% and 100% user's accuracy and 62% and 71% producer's
accuracy, respectively); however, Myrica confuses with Typha resulting in lower
accuracies for these types (Table 3-5). Table 3-6 shows that Sabal and herbaceous
classifications in image ww2046 resulted in poor accuracies.
Classification of the calibration images using the standard deviation texture
results in a very high level of confusion among classes (Figure 3-20). With the exception
of water, only hardwoods in image ww1085 show any indication of being successfully
identified. Even in the case of hardwoods, overall accuracy is very poor (Table 3-7).
While there is a 75% probability that a location identified as hardwood will be hardwood,
there is only a 10% probability that hardwoods will be identified. Water was identified
with 88% and 96% user's accuracy in images ww1093 (Table 3-8) and ww2042 (Table 3-
10), respectively. Producer's accuracies for water in those two images were 63% and
72% respectively. Accuracies for the other cover types were generally low (Tables 3-7
through 3-11).
Univariate Validation Image Classifications
When the mean and standard deviation textures where applied to the validation
images, the results showed no relationship to reality (Figures 3-21 and 3-22). In both
cases, the highest accuracies (excluding water) were for herbaceous ground cover in
image ww2008 (Tables 3-13 and 3-17). With user's accuracies of 87% for the mean
texture and 98% for the standard deviation texture, but producer's accuracies of only 12%
and 15%, herbaceous ground cover showed the same relationship of use and producer
accuracies that was seen for hardwoods using the standard deviation texture in ww1085.
That Schinus has a 93% user's accuracy in the standard deviation texture image 0212-528
(Table 3-16) is not meaningful since almost the entire image is Schinus; therefore,
anything that is classified as Schinus is correctly classified. More significant is the 4%
producer's accuracy that points to the large omission error in classification. It is not
surprising that Schinus in image 0212-528 was poorly discriminated by the mean and
standard deviation textures. The overexposed image has a direct effect on both measures.
Other cover types in the validation images have very low accuracies (Tables 3-12 through
3-19).
Discussion
Widely overlapping distributions of the original gray level values were expected
because of the high resolution of the imagery. A cover type has a mixture of distinct
spectral responses due to shadows, stems, species mixes, exposed ground, adjacent land
covers, and morphological variability. In larger (low-resolution) pixels, spectral
heterogeneity averages to characterize the overall spectral response of a class. In these
high-resolution images, the variance of spectral responses is not averaged and a broad
spectral distribution results.
For similar reasons, larger neighborhoods of analysis (analysis window sizes)
often discriminate between cover types better than small neighborhoods as was observed
in these analyses. The larger window accounts for more of the variance in the cover type.
The tradeoff is slower processing time and less precision in accurately classifying the
location of cover type edges.
In his trials with synthetic imagery, Weszka et al. (1976) reported that the simple
mean was the best discriminator among the first-order or second-order (Haralick textures)
statistics that he tried. That was also true for these images. While the mean gray level
value in a neighborhood often separated Schinus from other species in the imagery, it,
unfortunately, is not a gray level invariant texture and, therefore, is unlikely to be
consistent from image to image.
The co-occurrence matrices generated for these images showed no effective
separation of the cover types. The fine-grain versus coarse-grain distinctions that co-
occurrence textures emphasize may be valuable for separation of urban versus natural
cover types. However, the second-order gray-level variance about the natural cover type
images used here was not adequately different from one another. Sun and Wee (1983)
have applied these textures successfully to geologic features selected from Landsat-1
images. Taken by themselves, individual first and second-order textures do not
35
discriminate Schinus or the other tested cover types with adequate accuracy or
consistency.
ww1085
a,
a
9 -
0
0
a -
o
d
Co
o
0
0
d
cc.
o
(0
o -
o
0
d
ww1093
Sabal
Hardwoods
0 50 100 150 200 250
Infrared Band
ww1099
Schinus
Herbaceous
Shrub
It
/*,**
0 50 100 150 200 250
Infrared Band
ww2046
Sabal
Herbaceous
0 50 100 150 200 250
Infrared Band
a,
0-
C)
o
9 -
0
d
0 -
0
Figure 3-5. Spectral signatures from the infrared image band.
ISchinus
Herbaceous
iAvicennia .......
Water .--.---
it
0 50 100 150 200 250
Infrared Band
ww2042
; Schinus
pl Myrica
i1 Typha ..
S Water --.---
iI
ii
I
0 50 100 150 200 250
Infrared Band
ww1085
a00
0
0
0
a0
o .
0
o
0
0
0 .
0
o
ww1093
Sabal
Hardwoods
0 50 100 150 200 250
Red Band
ww1099
Schinus
Herbaceous
Shrub -----
0 50 100 150 200 250
Red Band
ww2046
Sabal
Herbaceous -
/
0 50 100 150 200 250
Red Band
00
0 -
0
0
o
0
0
0
9 -
0
0
o
Figure 3-6. Spectral signatures from the red image band.
iSchinus
iHerbaceous
iAvicennia ................
SWater .-----
lii
0 50 100 150 200 250
Red Band
ww2042
Schinus
Myrica
Typha ................
Water --.----
if\
i
0 50 100 150 200 250
Red Band
ww1085
aCO
9 -
0
o
0 .
0
0
0
0
0
CO
o
a>
0
0
0
00
*M
8
ww1093
Sabal
Hardwoods
0 50 100 150 200 250
Green Band
ww1099
Schinus
Herbaceous
Shrub -----
i\
0 50 100 150 200 250
Green Band
ww2046
Sabal
Herbaceous
I
0 50 100 150 200 250
Green Band
aO
0 -
SO
9 -
a -
0
0
0
d
Figure 3-7. Spectral signatures from the green image band.
P Schinus
n Herbaceous
ii Avicennia .......
i Water .-----
0 50 100 150 200 250
Green Band
ww2042
Schinus
Myrica
Typha ................
S Water --.
is
0 50 100 150 200 250
Green Band
ww2042 7x7 neighborhood
Schinus
Myrica
Typha ---
Water -.----
I
50 100 150 200 250
Mean
ww2042 7x7 neighborhood
Schinus i
Myrica
Typha ..........
Water ----- -
I
110 115 120 125 130
Skew
ww2042 7x7 neighborhood
0 10 20 30 40 50
Homogeneity
ww2042 7x7 neighborhood
*! Schinus
I 1 Myrica
ti \Typha .......
i Water -------
I .'
0 10 20 30 40 50
Standard Deviation
ww2042 7x7 neighborhood
1 Schinus
i Myrica
\ Typha ..
j Water --.---
0 2 4 6 8 10
Entropy
ww2042 7x7 neighborhood
Schinus
Myrica
Typha
Water ----.
-" .\^ 1
140 160 180 200
Simpson Diversity
Figure 3-8. ww2042 spectral signatures of the first-order textures with a window size
7x7.
ww2042 15x15 neighborhood
Schinus
Myrica
i Typha ----
Water -------
50 100 150 200 250
Mean
ww2042 15x15 neighborhood
Schinus i
Myrica --
Typha .....
Water -------
t*
110 115 120 125 130
Skew
ww2042 15x15 neighborhood
Schinus
ft Myrica
ITypha ------------.
Water ....
I Water --------
: \\
0 10 20 30 40 50
Homogeneity
O -
ww2042 15x15 neighborhood
Schinus
Myrica
TTypha ---
SWater -- ---
0 10 20 30 40 50
Standard Deviation
ww2042 15x15 neighborhood
Schinus
Myrica
i Typha ..
