Citation |

- Permanent Link:
- https://ufdc.ufl.edu/UFE0024362/00001
## Material Information- Title:
- Parameter Estimation Framework for Fusing Spatio-Temporal Data in Watershed Analysis and its Potential for RC-based Hardware Acceleration
- Creator:
- Nagarajan, Karthik
- Place of Publication:
- [Gainesville, Fla.]
- Publisher:
- University of Florida
- Publication Date:
- 2009
- Language:
- english
- Physical Description:
- 1 online resource (127 p.)
## Thesis/Dissertation Information- Degree:
- Doctorate ( Ph.D.)
- Degree Grantor:
- University of Florida
- Degree Disciplines:
- Electrical and Computer Engineering
- Committee Chair:
- Slatton, Kenneth C.
- Committee Co-Chair:
- George, Alan D.
- Committee Members:
- Principe, Jose C.
Graham, Wendy D. - Graduation Date:
- 8/8/2009
## Subjects- Subjects / Keywords:
- Architectural design ( jstor )
Correlations ( jstor ) Datasets ( jstor ) Design analysis ( jstor ) Hydrological modeling ( jstor ) Pipelines ( jstor ) Rain ( jstor ) Statistical estimation ( jstor ) Stream flow ( jstor ) Watersheds ( jstor ) Electrical and Computer Engineering -- Dissertations, Academic -- UF acceleration, designpatterns, fpgas, fusion, hydrology, spatiotemporal Suwannee River, FL ( local ) - Genre:
- Electronic Thesis or Dissertation
born-digital ( sobekcm ) Electrical and Computer Engineering thesis, Ph.D.
## Notes- Abstract:
- PARAMETER ESTIMATION FRAMEWORK FOR FUSING SPATIO-TEMPORAL DATA IN WATERSHED ANALYSIS AND ITS POTENTIAL FOR RC-BASED HARDWARE ACCELERATION Knowledge of the flow rates in rivers and major streams is critical for understanding and effectively managing watersheds because it bears on such things as predicting flood vulnerability and fluvial erosion of the landscape, as well as mitigating the damaging effects of floods and waterborne pollution. Spatially dense networks of in situ streamflow measurements are generally nonexistent because of the high cost of installing, calibrating, and maintaining the necessary equipment. A framework is therefore needed to predict streamflow from other available data. The relevant data are often of different modalities, however, such as images (e.g., digital topographic maps), time-series (e.g., an in situ streamflow guage), and spatio-temporal (e.g., daily radar-based precipitation estimates), and the estimation of parameters that exhibit complex spatio-temporal dependencies from such multi-modal measurements poses a significant challenge. Most data fusion algorithms explored to date work with data belonging to a single mode or ?family? possessing one-to-one sample mappings, such as images or time series data, and cannot be directly applied here. In this work, a simple yet efficient probabilistic framework based on Bayesian networks (BNs) is developed that provides a mechanism to incorporate spatial, temporal and spatio-temporal features. Graph topologies are generated based on physical models characterizing the underlying random process. Solving high-dimensional estimation problems over large data sets leads to computational demands that are often impractical to run even on high-end General Purpose Processors. This is true for the framework proposed in this work as well. FPGA-based reconfigurable computing has been successfully used to accelerate computationally intensive problems in a wide variety of scientific domains to achieve speedup over traditional software implementations. However, this potential benefit is quite often not fully realized because creating efficient FPGA designs is generally carried out in a laborious, case-specific manner requiring a great amount of redundant time and effort. Algorithm decomposition, performance prediction, and design reuse for parallel implementation all need to be performed in an efficient and structured manner in order to develop successful FPGA designs in a productive way. To address these challenges, an approach for pattern-based decomposition of algorithms for FPGA design and development is proposed. The major contributions of this work include (1) development of a scalable approach to fuse spatial, temporal and spatio-temporal watershed data via a BN to estimate streamflows; (2) prediction of uncertainties in streamflow estimates as a function of location (space), condition (time) and the particular mix of sensors; and (3) accelerating algorithm execution on FPGAs using pattern-based decomposition with significant increases in productivity via design reusability. ( en )
- General Note:
- In the series University of Florida Digital Collections.
- General Note:
- Includes vita.
- Bibliography:
- Includes bibliographical references.
- Source of Description:
- Description based on online resource; title from PDF title page.
- Source of Description:
- This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
- Thesis:
- Thesis (Ph.D.)--University of Florida, 2009.
- Local:
- Adviser: Slatton, Kenneth C.
- Local:
- Co-adviser: George, Alan D.
- Electronic Access:
- RESTRICTED TO UF STUDENTS, STAFF, FACULTY, AND ON-CAMPUS USE UNTIL 2010-02-28
- Statement of Responsibility:
- by Karthik Nagarajan.
## Record Information- Source Institution:
- University of Florida
- Holding Location:
- University of Florida
- Rights Management:
- Copyright Nagarajan, Karthik. Permission granted to the University of Florida to digitize, archive and distribute this item for non-profit research and educational purposes. Any reuse of this item in excess of fair use or other copyright exemptions requires permission of the copyright holder.
- Embargo Date:
- 2/28/2010
- Resource Identifier:
- 489094574 ( OCLC )
- Classification:
- LD1780 2009 ( lcc )
## UFDC Membership |

Downloads |

## This item has the following downloads: |

Full Text |

PAGE 1 1 PARAMETER ESTIMATION FRAMEWORK FOR FUSING SPATIO TEMPORAL DATA IN WATERSHED ANALYSIS AND ITS POTENTIAL FOR RC BASED HARDWARE ACCELERATION By KARTHIK NAGARAJAN A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSI TY OF FLORIDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 2009 PAGE 2 2 2009 Karthik Nagarajan PAGE 3 3 To my mom and dad PAGE 4 4 ACKNOWLEDGMENTS This work would n ot have been possible if not for the continued support, belief and guidance of many people. First, I want to thank my parents Smt. Vanitha Nagarajan and Shri. G. Nagarajan for their support, prayers, and the wonderful learning platform they have provided since early childhood. I am also thankful to my brother and sister in law for having provided moral support and the occasional fun that is ever so required in a graduate students life. I also must thank my wonderful wife Vidhya for giving her love and s upport. She has been patient, understanding, and sacrificed her standard of living and more importantly her time with me while I remained in school. I wish to thank my advisors, Prof. K. Clint Slatton and Prof. Alan D. George for their continued guidanc e and encouraging words. Without their input and financial support, this work would not have been made possible. Though tough at times, they have gradually molded me into a better researcher I also wish to thank Dr. Wendy Graham for helping me understa nd certain hydrology related problems and also for the time and effort she invested in refining the research contributions. I am also thankful to my friends at the Adaptive Signal Processing Laboratory and the Center for High performance Reconfigurable Co mputing for their inputs and support. This work was partially supported by the National Science Foundation (Grant No. EAR 0609968) and I/UCRC Program of the National Science Foundation (Grant No. EEC 0642422). PAGE 5 5 TABLE OF CONTENTS page ACKNOWLEDGMENTS ...............................................................................................................4 LIST OF TABLES ...........................................................................................................................7 LIST OF FIGURES .........................................................................................................................8 ABSTRACT ...................................................................................................................................10 CHAPTER 1 INTRODUCTION ..................................................................................................................12 11 Unified Data Fusion Framework ......................................................................................12 12 Rapid Computation on Hardware Platforms ....................................................................14 13 Contributions and Outline ................................................................................................17 2 BACKGROUND ....................................................................................................................24 21 Data Fusion Methodologies in Hydrology .......................................................................24 22 Study Site and Data Description ......................................................................................27 221 Observed Data ........................................................................................................28 222 Derived Data ..........................................................................................................29 23 Hardware Acceleration on FPGAs ...................................................................................31 24 Case study Algorithms .....................................................................................................34 241 Probability Density Function Estimation ...............................................................34 242 K Means Clustering ...............................................................................................37 243 Correlation .............................................................................................................38 244 Algorithm Similarity ..............................................................................................38 3 SPATIO TEMPORAL BAYESIAN NETWORK .................................................................42 31 Bayesian Network Fundamentals .....................................................................................44 311 Network Learning ..................................................................................................46 312 Network Topology .................................................................................................47 313 Network Inference .................................................................................................48 314 Performance Metrics ..............................................................................................49 32 Measurement of Incremental Information ........................................................................50 321 Feature Ranking .....................................................................................................50 322 Valid ating Information Gain Analysis ..................................................................52 33 Model Predictions at Training Stations ............................................................................52 34 Model Predictions at Intermediate Stations ......................................................................55 PAGE 6 6 4 HARDWARE ACCELERATION ON FPGAS .....................................................................68 41 Pattern based Algorithm Decomposition .........................................................................68 411 Pattern Description ................................................................................................69 411 1 Pipeline pattern ............................................................................................69 411 2 Datapath replication pattern ........................................................................70 411 3 Loop fusion pattern......................................................................................71 412 Quantifying Design Patterns ..................................................................................72 42 Sca lable and Portable Architecture for PDF Estimation ..................................................73 421 Performance Prediction ..........................................................................................74 422 1 D PDF Design .....................................................................................................77 422 1 Single core design .......................................................................................78 422 2 Dual core design ..........................................................................................79 423 2 D PDF Design .....................................................................................................79 424 Experimental Results .............................................................................................80 424 1 Speedup, power, and resource utilization ....................................................81 424 2 Data verification ..........................................................................................82 424 3 Scaling to multiple cores .............................................................................82 424 4 Portability considerations ............................................................................83 424 5 Scaling to multiFPGA systems ..................................................................84 43 Composite Patterns for Design Reuse ..............................................................................85 431 K M eans Clustering ...............................................................................................85 432 Correlation .............................................................................................................88 5 MULTISCALE ESTIMATION FRAMEWORK ...................................................................98 51 Test Area ........................................................................................................................100 52 Multiscale Estimation Framework .................................................................................100 53 Mixture of Experts (MOE) .............................................................................................103 54 Results and Analyses ......................................................................................................104 6 CONCLUSIONS AND FUTURE WORK ...........................................................................112 61 Message Passing .............................................................................................................113 61 Streamflow Estimation via Message Passing .................................................................114 LIST OF REFERENCES .............................................................................................................120 BIOGRAPHICAL SKETCH .......................................................................................................127 PAGE 7 7 LIST OF TABLES Table page 11 Elapsed time in estimating high dimensional PDFs using Parzen w indows. ....................20 21 Set of available features .....................................................................................................39 31 Comparison of computational complexity between traditional Bayesian method, RDBE and STBN ...............................................................................................................57 32 Feature ranking based on conditional entropy (CH) ..........................................................57 33 Performance metrics for cases 1 4 validating informati on gain analysis ..........................57 34 Performance metrics for networks N1 (Qnet) and N2 (WAM) ...........................................58 35 Performance metrics for N2 under 80% Trai ning Data .....................................................58 36 Performance metrics for N2 under 50% Training Data .....................................................58 41 List of primitive patterns and their description ..................................................................89 42 Parameters of primitive patterns ........................................................................................89 43 RAT input parameters for analysis of 1 D and 2 D PDF estimation algorithms ...............90 44 Speedup, power, and utilization for single core designs ...................................................90 45 Performance factors for single core designs ......................................................................90 46 Speedup and utilization for dual core designs ...................................................................91 47 Speedup and utilization for single core designs across platforms .....................................91 48 Speedup for single core 2 D PDF design across multiple FPGAs ....................................91 49 Speedup and utilization for PDF estimation, K means clustering, and Correlation (single core designs) ..........................................................................................................91 51 MSE obtained over the simulated image .........................................................................107 52 Incremental variance across scales and empirical F ( s ) values .........................................107 53 Average model weights obtained using MM MKS for real data.....................................108 61 Performance metrics at 2321500 after message passing ..................................................118 PAGE 8 8 LIST OF FIGURES Figure page 11 Fusing data measured across domains ...............................................................................21 12 Research concept diagram .................................................................................................22 13 Research roadmap. .............................................................................................................23 14 Availability of measurements at multiple scales ...............................................................23 21 A map of the Santa Fe River Watershed with its streams and rivers .................................40 22 Illustration of case study algorithms ..................................................................................41 23 Dataflow and computational structure in PDF estimation, K means clustering, and Correlation .........................................................................................................................41 31 Basic node structures in BNs .............................................................................................58 32 Sample network representation of a BN for embedding spatiotemporal data ..................59 33 Bayesian networks for streamflow estimation given spatio temporal data sources ..........59 34 Streamflow estimates at station 2321000 ..........................................................................60 35 Streamflow estimates at station 2321500 ..........................................................................61 36 Streamflow estimates at station 2322500 ..........................................................................62 37 Model validation at 2321000 .............................................................................................63 38 Model validation at 2321500 .............................................................................................64 39 Model validation at 2322500 .............................................................................................65 310 Streamflow estimation at intermediate points ....................................................................66 311 Streamflow and MAP estimates along the Santa Fe river on October 4th, 2006. ..............66 312 Observed versus predicted streamflow at USG S station 2321975 ....................................66 313 Model validation at 2321975 .............................................................................................67 41 Pattern based analysis of 1D PDF estimation algorithm ..................................................92 42 Development stages in system design ................................................................................92 PAGE 9 9 43 State transition diagram for a single core 1 D PDF design ...............................................93 44 Design details of 1 D PDF .................................................................................................93 45 Host centric execution flows .............................................................................................94 46 Architecture design for 2 D PDF estimation algorithm .....................................................94 47 State transition diagram for a single core 2 D PDF design ...............................................94 48 1D PDF estimates on GP P and FPGA ..............................................................................95 49 2D PDF estimates on GPP and FPGA ..............................................................................95 410 Multi FPGA system configuration ....................................................................................95 411 Architecture for K means clustering ..................................................................................96 412 Decomposition and design of correlation ..........................................................................97 51 F orest hydrologic cycle processes ....................................................................................108 52 Elevation image created by gridding last stop elevation points using an inverse distance weighting of study area ......................................................................................109 53 Mixture of Gaussians fitted over the histogram plot of fractal dimension ......................109 54 Fusion of Kalman filter bank estimates using Mixture of experts network ....................109 55 Fine scale simulated image with two terrain roughness ..................................................110 56 Weights obtained for smooth and rough filters at scales m=7 and 6 respectively ...........110 57 Filter weights under large and small P0 ...........................................................................111 58 Weights assigned to Model 1, 2, and 3 at various scales .................................................111 61 Illustration of message passing ........................................................................................118 62 Message passing framework for improving streamflow estimates ..................................118 63 Prediction at USGS 2321500 after message passing .......................................................119 64 Composite pattern for tree based algorithms ...................................................................119 65 Composite pattern for mesh based algorithms ................................................................119 PAGE 10 10 Abstract of Dissertation Presented to the Graduate School of the University of Florida in Partial Fulfillment of the Require ments for the Degree of Doctor of Philosophy PARAMETER ESTIMATION FRAMEWORK FOR FUSING SPATIO TEMPORAL DATA IN WATERSHED ANALYSIS AND ITS POTENTIAL FOR RC BASED HARDWARE ACCELERATION By Karthik Nagarajan August 2009 Chair: K. Clint Slatton Cochair: Alan D. George Major: Electrical and Computer Engineering K nowledge of the flow rates in rivers and major streams is critical for understanding and effective ly managing watersheds because it bears on such things as predicting flood vulnerability and fluvial eros ion of the landscape, as well as mitigating the damaging effects of floods and waterborne pollution S patially dense networks of in situ stream flow measurements are generally nonexistent because of the high cost of installing, calibrating, and maintaining the necessary equipment. A framework is therefore needed to predict streamflow from other available data. The relevant data are often of different modalities, however, such as images (e.g. digital topographic maps), time series (e.g. an in situ streamflow guage), and spatiotemporal (e.g. daily radar based precipitation estimates), and the e stimation of parameters t hat exhibit complex spatio temporal dependencies from such multi m odal measurements poses a significant challenge. Most data fusion algor ithms explored to