Water -------
I i
i
r i__
0 2 4 6 8 10
Entropy
ww2042 15x15 neighborhood
Schinus
Myrica
Typha ..
Water ----
i /
.... . ........... .
140 160 180 200
Simpson Diversity
Figure 3-9. ww2042 spectral signatures of the first-order textures with a window size
15x15.
ww2042 31x31 neighborhood
Schinus
Myrica
Typha --
Water --.---
r5
i ft
I /0 1
50 100 150 200 250
Mean
ww2042 31x31 neighborhood
Schinus I
Myrica i
Typha -........... -
Water ------ i
1j I
110 115 120 125 130
Skew
ww2042 31x31 neighborhood
Schinus
Myric ---
Typh ------
r Water -------
f
Ii
0 10 20 30 40 50
Homogeneity
o -
ww2042 31x31 neighborhood
0 10 20 30 40 50
Standard Deviation
ww2042 31x31 neighborhood
S ;hinus -
i! M rica
i T ha
\ ter -----
I |
0 2 4 6 8 10
Entropy
ww2042 31x31 neighborhood
Schinus
Myrica --. -t
Typha .......
Water -----.
i 1
i I .
140 160 180 200
Simpson Diversity
Figure 3-10. ww2042 Spectral signatures of the first-order textures with a window size
31x31.
ww2042 63x63 neighborhood
S Schinus
i IVyrica
I %--yp h a - - ---. -
I A' ,/. >- . _. ... .
50 100 150 200 250
Mean
ww2042 63x63 neighborhood
Schinus t ill
Myrica ----
Typha -----
Water ----.-
I ii
110 115 120 125 130
Skew
ww2042 63x63 neighborhood
iII Schiriu~
Myrida
Typha ---------------
SWater .-------
lll I
I "I
0 10 20 30 40 50
Homogeneity
o -
ww2042 63x63 neighborhood
Schinus
Myrica
Typha
I Water --.---
'i1:
i / .'... ......
0 10 20 30 40 50
Standard Deviation
ww2042 63x63 neighborhood
i! S hinus
i! -rica
T ha ...
Vter --.-----
A 1_
0 2 4 6 8 10
Entropy
ww2042 63x63 neighborhood
Schinus
Myrica --
Typha ----
Water ----
140 160 180 200
Simpson Diversity
Figure 3-11. ww2042 Spectral signatures of the first-order textures with a window size
63x63.
ww1085
Sabal
Hardwoods
50 100 150 200 250
Mean
ww1099
50 100 150 200 250
Mean
ww2046
Sabal
Herbaceous
I
I
50 100 150 200 250
Mean
ww1093
cO
0
0
a,
0 .
0o
o
50 100 150 200 250
Mean
ww2042
0 -
0
O
0
0
O
50 100 150 200 250
Mean
Figure 3-12. Mean texture signatures.
cO
0
C,
o d
0
C)
1"q
HerbJ ous
i A ................
1 vi a
W a te ;c --. -. --
Schinus
jVlyrica
Sypha ---
00
0
0
0
o
c -
0 .
0
0
d
(0
d
r
O
'>~
o O
O
0
(D
d
. O
'>~
, (
d
0
0
ww1093
ww1085
Sabal
Hardwoods
10 15 20 25 30 35 40
Standard Deviation
ww1099
Schinus
Shrub
Herb ---------
: I,
JI Ii
10 15 20 25 30 35 40
Standard Deviation
ww2046
Sabal
Herbaceous
,i
(D
>,6 O
0
d
(D
O
, (
d
10 20 30 40 50 60 70
Standard Deviation
Figure 3-13. Standard Deviation texture signatures.
1 Schinus
Avicennia
I Water .-----
hi S
10 15 20 25 30 35 40
Standard Deviation
ww2042
Schinus
Myrica
Typha .......
Water -------
10 15 20 25 30 35 40
Standard Deviation
(q
o
0
0
O
ww1085
Sabal
Hardwood s
S
I I
110 115 120 125 130 135 140
Skew
ww1099
er p ------------------
hru ----
erb
0
110 115 120 125 130 135 140
Skew
ww2046
SSabal
Herbaceous
0
o
110 115 120 125 130 135 140
Skew
Figure 3-14. Skew texture signatures.
ww1093
Schinusti
Avicenni
Herb ...
Water .-----
110 115 120 125 130 135 140
Skew
ww2042
110 115 120 125 130 135 140
Skew
ww1085
mo -
ww1093
mo -
Sabal
Hardwoods -
2 3 4 5 6
Entropy
ww1099
Sh inus
Slrub .
I rib ----------
~rb
i
I
I I
2 3 4 5 6
Entropy
ww2046
Sab i
Herc ous i
I iI
II I
I II
I
2 3 4 5 6
Entropy
Figure 3-15. Entropy texture signatures.
Schirus
Avicennia -
Herb
Water ...
2 3 4 5 6
Entropy
ww2042
Schinus
Myrica -
Typha ..............
Water --.----
I :I:
i :
2 3 4 5 6
Entropy
ww1085
ww1093
Sd
0
0
0
0 10 20 30 40
Homogeneity
ww1099
Schinus
Shrub
Herb
0 10 20 30 40
Homogeneity
0 10 20 30 40
Homogeneity
ww2042
00
0.
Schinus
Myrica
Typha .......
Water -------
0 10 20 30 40
Homogeneity
0 10 20 30 40
Homogeneity
Figure 3-16. Homogeneity texture signatures.
0
a>,
0
0
cO
0
o
d
ww2046
>,
s 5
0
0
0
ww1085
d
a, 5
Sd
o,
0
ww2046
d
a o
S
0
0
ww1093
d
o d
Sabal -
Hardwoods -
II
II
150 160 170 180 190 200
Simpson Diversity
ww1099
Schinus
Shrub
Herb .....-
Sii
150 160 170 180 190 200
Simpson Diversity
Schinus
Avicennia
Herb
Water .--.--
:_:
150 160 170 180 190 200
Simpson Diversity
ww2042
Schinus -
Myrica
Typha ----
Water -----.- I
'I
150 160 170 180 190 200
Simpson Diversity
150 160 170 180 190 200
Simpson Diversity
Figure 3-17. Simpson's Diversity texture signatures.
0
d
'>,
a, h
O4
o
0
0
,I
0
Sabal |
Herbaceous
II
II
I 1
t!
w wD2 E cIhls Fwnw 3 QyrslW- 54
3lhiw
J- 1 'i f L
WWM42 kkr y Pyo i- 3 Q ebrw-&m4
AiM
*w. rW42pha pnndw-= 7 CWnd'= c 4
wil
yvwB342 7Wpa rnd h= 3 QCrdm= 64
wnc~-~pr Rc~sxu-- A V,~ 6
Figure 3-18. Normalized Gray Level Difference Matrix contour plots for vegetation
covers in image ww2042.
mww2D2 SchnnuaF PFrwdow 7 awr:w m B4
,.q(>
WWZN2 hIca fTwmfkw 7a nMOw 54
' "
La.
c. ww1099
- SchiWnus terebinthifolius
S Sabal palmetto
-l Open Water
Avicennla germnnans
Herbaceous
Shrub
Typha spp.
I Myrica cerifera
I I Hardwoods
e. ww2U46
Figure 3-19. Classification of calibration images from the Mean texture.