date work with data belonging to a single mode or family possessing one to one sampl e mappings, such as images or time ser ies data and cannot be directly applied here. In this work, a simple yet efficient probabilistic framework based on Bayesian networks (BNs) is developed that provides a mechanism to incorporate PAGE 11 11 spatial, temporal and spatiotemporal features G raph topologies are generated based on physical models characterizing the underlying random process. Solving high dimension al estimation problems over large data sets lead s to computational demands that are often impractical to run even on high end General Purpose Processors. This is true for the framework proposed in this work as well. FPGA based reconfigurable computing ha s been successfully used to accelerate computationally intensive problems in a wide variety of scientific domains to achieve speedup over traditional software implementations. However, this potential benefit is quite often not fully realized because creat ing efficient FPGA designs is generally carried out in a laborious, case specific manner requiring a great amount of redundant time and effort. Algorithm decomposition, performance prediction, and design reuse for parallel implementation all need to be pe rformed in an efficient and structured manner in order to develop successful FPGA designs in a productive way To address these challenges, an approach for pattern based decomposition of algorithms for FPGA design and development is proposed. The major c ontributions of this work include (1) development of a scalable approach to fuse spatial, temporal and spatiotemporal watershed data via a BN to estimate streamflow s; (2) prediction of uncertainties in streamflow estimates as a function of location (space ), condition (time) and the particular mix of sensors ; and (3) a ccelerating algorithm execution on FPGAs using pattern based decomposition with significant increases in productivity via design reusability. PAGE 12 12 CHAPTER 1 I NTRODUCTION 11 Unified Data Fusion Framework Traditional data fusion is the process of putting together information in the form of measurements ( or features) collected from a heterogeneous set of sensors into a single composite picture of the environment. The composite picture, for examp le, might deal with parameter estimation or a classification problem wherein the results depend on the set of features. Such data fusion problems arise in a variety of contexts, most notably in remote sensing and geophysical applications. For example, f l ow of water through stream networks and rivers directly impacts flooding and transport of s ediments and pollutants (e.g., nutrients) in watershed systems. Hence knowledge of streamflow rates is critical for management of water quality, in stream flows and mitigation of flooding and drought events Unfortunately, spatially dense networks of in situ streamflow meters are not generally available and would be prohibitively expensive to deploy and maintain. Thus, an estimation framework is needed that utiliz es the available data to predict streamflow through a river network S patial data, such as surface topography, landcover, and soil type; temporal data, such as time series of measured stream flow and groundwater levels ; and spatiotemporal data, such as ra infall, all bear on streamflow Some of these quantities can be obtained from remote sensing imagery ; however, utilizing such disparate data types using traditional data fusion methods is problematic. In general, e stimation of parameters that exhibit complex spatial and temporal dependencies on multi modal measurements (features) poses a significant challenge. This is because most data fusion algorithms explored to date work with data samples belonging to a single mode or family of data, such as ima ge data at a particular point in time [ 1] or time series PAGE 13 13 data at a particular point in space. Certain algorithms do offer the possibility to fuse measurements of different types (e.g., spatial data with point measurements) [ 2 ]. In doing so, they still on ly capture the variations of the parameter being estimated spatially or temporally. In other words, most methods developed for fusing image data do not integrate point source and distributed time series data efficiently. Application specific deterministic models ( e.g. p hysically based hydrologic models ) have been used to predict streamflow in watersheds but often with significant uncertainty because numerous assumptions often have to be made for many unmeasured input and parameter values Purely statis tical methods (e.g., regression analysis) have been used extensively as an alternative to physically based models in an effort to account for unmeasured input parameters and/or parameter uncertainty. However, such approac hes often deal only with second or der moments (variance) and also impose calibration requirements for their extrapolation to new locations or conditions Probabilistic methods (e.g. traditional Bayesian inference) that ut ilize the joint (multivariate) probability density function (PDF) of all the features and parameters rather than merely a finite subset of its statistical moments can be used to incorporate (fuse) different data sets i n order to estimate streamflow. This approach however, suffers from a super linear increase in computa tional complexity as the number of data sets to be fused grows. This demonstrates the need for a unified estimation framework that provides the computational platform to fuse spatio temporal, spa tial and/or temporal data ( Figure 11), in conjunction with limited training data/measurements. When measurements for some desired input parameters are unavailable the framework should accommodate hidden or proxy variables through the use of physically based relationships. In this way, the information containe d in PAGE 14 14 our physical understanding of the system can leverage the information in the training data, while uncertainty and variability in the parameters can be accommodated to relate parameters whose precise deterministic relationships are not well known. 12 Rapid Computation on Hardware Platforms Machine learning is a general term that has been used to refer to a large and diverse set of methods for extracting patterns from data. In general, it encompasses portions of statistical and adaptive signal proc essing, probabilistic decision theory such as Bayes Rule, and meta heuristic strategies such as the genetic algorithm. M achine learning algorithms have become widely used for problems such as detection, estimation, and classification involving diverse da ta types, including time series, image, and video (spatiotemporal) data [ 3 ]. Raw image and video data do not explicitly contain high level information and thus generally require an abstraction process to extract useful information from the data. This pr ocess is often accomplished via a sequence of data segmentation algorithms [ 3, 4]. Clustering algorithms are the most common type of algorithms used for segmenting data points into groups and employ various measures of similarity. Probabilistic algorithm s can then be employed to represent the segmented data as PDFs rather than single data points (e.g., cluster centers). Such data representations are essential in many applications involving decision making based on probabilistic reasoning [ 5, 6, 7]. Alte rnatively, as pointed out in Section 1 1, certain estimation algorithms use statistical methods to model relationships between datasets as a function of the separation between them (dissimilarity). These methods are useful for applications such as data co ntent retrieval and indexing [ 8 ], data segmentation, and speech recognition systems [ 9, 10 ] that benefit from quantitative measures of the statistical relationships between features. The streamflow e stimation framework proposed in this work is driven by machine learning algorithms primarily based on a data driven approach wherein the computational burden is PAGE 15 15 heavily impacted by the volume of data being processed. Three such principal algorithms are the Parzen window based non parametric estimation of PDFs K means clustering and correlation While applying the estimation fr amework over small watersheds ( e.g., the 3585 km2 Santa F e River w atershed ) would result in moderate execution times, extending the framework to larger watersheds (e.g., the 28542 km2 S uwannee River basin) would lead to unre a sonable computational times (Figure 12) For example, Table 1 1 illustrates the computational burden incurred in terms of time elapsed (where optimized C code is executed on a 3.2 GHz single core Xeon processor) in computing PDFs of increasing dimension (number of feature s considered) using the Parzen window technique with Gaussian kernels. The total number of data samples was set to 204800 ( N ) and the support size of the PDF on each dimension was set to 256 ( n) T he high computational burden arises because most problems involve large volumes of data and require complex computations to be performed in a high dimensional space. To mitigate this problem, application researchers have investigated ways to approximate t he solution by either solving the problem under a reduced dimensional space or providing alternative methods like the Fast Gauss Transform (FGT) [1 1]. The FGT addresses the dimensionality issue to a certain extent by reducing the number of computations in each dimension. Irrespective of the approximation method used, the computation involved still grows exponentially with increasing dimensionality. Since the three algorithms mentioned ( Parzen estimation of PDFs K means clustering and correlation ) form an integral part of the estimation framework (and other numerous machine learning applications as well ) their fast and efficient execution is extremely desirable. Further, according to Amdahls law [12], significant improvements to the overall execution time of an application are obtained only when significant portions of it are improved. In other words, PAGE 16 16 overall efficiency gains are most likely to occur when the principal (or most often used) portions of the machine learning tasks are made more efficient. FPGAbased reconfigurable computing (RC) has been successfully used to accelerate computationally intensive problems in a wide variety of scientific domains to achieve speedup over traditional software implementations. D eveloping FPGA designs howev er, is often time consuming and not every algorithm is amenable to hardware acceleration. Preliminary analyses should be performed to predict the algorithms amenability to a hardware paradigm before undertaking a lengthy development process. In contrast to the general computing domain, there is a relatively modest variety of FPGA based hardware platforms from which a designer can choose. These platforms primarily vary in the type of FPGA they house and the manner in which data is communicated between the FPGA and host processor; variations on each platform are also available based on the constraints on the particular application design and performance. For example, decisions regarding bit precision, design architecture, memory utilization, and storage de pend on the FPGAs resource availability and the size of the hardware design. The presence of dedicated arithmetic blocks also dictates the implementation options for certain mathematical instructions. Thus, it is best to determine the expected performan ce attainable at an early stage when migrating an algorithm to an FPGA platform by employing simple performance prediction tools. Another barrier to successfully using FPGA based RC technology for algorithm acceleration is the prohibitive time and effort that is required to develop the necessary FPGA core designs in a custom, case specific manner. A more desirable approach would be to recognize and exploit common patterns in terms of computational structure and data flow across a variety of algorithms an d prove that those design patterns are indeed suitable for use in other FPGA implementations. The primary goal is hence to develop a concise set of design patterns PAGE 17 17 that can be formally used to represent the decomposition of an algorithm for parallel implementation. If the hardware design of an algorithm can be represented at a high level graphically via composite patterns (mixture of computation and communication patterns), then other algorithms having a similar high level representation can reuse the exis ting design with only minimal customization. Such a high level description is also desirable to aid inexperienced FPGA designers (e.g., domain scientists) so they can easily comprehend and reuse existing hardware designs while developing new applications. Pattern based descriptions of designs could be used to effectively disseminate and share hardware cores as well. Such an approach can lead to a significant increase in productivity via design reusability. Several important machine learning algorithms, such as PDF estimation using Parzen windows, K means clustering, correlation, and information theoretic measures, have very similar underlying algorithmic and dataflow structures and thus could potentially benefit from such an approach. In fact, the FGT a lgorithm can be decomposed for a parallel implementation in a similar fashion to that of the Parzen window technique and could potentially benefit from the architecture designed for the latter. In general, the similarity in dataflow in many machine learni ng algorithms is due to the fact that application data (input to the learning system) is compared with several representative values (parameters of the learning system) to perform inference. In addition to being structurally similar, most of these algorit hms are often applied to very large volumes of data and impose difficult constraints on memory. 13 Contributions and Outline In this dissertation, a unified and efficient framework to fuse spatiotemporal, spatial and/or temporal data for streamflow pr ediction is presented and key algorithms in the framework are accelerated on FPGAs in a productive manner via design reuse using patternbased algorithm decomposition. The topic is timely because the hydrology community is PAGE 18 18 currently seeking principled approaches to determine how best to instrument large watersheds in the United States with limited resources. The National Science Foundation (NSF) has termed such instrumented watersheds hydrologic observatories as an analogy to large astronomical observat ories in which NSF invests funds so that many different investigators can use the same resource. While the exact schedule for the creation of these hydrologic observatories remains uncertain, NSF has made a strong commitment to this vision by funding seve ral pilot projects [13] The work herein is designed to advance our knowledge of how such an observatory might be characterized algorithmically. The major contributions of this work include (1) development of a scalable approach to fuse spatial, tempora l and spatiotemporal data via a Bayesian Network (BN) to estimate streamflow in a watershed system; (2) prediction of uncertainties in streamflow estimates as a function of location (space), condition (time) and the particular mix of sensors ; and (3) a cce lerating algorithm execution on FPGAs using pattern based decomposition with significant increases in productivity via design reusability. Chapter 1 gave a basic overview of the problem of fusing data collected from disparate sources across various domains, some of the existing methodologies, their characteristics and scope for improvements. The issue of computational burden faced while handling large datasets and the choice of FPGAs as potential solutions were pointed out. The major contributions in the work and the application cas e studies that would be used for purposes of illustrating the proposed concepts were also presented (Figure 1 3) Chapter 2 discusses background on existing methodologies for solving data fusion problems in hydrology and the am enability of signal processing algorithms for implementat ion on FPGAs. D iscussions on relevant performance prediction methods and design patterns for RC PAGE 19 19 used in this work are briefly presented. C ase study algorithms (fr om the estimation framework) chosen for hardware acceleration are detailed as well. A brief description of the study site ( the Santa Fe River w atershed system) and the features used to predict streamflow are also provided. Chapter 3 describes the Spatio Temporal Bayesian Network ( STBN ) method that offers a unified platform to incorporate measurements observed across various domains. The network is then applied to t he problem of estimating streamflow in a watershed system taking into account the complex non linear dependencies of flow with a variety of other measurements made across space and time. When sufficient training data are available for the full dynamic range of the random variables in each dimension, o ptimal graph topologies for BNs can be obtained using optimization methods, such as simulated annealing [14] However, in many geophysical applications, such as ours, coincident measurements are only available for relatively limited time periods or spatial locations, thus we must exploit our physical knowledge of the hydrologic processes to determine reasonable BN topologies A n information theoretic methodology based on conditional entropy is then employed to help quantify the impact of adding nodes in the graphical model in terms of information gained While this approach doe s not strictly gu a rantee that the optimal BN topology is found, it does allow us to measure the benefit of one topology over another The framework also offers the flexibility of embedding knowledge from hydrologic models calibrated for the study area by introducing them as additional nodes in the network, thereby improving prediction accuracy. Posterior probabilities of streamflow estimates and the associated entropy values provide valuable information on the quality of predictions and offer directions f or future watershed instrumentation. Chapter 4 describes the pattern based decomposition, analysis, and implementation of a scalable and portable architecture for multi dimensional, non parametric PDF estimation using PAGE 20 20 Gaussian kernels on FPGAs. T he algori thms amenability to a hardware paradigm and expected speedups are predicted before implementation Multi core architectures are developed to further improve performance by scaling the design. Portability of the hardware design across multiple FPGA platf orms is also analyzed. After implementing the PDF algorithm, the value of pattern based decomposition to support reuse is demonstrated by rapid development of the K means and correlation algorithms. Chapter 5 presents a multiscale estimation framework that would prove useful in data fusion problems where measurements (image data in particular) are available at a wide range of scales. In such cases, there is a need to convert all the features to a common spatial resolution (through a downsample/upsampl e operation). Most often, this is a scale at which the spatio temporal features are measured ( Figure 1 4) Although this work does not extend the streamflow estimation framework to incorporate multi scale data, a p roof of concept study for multi scale anal ysis is provided and the actual integration is left as future work. Chapter 6 summarizes the work through conclusions and lists ideas for future work. Table 1 1. Elapsed time in estimating high dimensional PDF s using Parzen windows Dimension Time elapsed Increase factor 1 0.58 seconds 1 2 158.75 seconds 275 3 ~12.00 hours (estimated) 75000 4 ~139.00 hours (estimated) ~ 2 10 7 Estimated indicates the expected time elapsed based on knowledge of algorithm complexity The total numb er of data samples was set to 204800 ( N ) and the support size of the PDF on each dimension was set to 256 ( n ) PAGE 21 21 Figure 11. Fusing data measured across domains PAGE 22 22 Figure 12. Research concept diagram. Note: The key algorithms used in the estimation framew ork are PDF estimation, K means clustering and correlation (color coded as blue, gray, and red). While applying the framework to the SFW (upper left) results in manageable computational times, extending the framework to the Suwannee Basin (lower right) le ads to unreasonable times due to high data volume. Accelerating algorithm execution via parallel processing is a viable solution to reduce overall computational time. PAGE 23 23 Figure 13. Research roadmap Figure 14. Availability of measurements at multipl e scales PAGE 24 24 CHAPTER 2 BACKGROUND 21 Data Fusion Methodologies in Hydrology F low rates in rivers and major streams are governed by a complex set of physical parameters The subset of those parameters that are measured often comprise a wide range of scales and uncertainties making fusion of the data more difficult. In addition, some data are strictly temporal (e.g. time series of measured flow at a particular location), others are spatial (e.g. digital maps of soil type and topography), while still othe rs are spatio temporal (e.g. estimates of distributed daily rainfall from weather radars). To date, most of the work on streamflow estimation has been handled by using deterministic models of hydrologic processes that attempt to parameterize the function al relationships between quantities, such as rainfall, surface topography, antecedent soil moisture, and stream discharge [ 15, 16 ]. These physically based approaches, though appealing, are inherently limited because it is typically impossible to specify a ll the necessary boundary conditions at all times and locations with reasonable accuracy. Thus, such approaches often necessitate unreasonable assumptions of uniformity on various hydrologic parameters over the watershed. Statistical methods, such as regr ession, have been extensively used in conjunction with physically based models in an effort to account for parameter uncertainty. However, such approac hes often deal only with second order moments (variance) and also impose calibration requirements for th eir extrapolation to new locations or conditions. In [ 17, 18], the authors identify and propose strategies for improved site calibration using statistical methods. In [ 19], the authors support the claim that many existing physical models assisted by stat istical approaches to parameter estimation can potentially be too site specific or globally common. Multi resolution data fusion techniqu es have been applied to assimilate information from PAGE 25 25 multiple images [ 20]. Many investigators have developed specializ ed methods for fusing low and high resolution multispectral and hyperspectral images of the terrain [ 21, 22 ]. M ost methods developed for fusing image data do not integrate po int source and distributed time series data efficiently, y et many hydrologic modeling problems require both image and time series data. The ensemble Kalman Filter, first applied to ocean and atmospheric modeling in the mid 1990s, has been more recently applied to problems in hydrology and soil science [ 23, 24] to accommodate such mix ed mode data. These methods, however, can be suboptimal when the data and underlying processes are not Gaussian distributed, which is often the case. Attempts have been made to fuse data from multiple sources using Bayesian approaches to accommodate non Gaussian distributions [ 25, 26, 27 ]. A hierarchical approach based on Bayesian inference was proposed in [ 19] to estimate parameters of a physical model developed for predicting oxygen demand in estuaries. While the work assumed parametric models for fi tting the underlying data, the target variable was estimated based on non parametric formulations. In theory, non parametric Bayesian inference yields optimal estimates (minimum probability of error) because it utilizes the joint (multivariate) PDF rather than merely a finite subset of its statistical moments. However, estimating high dimensional PDF s is computationally very difficult due to the well known phenomenon known as the curse of dimensionality [ 28]. Information theory has been employed in man y areas to effectively handle large datasets. Authors in [29] address the curse of dimensionality issue in probabilistic methods by selecting a subset of features based on an information theoretic approach. However, t his eliminates information that mig ht help to characterize the parameter