Table 3-2. Contingency table for ww1085 Mean textures classification.
Reference ww1085 Mean row produce
0 1 2 3 4 5 6 7 8 9 count Accuracy
UnClassified 0 0 82 0 0 68 60 9 1 25 0 245 0.0%
Schinus 1 0 0 0 0 0 0 0 0 0 0 NA
Sabal 2 0 79 14 0 15 8 0 0 8 0 124 11.3%
Water 3 0 0 0 0 0 0 0 0 0 0 NA
Avicennia 4 0 0 0 0 0 0 0 0 0 0 NA
Herbaceous 5 0 0 0 0 0 0 0 0 0 0 NA
Shrub 6 0 0 0 0 0 0 0 0 0 0 NA
ll, 7 0 0 0 0 0 0 0 0 0 0 0 NA
Myrica 8 0 0 0 0 0 0 0 0 0 0 NA
Hardwood 9 207 85 5 1 209 76 951 787 489 221 3031 7.3%
column count 207 246 19 1 292 144 960 788 522 221 340
User Accuracy 0.0% 0.0% 73.7% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 100.0%
Overall Accuracy 6.9%
Table 3-3. Contingency table for ww1093 Mean textures classification.
Reference ww1093 Mean row Producer
0 1 2 3 4 5 6 7 8 9 count Accuracy
UnClassified 0 0 136 0 0 149 76 243 29 178 0 811 0.0%
Schinus 1 13 541 37 0 173 147 1 0 78 0 990 54.6%
Sabal 2 0 0 0 0 0 0 0 0 0 0 O NA
Water 3 144 0 0 257 0 0 15 32 3 24 475 54.1%
Avicennia 4 0 1 0 0 101 29 50 24 61 1 267 37.8%
Herbaceous 5 0 63 0 0 217 80 19 2 136 0 517 15.5%
Shrub 6 0 0 0 0 0 0 0 0 0 0 0 NA
il. 7 0 0 0 0 0 0 0 0 0 0 NA
Myrica 8 0 0 0 0 0 0 0 0 0 0 NA
Hardwood 9 0 0 0 0 0 0 0 0 0 0 NA
column count 157 741 37 257 640 332 328 87 456 25 3060
UserAccuracy 0.0% 73.0% 0.0% 100.0% 15.8% 24.1% 0.0% 0.0% 0.0% 0.0%
Overall Accuracy 32.0%
Table 3-4. Contingency table for ww1099 Mean textures classification.
Reference ww1099 Mean row Producer
0 1 2 3 4 5 6 7 8 9 count Accuracy
UnClassified 0 0 17 1 0 21 10 16 0 42 0 107 0.0%
Schinus 1 128 389 89 0 74 63 12 0 39 0 794 49.0%
Sabal 2 0 0 0 0 0 0 0 0 0 0 NA
Water 3 0 0 0 0 0 0 0 0 0 0 NA
Avicennia 4 0 0 0 0 0 0 0 0 0 0 NA
Herbaceous 5 108 353 174 0 383 207 11 0 143 0 1379 15.0%
Shrub 6 0 6 4 0 66 19 153 19 148 0 415 36.9%
lil't 7 0 0 0 0 0 0 0 0 0 0 NA
Myrica 8 0 0 0 0 0 0 0 0 0 0 NA
Hardwood 9 0 0 0 0 0 0 0 0 0 0 0.0%
column count 236 765 268 0 544 299 192 19 372 2695
User Accuracy 0.0% 50.8% 0.0% NA 0.0% 69.2% 79.7% 0.0% 0.0%
Overall Accuracy 27.8%
Table 3-5. Contingency table for ww2042 Mean textures classification.
Reference ww2042 Mean row Producer
0 1 2 3 4 5 6 7 8 9 count Accuracy
UnClassified 0 2 21 0 0 32 13 22 5 44 5 144 1.4%
Schinus 1 0 573 66 0 115 96 14 7 51 0 922 62.1%
Sabal 2 0 15 2 0 2 2 0 0 0 0 21 9.5%
Water 3 111 0 0 377 0 0 6 18 0 19 531 71.0%
Avicennia 4 0 0 0 0 0 0 0 0 0 0 NA
Herbaceous 5 0 0 0 0 0 0 0 0 0 0 NA
Shrub 6 0 0 0 0 0 0 0 0 0 0 NA
i/l/.t 7 2 19 0 0 80 29 152 77 266 10 635 12.1%
Myrica 8 0 26 0 0 94 35 17 0 263 0 435 60.5%
Hardwood 9 0 0 0 0 0 0 0 0 0 0 NA
column count 115 654 68 377 323 175 211 107 624 34 2688
User Accuracy 1.7% 87.6% 2.9% 100.0% 0.0% 0.0% 0.0% 72.0% 42.1% 0.0%
Overall Accuracy 48.1%
Table 3-6. Contingency table for ww2046 Mean textures classification.
Reference ww2046 Mean row Producel
0 1 2 3 4 5 6 7 8 9 Count Accuracy
UnClassified 0 249 276 112 1 118 54 88 37 107 7 1049 23.7%
Schinus 1 0 0 0 0 0 0 0 0 0 0 NA
Sabal 2 715 151 111 0 22 11 3 3 11 0 1027 10.8%
Water 3 0 0 0 0 0 0 0 0 0 0 NA
Avicennia 4 0 0 0 0 0 0 0 0 0 0 NA
Herbaceous 5 26 8 6 0 2 3 0 1 2 0 48 6.3%
Shrub 6 0 0 0 0 0 0 0 0 0 0 NA
,liph 7 0 0 0 0 0 0 0 0 0 0 NA
Myrica 8 0 0 0 0 0 0 0 0 0 0 NA
Hardwood 9 0 0 0 0 0 0 0 0 0 0 NA
column count 990 435 229 1 142 68 91 41 120 7 2124
User Accuracy 25.2% 0.0% 48.5% 0.0% 0.0% 4.4% 0.0% 0.0% 0.0% 0.0%
Overall Accuracy 17.1%
b. ww1093
C. WWIUYY
U. WWZU4Z
M Schinus terebinthifollus
- Sabal palmetto
M Open Water
Avicennia germinans
Herbaceous
Shrub
Typha spp.
M Myrica cerifera
SI Hardwoods
Figure 3-20. Classification of calibration images from the Standard Deviation texture.
e. ww2046
Table 3-7. Contingency table for ww1085 Standard Deviation textures classification.
Reference ww1085 Standard Deviation row Producel
0 1 2 3 4 5 6 7 8 9 Count Accuracy
UnClassified 0 28 14 86 0 6 1 0 0 26 84 245 11.4%
Schinus 1 0 0 0 0 0 0 0 0 0 0 NA
Sabal 2 90 1 14 0 0 0 0 0 1 18 124 11.3%
Water 3 0 0 0 0 0 0 0 0 0 0 NA
Avicennia 4 0 0 0 0 0 0 0 0 0 0 NA
Herbaceous 5 0 0 0 0 0 0 0 0 0 0 NA
Shrub 6 0 0 0 0 0 0 0 0 0 0 NA
,lia 7 0 0 0 0 0 0 0 0 0 0 NA
Myrica 8 0 0 0 0 0 0 0 0 0 0 NA
Hardwood 9 198 594 985 0 221 88 14 0 620 311 3031 10.3%
column count 316 609 1085 0 227 89 14 0 647 413 340
User Accuracy 8.9% 0.0% 1.3% NA 0.0% 0.0% 0.0% NA 0.0% 75.3%
Overall Accuracy
10.4%
Table 3-8. Contingency table for ww1093 Standard Deviation textures classification.
Reference ww1093 Standard Deviation row Producel
0 1 2 3 4 5 6 7 8 9 count Accuracy
UnClassified 0 37 98 152 11 77 92 25 0 255 64 811 4.6%
Schnus 1 0 257 466 0 39 1 0 0 194 33 990 26.0%
Sabal 2 0 0 0 0 0 0 0 0 0 0 NA
Water 3 6 14 56 298 7 23 9 0 26 36 475 62.7%
Avicenna 4 9 46 115 0 10 0 0 0 32 55 267 3.7%
Herbaceous 5 2 86 60 28 43 101 59 0 123 15 517 19.5%
Shrub 6 0 0 0 0 0 0 0 0 0 0 NA
7 0 0 0 0 0 0 0 0 0 0 NA
8 0 0 0 0 0 0 0 0 0 0 NA
Hardwood 9 0 0 0 0 0 0 0 0 0 0 NA
column count 54 501 849 337 176 217 93 0 630 203 306
UserAccuracy 68.5% 51.3% 0.0% 88.4% 5.7% 46.5% 0.0% NA 0.0% 0.0%
Overall Accuracy 23.0%
Table 3-9. Contingency table for ww1099 Standard Deviation textures classification.
Reference ww1099 Standard Deviation row Producel
0 1 2 3 4 5 6 7 8 9 count Accuracy
UnClassified 0 10 7 45 1 3 5 1 0 13 22 107 9.3%
Schinus 1 113 104 266 3 26 3 0 0 97 182 794 13.1%
Sabal 2 0 0 0 0 0 0 0 0 0 0 0 NA
Water 3 0 0 0 0 0 0 0 0 0 0 0 NA
Avicennia 4 0 0 0 0 0 0 0 0 0 0 0 NA
Herbaceous 5 39 131 257 222 86 254 156 0 141 93 1379 18.4%
Shrub 6 0 47 33 9 21 126 103 0 72 4 415 24.8%
Ilih 7 0 0 0 0 0 0 0 0 0 0 0 NA
Myrica 8 0 0 0 0 0 0 0 0 0 0 0 NA
Hardwood 9 0 0 0 0 0 0 0 0 0 0 0 NA
column count 162 289 601 235 136 388 260 0 323 301 2695
User Accuracy 6.2% 36.0% 0.0% 0.0% 0.0% 65.5% 39.6% NA 0.0% 0.0% _
Overall Accuracy
17.5%
Table 3-10. Contingency table for ww2042 Standard Deviation textures classification.
Reference ww2042 Standard Deviation row Producel
0 1 2 3 4 5 6 7 8 9 count Accuracy
UnClassified 0 2 35 57 0 4 0 0 0 29 17 144 1.4%
Schinus 1 17 257 344 0 60 0 0 0 186 58 922 27.9%
Sabal 2 0 2 11 0 0 0 0 0 0 8 21 52.4%
Water 3 1 20 45 384 10 15 18 0 13 25 531 72.3%
Avicennia 4 0 0 0 0 0 0 0 0 0 0 NA
Herbaceous 5 0 0 0 0 0 0 0 0 0 0 NA
Shrub 6 0 0 0 0 0 0 0 0 0 0 NA
/l'ih.t 7 21 124 136 15 49 53 44 0 150 43 635 0.0%
Myrica 8 0 120 125 0 40 26 1 0 123 0 435 28.3%
Hardwood 9 0 0 0 0 0 0 0 0 0 0 NA
column count 41 558 718 399 163 94 63 0 501 151 268
User Accuracy 4.9% 46.1% 1.5% 96.2% 0.0% 0.0% 0.0% NA 24.6% 0.0%
Overall Accuracy 28.9%
Table 3-11. Contingency table for ww2046 Standard Deviation textures classification.
Reference ww2046 Standard Deviation row Producel
0 1 2 3 4 5 6 7 8 9 count Accuracy
UnClassified 0 1016 1 7 0 1 1 0 0 1 22 1049 96.9%
Schinus 1 0 0 0 0 0 0 0 0 0 0 NA
Sabal 2 956 8 23 1 1 2 0 0 5 31 1027 2.2%
Water 3 0 0 0 0 0 0 0 0 0 0 NA
Avicennia 4 0 0 0 0 0 0 0 0 0 0 NA
Herbaceous 5 27 4 6 0 0 0 0 0 3 8 48 0.0%
Shrub 6 0 0 0 0 0 0 0 0 0 0 NA
,liph 7 0 0 0 0 0 0 0 0 0 0 NA
Myrica 8 0 0 0 0 0 0 0 0 0 0 NA
Hardwood 9 0 0 0 0 0 0 0 0 0 0 NA
column count 1999 13 36 1 2 3 0 0 9 61 212
User Accuracy 50.8% 0.0% 63.9% 0.0% 0.0% 0.0% NA NA 0.0% 0.0% _
Overall Accuracy 48.9%
b. ww2008
c. ww2060
c. ww2060
- Schinus terebinthifolus
- Sabal palmetto
- Open Water
Avicennla germfnans
Herbaceous
Shrub
Typha spp.
S Myria cerifera
I Hardwoods
Figure 3-21. Classification of validation images from the Mean texture.
Table 3-12. Contingency table for 0212-528 Mean textures classification.
Reference 0212-528 Mean Texture Classification row Producer
0 1 2 3 4 5 6 7 8 9 count Accuracy
UnClassified 0 791 156 80 0 37 24 0 0 8 0 1096 72.2%
Schinus 1 3400 88 93 0 6 7 0 0 0 0 3594 2.4%
Sabal 2 0 0 0 0 0 0 0 0 0 0 0 NA
Water 3 0 0 0 0 0 0 0 0 0 0 0 NA
Avicennia 4 0 0 0 0 0 0 0 0 0 0 0 NA
Herbaceous 5 0 0 0 0 0 0 0 0 0 0 0 NA
Shrub 6 0 0 0 0 0 0 0 0 0 0 0 NA
inl'.t 7 0 0 0 0 0 0 0 0 0 0 0 NA
Myrica 8 0 0 0 0 0 0 0 0 0 0 0 NA
Hardwood 9 0 0 0 0 0 0 0 0 0 0 0 NA
column count 4191 244 173 0 43 31 0 0 8 0 4690
User Accuracy 18.9% 36.1% 0.0% NA 0.0% 0.0% NA NA 0.0% NA
Overall Accuracy 18.7%
Table 3-13. Contingency table for ww2008 Mean textures classification.
Reference ww2008 Mean Texture Classification row Producer
0 1 2 3 4 5 6 7 8 9 count Accuracy
UnClassified 0 4 24 10 0 12 6 0 0 3 0 59 6.8%
Schinus 1 682 274 256 0 25 17 16 0 14 0 1284 21.3%
Sabal 2 0 0 0 0 0 0 0 0 0 0 0 NA
Water 3 0 0 0 0 0 0 0 0 0 0 0 NA
Avicennia 4 0 0 0 0 0 0 0 0 0 0 0 NA
Herbaceous 5 22 453 45 0 272 159 147 6 267 0 1371 11.6%
Shrub 6 0 0 0 0 0 0 0 0 0 0 0 NA
i/l/.t 7 0 0 0 0 0 0 0 0 0 0 0 NA
Myrica 8 0 0 0 0 0 0 0 0 0 0 0 NA
Hardwood 9 0 0 0 0 0 0 0 0 0 0 0 NA
column count 708 751 311 0 309 182 163 6 284 0 2714
User Accuracy 0.6% 36.5% 0.0% NA 0.0% 87.4% 0.0% 0.0% 0.0% NA
Overall Accuracy 16.1%
60
Table 3-14. Contingency table for ww2060 Mean textures classification.