being estimated. Another drawback to the method in [2 9 ] is that, in its current formulation, the estimation problem is solved either strictly in the temporal domain or spatial domain. PAGE 26 26 Formulating Bayesian inference problems on graphical models such as Bayesian networks ( B N s), is known to ameliorate the computational complexity issue. Existing B N approaches applied to hydrology problems are based on simple datasets [30] applied at different spatial locations [3 1] an d do not traditionally capture the impact of spatial nonstationarities on response variables. Further, they have been primarily limited to addressing qualitative issues like landscape quality (for pleasure and leisure) and ecological value (species and bi odiversity) [3 2 ], compliance violations at water treatment plants [3 3], and impact of pollutant transfer [3 4]. The authors in [ 27] present a Bayesian spatio temporal dynamic model for determining ocean surface winds by fusing multi resolution datasets in a hierarchical manner. Incorporating a complex spatio temporal covariance structure in the model captures the spatial nonstationarity in ocean wind fields. However, as indicated earlier, estimation of hydrologic parameters that exhibit complex (non line ar) spatial and temporal dependencies on multi modal measurements poses a significant challenge. Unlike most data fusion applications explored to date where the data samples all come from a single family of data, the measurements from which the conditio nal PDF s (likelihoods) must be learned include spatial, temporal, and spatiotemporal measurements that are often sparse. Thus new strategies for designing the graph structure that incorporate knowledge of the physical processes are required. In this wo rk a non parametric BN methodology is proposed that incorporates both remotely sensed spatial and in situ temporal data (features) extracted from a watershed to predict streamflow with quantified uncertainty. BNs are appropriate to model the intuitive un derstanding of physical hydrologic processes as simple causal relations quantified by probability densities They also provide capabilities to incorporate physical model (i.e., models PAGE 27 27 calibrated for the watershed under study) knowledge by adding them as n odes in the network topology. This can potentially lead to improved performance. Even though the method maintains manageable computational complexity while still allowing many features to be considered, it is desirable to quantify the usefulness of featu res towards estimation. In this work, i nformation theoretic concepts are employed to rank features (added as nodes to the B N ) based on the information they contribute. The proposed methodology is readily scalable and generalizable to other monitored wate rsheds because the primary prediction parameters (likelihoods in this case) are learned from the data itself. Posterior probabilities of streamflow estimates and entropy values provide valuable information in determining sites for future instrumentation. 22 Study Site and Data Description The study site for this work is the Santa Fe River w atershed (SFW) in North Central Florida, USA ( Figure 21). It shares many climatic, hydrologic geologic, and land use characteristics of the larger Suwannee River bas in of which it is a part but at a smaller scale (i.e. 3700 km2) It is thus an ideal test bed for examining scientific and engineering questions pertinent to the larger Suwannee basin. The SFW exhibits large variations in the relationships between rainf all and streamflow as a function of location due to changes in draining (contributing) area, soil type, geology, and vegetation ranging from farm land to forests. The SFW also experiences dramatic variations in the amount and frequency of precipitation, ex hibiting dry conditions from late winter to early spring, convective and tropical storms during the summer and early fall with high rainfall rates, and frontal systems in the late fall through late winter with low rainfall rates. The region commonly exper iences extreme events (e.g. flooding from tropical storms and droughts associated with the North Atlantic multi decadal oscillation) PAGE 28 28 The most distinguishing feature of the SFW, however, is the fact that the river traverses a major hydrogeologic divide the Cody Escarpment dividing the watershed into two major regions. The Cody Escarpment is the erosional edge of the confining unit, which includes clay rich Hawthorn Group rocks. On the upstream side of this boundary, the Hawthorn clay acts as a conf ining layer retarding the movement of surface water into the underlying Floridan Aquifer. Thus, a dense network of streams and rivers formed in this region transport surface water to the outlet of the watershed. On the downstream side of the boundary, the clay layer is absent, resulting in karst terrain ground water discharges to the surface through numerous springs. As the Santa Fe River crosses the boundary, it exhibits significant transitions in the manner and magnitude of surface and ground water inte ractions, with important implications for the fate and transport of water and pollutants [ 35, 36 ]. The SFW therefore serves as a good test site on which to develop hydrologic estimators because these varied characteristics cause it to exhibit relationship s that can be found at many other watersheds. However, the complexity also implies that many of the relationships between rainfall and streamflow are non Gaussian and non stationary in space and time 221 Observed Data Topography data from the USGS Nat ional Elevation Dataset (NED) [ 37] was used to determine the region of the watershed area that drains rainfall and runoff to the particular location in the stream wher e streamflow must be predicted referred to as th e contributing area for that location Streamflow measurements were obtained from the USGS NWIS database for locations indicated in Figure 2 1. Ground water levels (GWL) from monitoring wells (Figure 2 1) were provided by the Suwannee River Water Management District (SRWMD) at sparse locatio ns These values were then linearly interpolated spatially to obtain ground water levels ( Gx y) at all points across the watershed. Rainfall estimates ( Rx y) are obtained at moderate spatial PAGE 29 29 resolution (260m) and high temporal resolution (daily) from the National Weather Service NexRAD database for February 2001 through December 2006. Wetland classes and coverage (land cover) were provided by the National Wetland Inventory (NWI) and underlying soil characteristics (soil drainage) was obtained from the So il Survey Geographic (SSURGO) Database at 260m resolution 222 Derived Data The Watershed Assessment Model ( WAM ) is a watershed scale model ( Soil and Water Engineering Technology, Inc. 2008) that predicts stream flow as well as the transport and transf ormation of particulate and soluble phosphorus, particulate and soluble nitrogen, total suspended solids and biochemical oxygen demand in the stream system The conceptual hydrologic model underlying WAM includes rainfall, evapotranspiration, overland flow groundwater flow and river flow The model takes Geographic Information System (GIS) layers for landuse, soils, topography, streamflow basin boundaries, and meteorological data such as rainfall as inputs. It uses a GIS raster or grid cell representation of a watershed Based on soils topography and landuse in each grid cell, the GLEAMS field scale model [41] is run to generate overland and groundwater flow (and associated water quality constituents) produced from each grid cell on a daily basis. The mo del delivers the daily overland and groundwater flows and constituents to the nearest down gradient stream based on predetermined flow paths and empirical flow velocities and hydrographs that are different for surface water and groundwater but are assumed constant over the modeled domain in both space and time There is no cell to cell interaction as the water and its constituents are ro uted to the stream. Complex p hysically based models such as WAM, require very specific input data and are usually employ ed for long term scenario analyses. When applied to watershed s where large amounts of input data are available for calibration the WAM model can perform well for PAGE 30 30 ran king of alternative future land use and water management practices, but typically is not a ccurate enough for streamflow forecasting However, a framework that utilizes information from physical models, whenever available, is desirable. In this work, we analyze the utility of WAM for improving real time streamflow networks by incorporating it as a node in the Bayesian network. In particular, streamflow predictions from the WAM mo del configured for the SFW by Sriva stava et al [41] streamflow predictions were evaluated for incorporation into the Bayesian Network. WAM predictions of streamflow c oncentrations for nutrients, sediments and BOD were not utilized in this study. Predictions from a calibrated hydrologic model are not always available and therefore we also analyzed the utility of using a simpler estimate of surface runoff using the SCS C urve number method as a feature in the estimation framework. This is more computationally efficient than considering land cover soil drainage and antecedent moisture as individual features in the estimation framework. The Soil Conservation Services (S CS) Runoff Curve Number ( Equation 21) method [ 38] was employed to generate daily runoff contributions ( Qx y) across the contributing area at point ( x ,y ) where Ia=0.2 s s =(1000/ Cx y) 10. otherwise I R if S I R I R Qa y x a y x a y x y x0 ) (,, 2 (2 1) Cx,y and Rx,y are the curve number and current days rainfall at location ( x .y ) in the watershed, respectively. Wetland classes land cover and underlying soil characteristics (soil drainage) were employed to estimate Cx y. Since curve number is known to vary with time due to fluctuations in antecedent soil moisture, commonly used adjustments for curve number under dry and wet antecedent moisture conditions ( which can be found in the hydrologic literature, such as [ 39, 40] ) were used Antecedent moisture condition for a particular day was de termined by PAGE 31 31 accumulating rainfall from the preceding 5 days ( RA). Further, flow at the outlet station is affected by recent runoffs (through rainfall events) proximal to that point and also by earlier events at points farther upstream. Time dependencies between rainfall and streamflow we re inferred from second order statistics and are used to determine lags ( lx y) experienced by a surface flow component originating at point ( x ,y ) to reach the watershed outlet. Peak t emporal cross correlation times (i.e., lag at which cross correlation is maximum) between every rainfall pixel within the contributing area ( Rx y) and flow at the outlet were computed as the lag ( lx y) corresponding to that pixe l. Runoff s ( Qx y) over different portions of the basin are groupe d by their corresponding time lags ( lx y) to generate net runoff ( Qnet) for the contributing area (Equation 2 2) y xl y x y x netl t Q t Q,) ( ) (, (2 2) Table 21 summarizes the potential inputs for the Bayesian models utilized in this work. The observed features are generally widely available for watersheds in the United States and in Europe and are known to strongly impact streamflow. The derived features are also relatively easy to estimate for most watersheds in the United State and Europe. All data were obta ined through the University of Florida Water Institute [ 41 ]. 