Reference ww2060 Mean Texture Classification row Producer
0 1 2 3 4 5 6 7 8 9 count Accuracy
UnClassified 0 456 240 228 6 150 47 155 61 306 39 1688 27.0%
Schinus 1 180 51 20 0 17 13 3 0 11 0 295 17.3%
Sabal 2 0 0 0 0 0 0 0 0 0 0 0 NA
Water 3 960 2 0 87 7 4 109 111 30 61 1371 6.3%
Avicennia 4 0 0 0 0 0 0 0 0 0 0 0 NA
Herbaceous 5 0 0 0 0 0 0 0 0 0 0 0 NA
Shrub 6 0 0 0 0 0 0 0 0 0 0 0 NA
il't 7 0 0 0 0 0 0 0 0 0 0 0 NA
Myrica 8 0 0 0 0 0 0 0 0 0 0 0 NA
Hardwood 9 0 0 0 0 0 0 0 0 0 0 0 NA
column count 1596 293 248 93 174 64 267 172 347 100 3354
User Accuracy 28.6% 17.4% 0.0% 93.5% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0%
Overall Accuracy 17.7%
Table 3-15. Contingency table for ww2091 Mean textures classification.
Reference ww2091 Mean Texture Classification row Producer
0 1 2 3 4 5 6 7 8 9 count Accuracy
UnClassified 0 122 327 105 9 75 40 52 28 97 10 865 14.1%
Schinus 1 0 0 0 0 0 0 0 0 0 0 0 NA
Sabal 2 1496 147 180 0 12 10 1 0 7 0 1853 9.7%
Water 3 469 14 1 19 25 13 126 337 61 421 1486 1.3%
Avicennia 4 0 0 0 0 0 0 0 0 0 0 0 NA
Herbaceous 5 0 0 0 0 0 0 0 0 0 0 0 NA
Shrub 6 0 0 0 0 0 0 0 0 0 0 0 NA
i/t 7 0 0 0 0 0 0 0 0 0 0 0 NA
Myrica 8 0 0 0 0 0 0 0 0 0 0 0 NA
Hardwood 9 671 455 273 0 74 46 18 1 38 0 1576 0.0%
column count 2758 943 559 28 186 109 197 366 203 431 5780
User Accuracy 4.4% 0.0% 32.2% 67.9% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0%
Overall Accuracy 5.6%
d. ww2091
M Schinus terebinthifolius
- Sabal palmetto
M Open Water
Avicennla gerrmnans
Herbaceous
Shrub
Typha spp.
SMykrica cerifer
SI Hardwoods
Figure 3-22. Classification of validation images from the Standard Deviation texture.
62
Table 3-16. Contingency table for 0212-528 Standard Deviation textures classification.
Reference 0212-528 Standard Deviation Classification row Producer
0 1 2 3 4 5 6 7 8 9 count Accuracy
UnClassified 0 982 10 40 1 1 1 2 0 10 49 1096 89.6%
Schinus 1 2252 129 414 58 48 65 60 0 151 417 3594 3.6%
Sabal 2 0 0 0 0 0 0 0 0 0 0 0 NA
Water 3 0 0 0 0 0 0 0 0 0 0 0 NA
Avicennia 4 0 0 0 0 0 0 0 0 0 0 0 NA
Herbaceous 5 0 0 0 0 0 0 0 0 0 0 0 NA
Shrub 6 0 0 0 0 0 0 0 0 0 0 0 NA
i/l. 7 0 0 0 0 0 0 0 0 0 0 0 NA
Myrica 8 0 0 0 0 0 0 0 0 0 0 0 NA
Hardwood 9 0 0 0 0 0 0 0 0 0 0 0 NA
column count 3234 139 454 59 49 66 62 0 161 466 4690
User Accuracy 30.4% 92.8% 0.0% 0.0% 0.0% 0.0% 0.0% NA 0.0% 0.0%
Overall Accuracy
23.7%
Table 3-17. Contingency table for ww2008 Standard Deviation textures classification.
Reference ww2008 Standard Deviation Classification row Producer
0 1 2 3 4 5 6 7 8 9 count Accuracy
UnClassified 0 9 6 22 0 6 0 1 0 7 8 59 15.3%
Schinus 1 278 155 539 10 22 4 0 0 90 186 1284 12.1%
Sabal 2 0 0 0 0 0 0 0 0 0 0 0 NA
Water 3 0 0 0 0 0 0 0 0 0 0 0 NA
Avicennia 4 0 0 0 0 0 0 0 0 0 0 0 NA
Herbaceous 5 142 139 183 277 62 199 111 0 167 91 1371 14.5%
Shrub 6 0 0 0 0 0 0 0 0 0 0 0 NA
i/l. 7 0 0 0 0 0 0 0 0 0 0 0 NA
Myrica 8 0 0 0 0 0 0 0 0 0 0 0 NA
Hardwood 9 0 0 0 0 0 0 0 0 0 0 0 NA
column count 429 300 744 287 90 203 112 0 264 285 2714
User Accuracy 2.1% 51.7% 0.0% 0.0% 0.0% 98.0% 0.0% NA 0.0% 0.0%
Overall Accuracy 13.4%
63
Table 3-18. Contingency table for ww2060 Standard Deviation textures classification.
Reference ww2060 Standard Deviation Classification row Producer
0 1 2 3 4 5 6 7 8 9 count Accuracy
UnClassified 0 377 122 286 39 64 221 162 0 220 197 1688 22.3%
Schinus 1 153 8 50 0 6 6 4 0 17 51 295 2.7%
Sabal 2 0 0 0 0 0 0 0 0 0 0 0 NA
Water 3 121 49 73 896 21 59 28 0 57 67 1371 65.4%
Avicennia 4 0 0 0 0 0 0 0 0 0 0 0 NA
Herbaceous 5 0 0 0 0 0 0 0 0 0 0 0 NA
Shrub 6 0 0 0 0 0 0 0 0 0 0 0 NA
il.t 7 0 0 0 0 0 0 0 0 0 0 0 NA
Myrica 8 0 0 0 0 0 0 0 0 0 0 0 NA
Hardwood 9 0 0 0 0 0 0 0 0 0 0 0 NA
column count 651 179 409 935 91 286 194 0 294 315 3354
User Accuracy 57.9% 4.5% 0.0% 95.8% 0.0% 0.0% 0.0% NA 0.0% 0.0%
Overall Accuracy
38.2%
Table 3-19. Contingency table for ww2091 Standard Deviation textures classification.
Reference ww2091 Standard Deviation Classification row Producer
0 1 2 3 4 5 6 7 8 9 count Accuracy
UnClassified 0 748 8 28 4 3 1 1 0 13 59 865 86.5%
Schinus 1 0 0 0 0 0 0 0 0 0 0 0 NA
Sabal 2 1606 3 63 0 0 0 0 0 4 177 1853 3.4%
Water 3 172 17 41 1120 13 32 18 0 30 43 1486 75.4%
Avicennia 4 0 0 0 0 0 0 0 0 0 0 0 NA
Herbaceous 5 0 0 0 0 0 0 0 0 0 0 0 NA
Shrub 6 0 0 0 0 0 0 0 0 0 0 0 NA
il.t 7 0 0 0 0 0 0 0 0 0 0 0 NA
Myrica 8 0 0 0 0 0 0 0 0 0 0 0 NA
Hardwood 9 1228 11 108 0 0 0 0 0 13 216 1576 13.7%
column count 3754 39 240 1124 16 33 19 0 60 495 5780
User Accuracy 19.9% 0.0% 26.3% 99.6% 0.0% 0.0% 0.0% NA 0.0% 43.6%
Overall Accuracy 37.1%
CHAPTER 4
EDGE TEXTURES
Introduction
When considering visual discrimination, edges are often the first element that the
eye is drawn to in an image. Myler and Weeks (1993) asserted that edges contain most of
the information within an image. As much as color or gray-tone, edges define the
"texture" of an object by its structure. Edges occur at abrupt changes in image tone and so
are comprised of high-spatial-frequency information because spatial frequency increases
as spatial information increases per unit area. Hlavka (1987) found that one could
distinguish urban and rural areas with edge density textures applied to high altitude
scanner imagery.
In addition to density, edge has other properties that may be useful to texture
analysis as well. Edge strength or magnitude is a measure of how strongly the image tone
changes across the edge. Edge magnitude measures absolute difference. Another measure
is the rate of change across the edge--how abruptly the tonal change occurs. This last
measure can also be considered the magnitude of spatial frequency or change per unit
area.
This chapter considers textures measuring edge density, edge magnitude, and
edge rate of change from images generated by high-pass convolution filters. The
proportion of the images considered to be in the edge is also affected by a lower spectral
threshold value that will be examined.
Methods
Two edge filters were tested, the Laplacian filter and the Sobel filter. The
Laplacian filter enhances abrupt changes in intensity across the image and is non-
directional. The Sobel filter also enhances tonal edges but only in one direction,
horizontal or vertical, at a time (Myler and Weeks 1993). Passing a 3x3 filter over the
image and multiplying the filter values by the image values created Laplacian and Sobel
edge images. The filter values are shown in Figure 4-1. The Sobel filter requires two
passes, one for vertical components and one for horizontal components. The overall
Sobel edge is then computed as the square root of the sum of the squares of the vertical
edge and the horizontal edge.
.1 .1 1 2 1 -1 1
-1 8 -1 0 0 0 -2 0 2
-1 -1 -1 -1 -2 -1 -1 0 1
A B C
Figure 4-1. Values used in 3x3 edge filters. (A) Laplacian edge, (B) horizontal Sobel
edge, and (C) vertical Sobel edge.
To measure density, magnitude, and rate of change in edge pixel values, the
author created six image textures. Laplacian and Sobel edge images are produced as
intermediate edge images from the infrared band of the original images. The texture
images are then produced from characterization of the edge images. The six edge textures
are as follows:
1. Laplacian edge density. The relative density of edges in a moving window
from the Laplacian edge image. Relative density is simply the count of
pixel values above a chosen threshold divided by the total number of
pixels in the moving window. The relative density is multiplied by 100 to
work with integer results.
2. Sobel edge density. The relative density of edges in a moving window
from the Sobel edge image. Relative density is computed in the same
manner as the Laplacian edge density.
3. Median Sobel edge magnitude. The median value in a moving window of
the Sobel edge image.
4. Mean Sobel edge magnitude. The mean value in a moving window of the
Sobel edge image.
5. Sobel edge standard deviation. The standard deviation of the Sobel edge
image in a moving window.
6. Sobel Edge Slope. Computes the average rate of change in edge image
pixel values from the Sobel edge image in a moving window. Pixel value
differences are taken in the horizontal direction and in the vertical
direction and the maximum difference at each pixel in the moving window
is used in the calculation of rate of change. Rate of change is measured as
Z[max Ixy I ,1 I-I 1 V I N 100 where I is the spectral
intensity value of the Sobel edge image at coordinate pair x,y. and Nis the
total number of pixels in the moving window neighborhood.
A seventh texture, image standard deviation, is also included. Image standard
deviation is the standard deviation in a moving window of the original infrared image
values. Standard deviation is another approach to measuring edge strength. It is generally
expected that standard deviation will increase as a neighborhood window moves toward
an edge because the heterogeneity of pixel values increases. This texture is identical to
the standard deviation texture presented in Chapter 3 but is presented here for contrast to
the other edge detectors, particularly the Sobel edge standard deviation texture. The
Laplacian and Sobel edge images and the edge textures were created in the EdgeTextures
program listed in the Appendix.
Decision rules were created for each of the edge textures to test texture signatures
against the calibration and validation images. The decision rules were used to classify
each of the images. Creation of the decision rules, classification and accuracy assessment
proceeded as described in Chapter 3.
Results
Image Signatures
Figures 4-2 and 4-3 illustrate the effect of running a Sobel and Laplacian edge
filter on the calibration images and the visual effect of using a range of threshold values
on the image. The images were sensitive to the threshold value used. A threshold in the
20% range for the Sobel images and approximately 55% for the Laplacian images
seemed visually to give the best image definition. The Sobel images appeared visually to
have more range of edge responses and class separation than the Laplacian images.
Figure 4-4 illustrates the edge image histograms for image wwl085. The wwl085
histograms were characteristic of all the images. The Laplacian edge images had a sharp,
symmetrical peak centered on zero where the filter passes over an edge. It is this
characteristic that gave rise to a threshold near 50% (equal to a pixel value of zero)
displaying the best edge effect. The Sobel edge histogram, on the other hand, was
asymmetric and broader. The peak (mode) of the Sobel edge histogram again was
consistent with the threshold value that best displays edge differences. While the
Laplacian algorithm had a sharper response to edge, the broader histogram of the Sobel
edge allowed a measure of edge strength. It is for this reason that the Sobel edge was
used for all the textures that measure magnitude and variance.
Sensitivity tests of Sobel edge textures suggest that the most successful separation
of vegetation classes was with using a threshold of 10%. The Laplacian edge density
texture provided the most satisfactory results at a threshold of 55%. Sensitivity tests were
also conducted to select a suitable window size for analysis. As for the first order
statistical textures, the window size with the best class separation was the largest one
tested, 63 x 63 pixels. As the window size decreases, the variance of the signatures
increases.
Additionally, texture signatures were computed for both infrared and green bands
of each image. Only infrared band results are reported because they were consistently
superior to the green band results. The one exception was improved discrimination of
Typha in the green band of image ww2042.
Figure 4-5 graphs Laplacian Edge Density signatures for the calibration images
using the 55% threshold and the infrared band. As for the Sobel Edge Densities textures
(Figure 4-6), Laplacian Edge Densities discriminate Schinus and other vegetation types.
The Laplacian Edge Density did not have as good a separation of the signatures as the
Sobel Edge Density textures for some of the images and the Sabalpalmetto, herbaceous,
and hardwood classes did not discriminate at all in images ww1085 and ww2046.
However, herbaceous ground cover was more clearly separated in image ww1093 with
the Laplacian Edge Density than with the Sobel Edge Density. The absolute value of the
signatures for Laplacian Edge Density was not consistent among the images.