23 Hardware Acceleration on FPGAs The three algorithms ( non parametric PDF estimation, K means clustering, and correlation ) targeted for hardware acceleration in addition to being an integral p art of the framework proposed in this work, form an important core in many applications The Parzen window based non parametric PDF estimation algorithm is used for many applications in pattern recognition, such as fingerprint recognition, Bayesian classi fication, feature extraction [ 7], bioinformatics [ 5], networking [ 42 ], stock market prediction [ 43], and image processing [ 6 ]. The PAGE 32 32 authors in [ 9] employ cross modal correlations for clustering features in an image audio dataset. Clustering and feature ex traction methods are applied for video content retrieval and analysis in [ 44]. A novel video segmentation technique was presented in [ 4] based on correlating audio and video data. A simplified probability based methodology using histograms was proposed i n [ 8] for video analysis. Hence, a large number of applications can benefit from fast computations of PDF estimation, K means clustering and correlation, especially in time critical missions. FPGAs lie between general purpose processors and ASICs on the spectrum of processing elements in that they are highly flexible (like processors) and also have potential for high performance (hardwar e acceleration). Commonly used deterministic benchmarking cores, such as Fast Fourier Transforms (FFTs) [ 45], convolut ion, Lower Upper triangular matrix Decomposition (LUD) [ 46], and Basic Linear Algebra Subprograms (BLAS), have been effectively implemented on FPGAs by virtue of their inherent multi level functional and data parallelism. In addition to these benchmarking cores, significant performance gains have been achieved with FPGAs by selecting, developing, and testing diverse applications from various fields in signal and image processing. Previous works have also discussed the feasibility of targeting some structu rally similar algorithms for FPGAs. In [ 47], the authors use a Gaussian kernel estimator as one of their case studies in presenting a MATLAB to VHDL application mapper. It is a subset of the algori thm considered here, in that the streamflow estimation fr amework employs Gauss ian kernels to estimate PDFs and is hence more computationally intensive. In [ 48], the authors present simpler algorithmic variants of the K means algorithm by considering Manhattan and Max distances instead of Euclidean distance for an efficient hardware implementation. Significant acceleration of the algorithm with acceptable accuracy was achieved in clustering a sample dataset. This work is different from [ 48] in that Euclidean PAGE 33 33 distance is used for implementing K means (which is t he preferred method in most machine learning applications) and, moreover, the primary motivation is to accelerate a set of machine learning algorithms used in different stages of estimation by reusing hardware designs leading to shorter design times. Prev ious work in parallel processing for large data volumes is also relevant to this work. An object recognition application was developed in [ 49] by traversing huge volumes of data in a parallel fashion. The paper showcased how a data parallel programming m odel can be used to solve a general set of data intensive algorithms. In many of these examples, the applications could achieve significant performance gains if there was a way to rapidly compute multidimensional PDFs. PDF estimation enables the computa tion of metrics such as noise, error rates, and uncertainties that are very important for researchers in the application domain [ 5,6,7,50]. The authors in [ 51] and [ 52] had developed several signal processing applications on FPGAs in order to obtain speed up and discussed tradeoffs in solution accuracy and resource utilization. However, these papers only predict performance factors in a relatively limited setting. Various performance prediction tools have been proposed [ 53, 54] that base their techniques on parameterizing the algorithm and target FPGA platform. Algorithms are decomposed and analyzed to determine their total size (in terms of resources) and computational density. Computational platforms are primarily characterized by their memory structur e capacity and interconnect bandwidth. In particular, the RC Amenability Test (RAT) presented in [ 53] is a simple methodology that suggests a step by step procedure to predict the performance, in terms of speedup, of a specific design for an application on a specific platform before coding begins. This pre implementation test also helps the designer to understand the strengths and weaknesses of a particular algorithm towards parallel implementation. Communication and computation parameters in RAT are q uantified based upon algorithm and PAGE 34 34 FPGA platform characteristics. An estimate of performance in terms of execution time in hardware is then derived from the analytical models embedded in RAT. Comparing the hardware prediction against a known execution ti me for a software baseline leads to the overall speedup estimation. In this work, we use RAT to predict performance of an algorithm on multiple FPGA platforms. B uilding a good reconfigurable design requires skill and effort, and is often accompanied by a steep learning curve for designers without prior FPGA design experience. In [ 55], the authors emphasize the importance of identifying and cataloging an exhaustive list of design patterns to solve recurring challenges in RC and increase productivity. In this work, the utility of design patterns is validated by identifying and classifying relevant primitive patterns. The patterns are then quantified and applied for algorithm decomposition and performance prediction using RAT. 24 Case study Algorithms Kn owledge of the algorithms data flow and computational complexity is essential in order to make strategic decisions during FPGA design development. A n overview of the algorithms used as case studies in this work is presented and commonalities in their str uctures are investigated in this section 241 P robability Density Function Estimation The common parametric forms of PDFs (e.g., Gaussian, Binomial, Rayleigh distributions) represent mathematical idealizations and, as such, are often not well matched to densities encountered in practice. In particular, most of these classical parametric densities are unimodal (having a single local maximum), whereas many practical problems involve multimodal densities. Furthermore, because of correlations among the data features, it is unusual for high dimensional density functions to be accurately represented as a simple product of one dimensional densities, the latter of which are significantly easier to compute. Thus, an estimate PAGE 35 35 of the joint (multi dimensional), non parametric PDF (i.e. a PDF that assumes no particular functional form) is often required. The computational complexity of the Parzen window based PDF estimation algorithm is O( Nnd), where N is the total number of data points, n is the number of discrete points at which the PDF along a dimension is estimated (i.e. bins) and d is the number of dimensions. Mathematically, the probability that point i falls in a ddimensional space is given by Equation 2 3 where h is the bin size, which acts as a tuning pa rameter for resolution of the PDF estimate, the set ( xj,yj) represents the d subsets of bin centers at which the PDF is estimated, and the set ( xi,,yi) represents the ith input data point in a ddimensional space, where i ranges from 1 to N 1 1 11 111 () ...(,,)...(,,) ...d d dn n j j ii d jjpi xxhyyh nn (2 3) The kernel function could be as simple as a histogram function, a more general rectangular kernel, or the widely favored Gaussian kernel. The first two cases fall under the class of nave estimators. In the first case (histogram function), the data range is divided into a set of successive and non overlapping intervals (bins), and the frequencies of occurrence in the bins are plotted against the discretized data range. The rectangular kernel case is similar except that overlapping intervals are permitted. In either case, the bin size should be chosen such that a sufficient number of observations fall into each bin. The resulting PDF estimate depends on the bin size as well as the discretized range of the dataset and is discontinuous at the bin boundaries. Although nave estimators yield discontinuous results, the construction can be easily generalized to achieve continuous PDF estimates by employing different kernel functions. The smooth reproduction of the und erlying Gaussian process in Figure 22A clearly shows the adv antages of Gaussian kernels in this aspect and is thus the motivation for its use here. The PAGE 36 36 mathematical formula for a Gaus sian kernel and t he resulting expression for the 1 D Parzen window PDF estimate are defined in Equations 2 4 and 25 respectively wh ere h plays the role of the standard deviation 2 22 ) (2 1 ) (h x x j ij ie h h x x (2 4) n j h x x jj ie hn x p1 2 ) (2 22 1 ) ( (2 5) Computing complex exponentials is challenging in hardware because it requires significant hardware resources. To make the algorithm more suitable f or hardware, a truncated second order Taylor series expansion replaces the exponential function for the development of the core in hardware (Equation 2 6). 22 22()() 1 1 ,for1 0 (,) 22 2 0,otherwisejj ii j ixx xx xx hh h (2 6) The degree to which the quadratic (Equation 2 6) approximates the true Gaussian kernel (Equation 2 4) decreases for data points that yield large values of | xixj|. T hat condition is avoided by giving zero weight to points that cause the argument 1( xixj)2/2 h2 to be less than zero, which occurs when | xixj| is greater than h. The plot in Figure 22A for the Gaussian kernel employ s the approximation given by Equation 26. Additional management of approximation error is achieved by pre scaling the kernel to unity variance prior to computations on the FPGA. The datasets xi and xj are scaled by h before being transferred from the processor to the FPGA, which reduces the factor 1 ( xixj)2/2 h2 to 1 ( xixj)2, thus making the computational error due to the use of the Taylor series representation on the FPGA independent of the particular choice of h. PAGE 37 37 The dynamic range of the dataset can further be reduced by considering only the higher order bits during computation. This method is a simple task to implement since FPGAs are extremely efficient at performing bitlevel manipulat ions. 242 K M eans Clustering In data analysis problem s, given a training set of points, unsupervised learning algorithms are often applied to extract structure from them. Typical examples of unsupervised learning tasks include the problem of image and text segmentation [ 3 ]. A simple way to represent data is to specify similarity between pair of objects. If two objects share the same structure, it should be possible to reproduce the data from the same prototype. This idea underlies clustering methods that form a rich subclass of unsupervised algorithms. The K means algorithm is a popular unsupervised clustering algorithm that partitions N data points into m clusters by finding m mean vectors 1, 2,, m representing the cen troids of the clusters (Figure 22B shows an illustration of data points represented by two features partitioned into 4 clusters). It is an iterative algorithm that tries to minimize the total intra cluster variance or the squared error function (Equation 2 7) where there are m clusters Sj, j = 1, 2,, m and i is the centroid of all the points xi in Sj. 2 1ijm ij jxSVx (2 7) To begin with, the data points are partitioned into m initial sets, either at random or us ing some heuristic approach. The centroid of each set is then computed and a new partition is constructed by associating each data point with the closest centroid. The centroids are then recalculated for the new clusters. This process is repeated until data points no longer switch PAGE 38 38 clusters or alternatively centroids remain unchanged. The computational complexity of the algorithm is O( NdmT ) where d is the number of features representing the data point and T is the number of iterations. 243 Correlation Correlation computes a measure of similarity between two datasets (e.g., audio signals) as they are shifted by one another across time. Correlation is often used to measure the delay between two signals having the same shape with one delayed relative to the other. Autocorrelation is the correlation of a signal with itself while cross correlation is the correlation between two different signals. The correlation result reaches a maximum at the time when the two signals match best. For example, if the two signals are identical, this maximum is reached at t = 0 (i.e., no delay). Autocorrelation is useful for finding repeating patterns in a signal, such as determining the presence of a periodic signal that has been buried under noise, or identifying the fun damental frequency of a signal that does not actually contain the fundamental frequency component but implies it with many harmonic frequencies. Mathematically the correlation between two discrete signals xi and yi with n samples each, where is the delay and m is the maximum delay is given in Equation 2 8. n m m y x Rn i i i ; 0 ; ) (1 (2 8) 244 Algorithm Similarity Although the algorithms overviewed in the previous sections address different problems, involve different computations, and appear at different stages during data analysis, the underlying data flow within each algorithm has significant similarity. The data flow in these algorithms follows an exhaustive data permutation pattern wherein the interaction between every sample in two input vectors I1 and I2 affect s the value of every sample in an output vector O In PAGE 39 39 PDF estimation, it is between the data points ( I1) and the points at which the PDF is estimated ( I2). In K means clustering it is the Euclidean distance between the data point s ( I1) and the cluster centers ( I2), and in correlation it is the multiplication of two signals ( I1 and I2) shifted over different time delays. Figure 23 illustrates the dataflow and computations involved in each of the three algorithms in a common fram ework. Table 2 1. Set of available features Feature Notation Resolution Type Streamflow at neighboring station TNN Daily Temporal Ground water level GWL Daily Temporal Ground water level G x y 260m Spatial Rainfall R Daily Temporal Rainfall R x y 26 0m Spatial Runoff Q net Daily Temporal Runoff Q x y 260m Spatial WAM estimates WAM Daily Temporal PAGE 40 40 Figure 21. A map of the Santa Fe River Watershed with its streams and rivers [ 41]. Note: Two distinct hydrogeologic regimes are present: the confined ea stern region and the unconfined western region. Locations of streamflow measuring stations are indicated. PAGE 41 41 A 50 100 150 200 250 0 0.002 0.004 0.006 0.008 0.01 No. of binsProbability of data in a bin Histogram rectangular Taylor approx B Figure 22. Illustration of case study algorithms. A ) 1 D PDF estimation us ing three different kernels B ) K means clustering. Note: For the Gaussian kernel, the Taylor series approximation is used. In general, PDFs can be highly non Gaussian (e.g., bimodal) as shown here. Parzen estimates using rectangular, and in particular Gaussian kernels, generally provide estimate s of the true underlying PDF that are far superior to the histogram function. In the case of data clustering, it is desired to partition complex data sets in groups that share similar feature values Figure 23. Dataflow and computational structure in PDF estimation, K means clustering, and Correlation -5 0 5 10 15 20 25 -5 0 5 10 15 20 Feature 1Feature 2 cluster 1 cluster 2 cluster 3 cluster 4 PAGE 42 42 CHAPTER 3 SPATIO TEMPORAL BAYESIAN NETWORK Probabilistic methods that ut ilize the joint (multivariate) PDF of all the features and parameters rather than merely a finite subset of its statistical mome nts can be used to incorporate (fuse) different data sets i n order to estimate streamflow. Bayesian estimation incorporates probabilities based on multiple features that influence flow The many to one mapping of a d 1 feature vector x to streamflow cons titutes a probabilistic data fusion problem, where d denotes the number of data types used as features (Table 2 1) A d iscrete formulation of Bayes rule is used in Equation 3 1 for which streamflow is segmented into fifteen levels or classes ci, i =1,,15. The ranges of stream flow magnitudes that correspond to these classes are independently determined using k means clustering for the three long term flow monitoring stations (2321000, 2321500, 2322500) in the SFW. Fifteen levels ( k =15) were determined to give a reasonable trade off between precision of the estimated flow and maintaining a sufficient number of samples in each class for reliable estimation of the likelihoods via Parzen windowing. T he likelihoods p( x | ci) of the training flow data are thu s estimated in feature space for each flow class. Prior probabilities p ( ci) for the classes are computed based on class memberships obtained from k means clustering. The a posteriori probabilities p( ci| x ) are then computed based on these PDF s for each te st data point and the flow level is labeled according to the class with the maximum a posteriori (MAP) probability. 1()() (),()() ()k ii ii ipxcpc pcx pxpxc px ( 31) In theory this direct application of Baye s rule is optimal with respect to minimizing the probability of error of misclassification Since it is not dependent on physically based equations, this method does not require as much a priori knowledge of the particular hydrologic processes PAGE 43 43 that dominate in a given watershed as do physical models such as WAM. I n that sense, it is more easily applied to different watersheds where different combinations of input data may be available. Of course, the accuracy depends on the particular data types available. In general, the more features available, the better. The critical limitation with this approach arises when there are many features, forcing the estimation of the joint likelihoods in to a high dimensional space. Computational complexity and memory requirements increase super linearly with dimension [ 56] making it often impractical to implement a direct Bayesian classification for joint PDFs of greater than three dimensions over large areas and long time periods. While it is possible to decompose Equation 3 1 from a ddimensional problem into d 1dimensional pro blems to ameliorate the memory demands, this socalled nave Bayesian implementation is far from optimal because it ignores all statistical relationships among the features. Some data driven methods (the Reduced Dimension Bayesian Estimate RDBE) mitigate the increase in computational complexity by using a reduced feature data set, thus tolerating greater error [ 29] BNs offer a formal method for decomposing high dimensional likelihoods into lower dimensional factors by invoking the well known conditiona l probability theorem [ 57, 58 ] thereby reducing computational complexity A comparison of the computational complexity involved in a traditional Bayesian c lassification, RDBE based and BN based methods is given in Table 3 1. I t is also desirable to use a reduced set containing only the most informative features A method determining the information gained by incorporating a new feature in the network is hence needed ( Section 3.3) Such a methodology would also offer a platform to analyze the potential benefit of including physical model estimates as nodes in the BN. Further, well informed tradeoffs can be made regarding computation al complexity and accuracy. PAGE 44 44 Hidden Markov models (HMMs) are also a frequently used tool for modeling time series data tha t belongs to a subset of the more general BNs. It extends the concept of discrete markov models in that the observation is a probabilistic function of the state (i.e., streamflow). The resulting model is a doubly embedded stochastic process with an under lying stochastic process that is not directly observable; it is hidden. The hidden states are assumed to be discrete and can only be observed through another set of stochastic processes that produce the sequence of observations. The advantage of the BN f ramework over H MMs is that it allows for an ar bitrary set of hidden variables with arbitrary independence assumptions. If the conditional independence assumption results in a sparse network, this may result in an exponential decrease in the number of para meters required to represent a probability distribution. Often there is a corresponding decrease in the computational load as well. Hence BNs offer better statistical and computational efficiency 31 Bayesian Network Fundamentals A Bayesian network is a Directed Acyclic Graphical (DAG) model where nodes represent random variables (RVs) and arcs between nodes specify the conditional dependencies between RVs. If there is an arc from node A to another node B A is called a parent of B and B is a child of A The set of parent nodes of a node Xi is denoted by parents( Xi). A directed acyclic graph is a B N relative to a set of variables if the joint distribution of the node values can be written as the product of the local distributions of each node and its pare nts [ 58] ( Equation 3 2). n i i i NX parents X P X X X P1 2 1,..., (3 2) If node Xi has no parents, its local probability distribution is said to be unconditional otherwise it is conditional If the value of a node is observed, then the node is said to be an evidence nod e. The graph encodes independencies between variables. Conditional independence PAGE 45 45 is represented by the graphical property of dseparation: If two sets of nodes X and Y are dseparated in the graph by a third set Z then the corresponding variable sets X and Y are independent given the variables in Z The minimal set of nodes which dseparates node X from all other nodes is given by X s Markov blanket. dseparation is defined as follows. A path p is said to be dseparated (or blocked) by a set of nodes Z if a nd only if Path contains a chain p m j or a fork i m j such that the middle node m is in Z Path contains an inverted fork (or collider) i m j such that the middle node m is not in Z and no descendant of m is in Z A set Z is said to dsepar ate x from y in a directed acyclic graph G if all paths from x to y in G are dseparated by Z The d' in dseparation stands for 'directional', since the behavior of a three node link on a path depends on the direction of the arrows in the link. This property offers a formal method for decomposing high dimensional likelihoods into lower dimensional factors by invoking the well known conditional probability theorem and the effects of causality. The fundamental connections in a BN that would be used in this work and commonly found in literature are illustrated in Figure 3 1. In most of the applications where BNs are employed, the nodes are either qualitative representations of some random phenomenon/occurrence or a quantitative measure of a RV observ ed in a particular domain (time or space but not both ). A BN framework that offers the capability to embed both spatial and temporal RVs in proposed in this work as illustrated in Figure 3 2. The set of immediate neighbor nodes of any node A should be observed i n the same domain as that of A Domain transformation is obtained through aggregate functions o TYPE I Si mple statistical function (e.g., sum, minimum/maximum, mean ) values o TYPE II Physical model based m appings depending on the application o TYPE III Pr obability distribution based transformations PAGE 46 46 In the sections to follow, the proposed BN framework is applied to the problem of estimating streamflow in the Santa Fe watershed system. The application, in addition to serving as a proof of concept, would al so help bring out some of the advantages in choosing a BN formulation 311 Network Learning Probabilistic methods can readily be applied to different monitored watershed s because they rely on learning the parameter PDF s from the available data in a str aightforward manner using methods such as Parzen windowing [ 59]. As with any learning approach, the accuracy of the PDF estimates depends on how well the particular data sets sample the probability space. In general, time series data used for training sh ould have daily time steps and be at least a few years in duration to capture annual variations in streamflow. Only three of the flow monitoring stations in the SFW (2321000, 2321500, 2322500) had continuous records of measured streamflow over the time pe riod that the NexRAD data was available. Thus, these three stations were used to provide training data An ideal scenario to estimate PDF s using parzen windows is the availability of infinite number of data samples (i.e. n h to a significantly small value (i.e. hn 0) in Equation 2 3. However, the application under study is characterized by historically short datasets requiring better strategies in the selection of a suitable h. For fixed n, small values of h imply higher variance leading to a noisy appearance of the resulting PDF estimate. On the other hand, large values of h represent an excessively smoothed version of the true underlying PDF While there are approaches [ 60] that discuss optimal ways of s electing kernel sizes, they are either primarily targeted towards the representation of individual PDF s or are computationally very demanding [ 60 ]. The objective in this work is to be able to represent every possible sample of the feature vector x within the range of the training set, with a streamflow class realized by at least one of the training stations (i.e. for all [1,15] and PAGE 47 47 j [2321000,2321500,2322500], there exists a p ( x | ci,sj) h is initialized to some small value and in cremented gradually until the objective is satisfied. In general, it was observed that smaller kernel sizes are better able to capture short term rainfall and flow events and thus tend to work well for stations in the upstream confined regions of the SFW. Slightly larger kernels perform better for stations in the unconfined region where flow depends on longer transit times for rain falling over a larger contributing area to impact flow at the station. 