Figure 4-6 graphs the signatures of the Sobel Edge Density textures. Sabal
palmetto and hardwoods discriminate in image ww1085. Schinus, Avicennia, herbaceous
ground cover, and water discriminated in image ww1093, although considerable overlap
remains. Water was a single spike at zero and, thus, does not show up in the graph. In
image ww1099, the Sobel edge density texture provided separation for Schinus, shrub,
and herbaceous ground cover with some confusion between shrub and herbaceous
classes. Image ww2042 shows distinct distribution peaks for water, Myrica, and Schinus
with some confusion between Myrica and Schinus. Typha in image ww2042 confuses
with both Myrica and Schinus. The bi-modal distribution for Typha represents a mostly
water component of the signature and a mostly leaf surface component. Sabalpalmetto
and herbaceous ground cover separated for image ww2046. Across all the images
containing Schinus, it always had the highest densities of Sobel edge; however, the
absolute value of the signature range was not consistent from image to image.
Sobel Edge Magnitude was tested as a texture by recording the average and
median magnitude in a moving window, but only the average magnitude is presented
because the results were indistinguishable and the average requires less computation.
Figure 4-7 graphs all the calibration images with the Average Sobel Edge Magnitude
texture. The resulting signatures closely track the results of the Sobel Edge Density
textures (Figure 4-6).
Sobel Edge Slope texture (Figure 4-8) also had similar patterns of signature
discrimination to those seen in the Sobel Edge Densities texture (Figures 4-6). The range
of signature values for Schinus remained fairly stable with the mode falling above 2500
and below 3000 for all the images containing Schinus. In the Sobel Edge Slope textures
for image ww1093, the signatures for herbaceous ground cover and Avicennia do not
discriminate.
Sobel Edge Standard Deviation (Figures 4-9) differs from Image Standard
Deviation (Figure 4-10) in that it only measures the variance values of pixel that are
considered edge pixels. Sobel Edge Standard Deviation does a better job of
discriminating the vegetation classes than Image Standard Deviation. It, again, produced
almost identical, but noisier, signature appearance to the Sobel Edge Magnitude texture.
Sabal palmetto and herbaceous ground cover in image ww2046 did not separate,
however, as they did in the previous Sobel edge textures.
Figure 4-10 graphs the Image Standard Deviation signatures. The Image Standard
Deviation texture is the same algorithm as the first order Standard Deviation in Chapter
3. It is repeated here for contrast to the other edge detectors. Only water was completely
discriminated with this texture. However, as noted in Chapter 3, the herbaceous ground
cover was separated from Schinus in ww1093 and in image ww1099 both herbaceous
ground cover and shrub separated from Schinus, though not from each other.
Univariate Calibration Image Classifications
The Laplacian edge density texture classification shown in Figure 4-11 produced
both high user and producer accuracies in the discrimination of hardwoods in image
ww1085 (Table 4-1). This classification was also somewhat successful in discrimination
of Schinus with a tendency to overestimate its abundance, particularly in image wwl093.
Other vegetation types were poorly discriminated (Table 4-1 through Table 4-5).
The Sobel edge density texture classification (Figure 4-12 and Table 4-6 through
Table 4-10) produced poor results. Hardwoods (Table 4-6) and Schinus (Tables 4-7, 4-8,
and 4-9) both had high user's accuracies because when those classes are identified they
are mostly correct, but omission of large extents of Schinus and hardwoods result in low
producer's accuracies. Figure 4-12 illustrates the confusion when these classes are
mapped. Sabal and hardwoods were most likely to be confused with the Schinus class.
The Sobel edge magnitude classification preformed better in the identification of
Schinus. Accuracy Assessments of the images containing Schinus report a high user's
accuracy but low producer's accuracy (Tables 4-12, 4-13, and 4-14). Although user's and
producer's accuracies for Schinus are very similar to the Sobel edge density texture
classification results, the Sobel edge magnitude classification exhibits better cohesiveness
(Figure 4-13). That is, core areas of Schinus were generally well discriminated as
contiguous areas with most misidentification at the edge of the vegetation type. This is in
contrast to the Sobel edge density classification in which there was more
misclassification within the core areas of Schinus (Figure 4-12). Sabalpalmetto was most
likely to be confused with the Schinus class. Table 4-11 shows hardwoods with high
user's accuracy (97%), which is also similar to the results of the Sobel edge density
classification; however, the producer's accuracy is even lower (16% versus 29%). Water,
as is true with most of the textures, discriminates accurately. Other vegetation types in the
images were not discriminated (Tables 4-11 through 4-15).
Schinus had very similar user's accuracies when the Sobel edge slope texture is
used for classification (Tables 4-17, 4-18, and Table 4-19). Compared to the Sobel edge
magnitude classification, less of the core Schinus was detected (Figure 4-14) resulting in
10% to 20% lower producer's accuracies. Sabal and hardwoods were most likely to be
confused with the Schinus class. Hardwoods again have a 97% user's accuracy and a
considerably improved producer's accuracy of 54% (Table 4-16). The herbaceous class
also had a user's accuracy above 90% for image ww1099 (Table 4-18) but not in image
ww1093 (Table 4-17), and both images had a low producer's accuracy for herbaceous
ground cover.
The Sobel edge standard deviation texture classification (Figure 4-15)
discriminates Schinus with less accuracy than the other Sobel edge textures. User's
accuracies for images containing Schinus are 70%, 66%, and 89% (Tables 4-22, 4-23,
and 4-24) compared to 95%, 74%, and 94% for the same images classified with the Sobel
edge magnitude texture (Tables 4-12, 4-13, and 4-14). Producer's accuracies are about
the same. Again, Schinus was primarily confused with Sabalpalmetto. The Sobel edge
standard deviation texture classification accuracies for hardwoods (Table 4-21) and
herbaceous (Tables 4-22 and 4-23) are similar to the Sobel edge slope classification;
however, the hardwood class has a lower producer's accuracy.
Univariate Validation Image Classifications
Laplacian edge density classifications are not presented because Laplacian edges
were very uniformly distributed across the images, offering no discrimination of
vegetation classes. All of the classifications using decision rules from the calibration
images produced blank results. That is, all, or the majority, of the pixels in an image
classified to zero.
The Sobel edge density texture classification and the Sobel edge standard
deviation texture classification (Figure 4-16 and Figure 4-19; Table 4-26 through Table
4-29 and Table 4-38 through Table 4-41) had no success in identifying any of the
vegetation classes in the images. The Sobel edge density classification for image ww2091
had a 83% producer's accuracy for Sabalpalmetto; however, much of the correct
classification of Sabal results from a large commission error as seen in the 56% user's
accuracy (Table 4-29). With a 93% user's accuracy for image 0212-528 and 100% user's
accuracy in image ww2008, Sobel edge standard deviation appears to do well in the
identification of Schinus (Table 4-38 and Table 4-39); however, only 55 of 3594 pixels of
Schinus are identified in 0212-528 (1.5%) and the 100% user's accuracy in ww2008 is
from the identification of only 2 of the 1284 Schinus pixels in that image. The Sobel
magnitude texture and the Sobel edge slope texture classifications could only
successfully identify water (Figure 4-17 and Figure 4-18; Table 4-30 through Table 4-
37). The Sobel edge magnitude classification, like the Sobel edge standard deviation
classification, has high user's accuracy for Schinus in images 0212-528 and ww2008, but,
again, very few pixels of Schinus are actually identified and that high omission error is
reflected in the low producer's accuracies (Table 4-30 and Table 4-31).
Discussion
The edge texture classifiers showed promise for discrimination of vegetation
classes, particularly Schinus, when the signatures were observed for individual images.
Sobel edge textures are effective with all images but not always with the same threshold
(which can be very sensitive). Many of the Sobel edge textures gave similar, often good,
results, but there are specific differences in some textures getting better discrimination of
a specific vegetation class--thus the utility of a multivariate approach to be discussed in
Chapter 5.
However, when the signatures from the calibration images were combined to
generate an overall decision rule, the signature for any particular vegetation class
broadened to the extent that it no longer provided the separation needed to adequately
prevent confusion among the classes. Sobel edge density, Sobel edge magnitude, and
Sobel edge slope all had user's accuracies above 90% for Schinus in two of the three
calibration images containing Schinus (Table 4-7 through Table 4-9, Table 4-12 through
Table 4-14, and Table 4-17 through Table 4-19). Low producer's accuracies, however,
result in substantial omission.
Class confusion was particularly apparent when images not used in the calibration
were tested against the decision rules from the calibration images. Examination of the
Laplacian and Sobel edge images created from the validation images suggests that lower
edge thresholds may be required in many of these images to obtain satisfactory results.
If the signatures are developed in a supervised classification for individual
images, it is apparent that the edge textures should provide an excellent tool for
classification of the vegetation types tested. It was hoped, however, that the edge textures
would provide information relatively free from spectral variation among images and that
the edge textures among images would be reasonably constant. Those conditions should
75
allow unsupervised classification of vegetation classes based on the signatures of a few
calibration images. That was clearly not the case.
edges.
Threshold = 15%
Image ww2042, all edges.
0- i Aw~
Threshold = 25%
Threshold = 35%
Figure 4-2. Example Sobel edge images
ww1093, all edges.
Threshold = 50%
Threshold = 55%
Threshold = 60%
Figure 4-3. Example Laplacian edge images.
30000 -
25000-
20000 -
S15000 -
10000-
5000-
0-
b.
0 50 100 150 200 250
V1
-200 -100 0 100 200
V1
Figure 4-4. Sobel (a) and Laplacian (b) image histograms from edge images ofwwl085.
15000-
10000-
"
ww1085
Sabal -
Hardwoods
Iii
ill
I
I 1 L.
10 20 30 40 50
Laplacian Edge Density @ 55%
ww1099
Schinus
S Shrub
Herb ..............
i II
10 20 30 40 50
Laplacian Edge Density @ 55%
ww2046
Sabal
Herbaceous
" ll
1
10 20 30 40 50
Laplacian Edge Density @ 55%
ww1093
Schinus
Avicennia
Herb
Water -
10 20 30 40 50
Laplacian Edge Density @ 55%
ww2042
i j Schinus
i Myrica
ii i Typha ................
S Water ....-
iI!
10 20 30 40 50
Laplacian Edge Density @ 55%
Figure 4-5. Laplacian Edge Density texture signatures.
ww1085
Sabal
Hardwoods -
I I
0 20 40 60 80 100
Sobel Edge Density @ 10%
ww1099
0 20 40 60 80 100
Sobel Edge Density @ 10%
ww2046
Sabal
Herbaceous
0 20 40 60 80 100
Sobel Edge Density @ 10%
ww1093
(n
(n
w,
0
Schinus
Avicennia
Herb ..
Water .-----
iihr
0 20 40 60 80 100
Sobel Edge Density @ 10%
ww2042
Schinus
Myrica
Typha .......
Water -----
"- .. . .." \ ",.
0 20 40 60 80 100
Sobel Edge Density @ 10%
Figure 4-6. Sobel Edge Density texture signatures
Schinus
Shrub
Herb ----
c..
. /' I "
ww1085
Sabal -
Hardqoot
'I
I
20 40 60 80
Sobel Edge Magnitude @ DN 10
wwl 099
Schinus
rub
rb c o s ----------------
I-frbaceous
ww2046
ii
I';
20 40 60 80
Sobel Edge Magnitude @ DN 10
ww2046
Sabal
Herbaceous
^Y. .I \K\
ww1093
Schinus
Avice r ia
Herb eous ................
.t i, - - - -
20 40 60 80
Sobel Edge Magnitude @ DN 10
ww2042
Schinus
Myrica
Typha ................
Water ---.--
-."". -:
20 40 60 80
Sobel Edge Magnitude @ DN 10
20 40 60 80
Sobel Edge Magnitude @ DN 10
Figure 4-7. Sobel Edge Magnitude texture signatures.
ww1085
Sabal
Hardwoods
/ \
./ \ M
1000 1500 2000 2500 3000 3500
Sobel Edge Slope @ 10%
ww1099
Schinus
Shrub
Herbaceous -------
Herbaceous
A
I1
:I
1000 1500 2000 2500 3000 3500
Sobel Edge Slope @ 10%
ww2046
Sabal
Herbaceous
1000 2000 3000 4000 5000
Sobel Edge Slope @10%
ww1093
Schinus
Avicennia
Herbaceous .......
Water .--.---
i I
1000 1500 2000 2500 3000 3500
Sobel Edge Slope @ 10%
ww2042
Schinus
Myrica
Typha .......
Water --.---
ii
i \
V..
1000 1500 2000 2500 3000 3500
Sobel Edge Slope @ 10%
Figure 4-8. Sobel Edge Slope texture signatures.
ww1085
0
0
,
,0
0 .(
O
O
(D
(D
d
>,
o (M
0
O
ww1093
Sabal
Hardwoods ---
I., I
-LI
0 5 10 15 20 25 30
Sobel Edge Std. Dev.@ 10%
ww1099
Schinus
Shrub
Herbaceous -----
ii ,
... tj
0 5 10 15 20 25 30
Sobel Edge Std. Dev.@ 10%
ww2046
Sabal
Herbaceous
0 10 20 30 40 50 60
Sobel Edge Std. Dev.@ 10%
(D
0
>,
()
(0
O
(,
o (M
Figure 4-9. Sobel Edge Standard Deviation texture signatures.
Schinus
Avicennia -- n
Herbaceous a......
Water -
II ,
Li
0 5 10 15 20 25 30
Sobel Edge Std. Dev.@ 10%
ww2042
Schinus
Myrica
Typha .......
Water --.---
tl
i i
0 5 10 15 20 25 30
Sobel Edge Std. Dev.@ 10%
ww1085
(q
d
,0
O
O
O
ww1093
(D
O6
o C
O
O
0
Sabal
Hardwoods
10 15 20 25 30 35 40
Standard Deviation
ww1099
Schinus
Shrub
Herb ---------
I,
JI Ii
10 15 20 25 30 35 40
Standard Deviation
ww2046
Sabal
Herbaceous
i
10 20 30 40 50 60 70
Standard Deviation
1 Schinus
Avicennia
I Water .-----
hi f
10 15 20 25 30 35 40
Standard Deviation
ww2042
Schinus
Myrica
Typha .......
Water -------
10 15 20 25 30 35 40
Standard Deviation
Figure 4-10. Image Standard Deviation texture signatures.
(0
o
0
0
0,
d
d
(D
0
a
0
(q
d
,0
O
O
d
|