312 Network Topology Structure of the networks use d for estimating streamflow is determined based on the physical knowledge of the underlying process es The impact of nodes forming the BN is quantitatively analyzed later in Section 3 4. Runoff is a spatio temporal process impacted by the joint interaction of spatial rainfall, land cover, and soil moisture (i.e. percentage of rainfall intercepted at a point ( x ,y ) contributing to surface runoff is affected by the land cover and level of antecedent soil moisture at that point). Inclusion of SCS estimated runoff and WAM streamflow predictions as nodes in the BN i s redundant since both attempt to capture the impact of spatial rainfall, landuse, and soil moisture on streamflow. However, estimates from a calibrated hydrologic model are not always available an d so we analyze the performance of two BNs: one that excludes WAM (N1) and another that includes WAM (N2) as shown in Figure 3 3. In the absence of land cover and soil drainage measurements that characterize curve number ( Cx y), runoff values in network N 1 cannot be directly computed using the SCS method. To address this issue, SCS estimated runoff (from training period) is conditioned and inferred from antecedent ra infall and spatial groundwater levels as these are known to primarily impact runoff and als o more readily available Note that groundwater level was used as a surrogate for soil moisture since it is more commonly measured. In N1, Qx,y at a particular point on the terrain is conditioned on antecedent rainfall RA (in this work, rainfall is accum ulated over the preceding 5 PAGE 48 48 days) and ground water level Gx,y at that point (handles spatial nonstationarity). Note that groundwater level was used as a surrogate for soil moisture since it is more commonly measured and thus more readily available. Qnet at time t is the sum of all runoff s Qx,y over the basin as discussed in Section 2 2. Streamflow F at a current station is then conditioned on Qnet (or WAM in case of N2), streamflow at the station nearest to it ( TNN ), and ground water level at a well clos est to the station ( GWL ). Conditional probability tables (CPTs) for all parentand child node connections in the networks are then computed in a straightforward manner. Computing these CPTs from the training data can be viewed as a learning step. 313 Network Inference Given spatial rainfall data, spatial and temporal ground water level (or WAM estimates), and flow at a nearby station, the most probable Qx,y Qnet, and F values are inferred from the graph. The expected value of estimated flow on a par ticular day d is given in Equation 3 3 where p( ci) is the posterior for the ith class i is the mean flow value for the ith class and i [1,15] 1()()k ii iEfpc (3 3) The classes and their corresponding i values were determined by segmentin g the training data using k means clustering. The error in model prediction ( epred) can be defined as the deviation of the maximum a posteriori flow from the expected flow (Equation 3 4). The actual error ( eactual) is defined as the deviation of observed flow from the expected flow (Equation 3 5). Reliability of network performance can be inferred by comparing actual errors and expected errors (suggested by the model). PAGE 49 49 1 1 1() ()k pred iiMAPMAP i k ii iepcp pciMAP (3 4) 1()k actual iiobs iepcf (3 5) 314 Performance Metrics In addition to probabilistic performance measures like error entropy ( H ) and maximum a posteriori probability, the widely used Nash Sutcliffe coefficient ( NSC ) is employed as a goodness of fit measure to analyze performance [61] (Equation 3 6). j j j N i mean i N i i ie p e p H O O P O NSC ) ( log ) ( ; ) ( ) ( 11 2 1 2 (3 6) In Equation 3 3, Oi is observed streamflow, Pi is predicted streamflow, Omean is the mean of observed streamflow, and N is the total number of data points. Since NSC utilizes some form of deviation from the mean, it can be e xpected to provide reasonable indications of accuracy primarily when the underlying PDF is uni modal and not heavily skewed. Therefore, additional metrics based on the signed error are also examined. The signed error is defined as predicted minus observe d streamflow values, so that a positive error implies overestimation. The mean (bias) and standard deviation of the error are computed. Finally, the error PDF is estimated via the normalized Parzen windowing procedure described in Section 2 41, thus ena bling the calculation of error entropy. A Gaussian kernel with size h=1 and N =128 was used in all cases to estimate the error PDFs. PAGE 50 50 32 Measurement of Incremental Information When estimating streamflow at a location with no in situ data, it is important t o know which features will contribute the most information to limit the dimensionality and complexity of the problem. One way to pick the most informative features is a system of feature ranking based on measures such as mutual information (MI) or conditi onal entropy (CH) amongst the flow classes. The feature combinations that provide the most separation among all the classes are considered best. The MI gives a quantitative way of measuring this separation. Given a feature vector x and class c the best subset of x is the set that gives the largest MI between the feature PDFs and class membership based on the training data. Another way to approach this problem is to minimize CH of streamflow classes with respect to the feature vector The Shannon MI, e xpressed as I ( x ; c )= H ( x ) H ( c | x ) reveals a dependence on conditional entropy. When relatively few training samples are available for one or more classes, however, the estimate of H ( c ) can become biased. Using the CH of streamflow class c given feature vec tor x avoids this dependence on the sole entropy of c in MI methods. 321 Feature Ranking In the proposed approach the feature nodes are ranked using CH and the performance of networks are compared using the suggested feature s. The hydrologic makeup o f the study site is heterogeneous making it difficult to find a single BN to accurately predict stream discharge at all locations in the watershed. T hree USGS flow measurement stations in the watershed are use d as test sites. Stations 2321000 and 2321500 are in the confined area, while station 2322500 is in the unconfined area. The feature ranking algorithm is applied separately to each of the three stations th at aid in obtaining an optimal yet simple structure at each location. Temporal features in Ta ble 2 1 are considered for this analysis (rainfall R was not included as its effect was embedded in WAM and Qnet). PAGE 51 51 First, the information metric (i.e. CH) is used to determine the single most informative feature at each of the test stations (rankings sh own in First Feature rows of Table 32 with the best ranked feature in column Rank 1 ). T he remaining features each paired with the best feature from the previous step (Second Feature in Ta ble 3 2) are ranked in the next step T his process is cont inued until either all features have been ranked or the required number of features has been ranked Each time a feature is added to the network, the information metrics are computed for the entire network. The decrease in values along the columns in Tab le 32 indicates features with additional information are being added to the network, reducing CH The features under column Rank 1 in Table 3 2 are those chosen as the best three features at each training station. Since streamflow was divided into 15 classes, the highest CH value of class separability is 2.7 (i.e. loge15). At stations 2321000 and 2321500, most of the information is contained in feature TNN. This is because these stations are located in the upstream confined area of the watershed wh ere rainfall has an immediate and shortlived effect on discharge. In such cases where groundwater does not contribute significantly to streamflow, upstream flow is a very good indicator of flow at a downstream station. Adding groundwater and WAM contrib ute only a 1.4% reduction in CH, while TNN eliminates 41.6% of the uncertainty at 2321000. At 2321500, TNN and GWL eliminate 33.8% uncertainty and adding WAM eliminates only 0.5% more. Low CH values and smaller incremental information gains at these two stations indicate that a complex BN is probably not necessary. Station 2322500 lies in the unconfined region, where groundwater is the dominant contributor to streamflow. Here, flow at station 2321500 was used as the TNN feature which lies in the confin ed region. Due to this relationship, TNN is not chosen as the best feature. PAGE 52 52 Streamflow in the unconfined region is affected by a number of variables making it hard for one feature to adequately characterize it (net reduction of 30.2% in CH values with 3 features). 322 Validating Information G ain Analysis In order to validate the results of information gain analysis, performance of networks constructed from combinations of features were studied. The study would also warrant the use of CH as a metric f or information analysis in the strategic construction of simpler networks leading to reduced but acceptable performance. The following set of scenarios was considered. CASE 1. S ingle node network containing the rank 1 feature CASE 2. S ingle node networ k containing a feature that was not ranked best CASE 3. T wo node network that contains the rank 1 feature CASE 4. T wo node network containing features not ranked best Table 33 summarizes the results obtained for the case study scenarios at each of the three training stations for the testing period. Low CH values and smaller incremental information gains at stations 2321000 and 2321500 indicate that a complex network is probably not necessary in confined regions for achieving satisfactory performance. Addition of TNN to WAM at 2322500 leads to significant improvements in performance supportive of the large reduction in CH values in Table 3 2. Performance metrics for Case 4 networks decreased dramatically at all three stations validating the efficacy o f the proposed information gain analysis. 33 Model Predictions at Training Stations It is impractical to assume the availability of a calibrated hydrologic model for all watersheds, and is hence important to analyze the performance of the two networks (F igure 33) over the same watershed to infer if they can be used interchangeably. To gauge network performance and validate the findings of information gain analysis, the performance metrics mentioned in Section 3 2 are employed to ascertain estimation acc uracy. PAGE 53 53 T he probabilistic model s w ere trained using data obtained from 2001 2004 and tested during the period 2005 2006 in order to ascertain its predictive potential. Table 34 shows the performance of the network incorporating Qnet ( N1 ) and the net work incorporating WAM (N2) in terms of the metrics presented in Section 324. To avoid misjudgments, NSC entropy ( H ), and mean absolute error ( e ) metrics should always be analyzed in tandem while evaluating the performance of a model. For example, a m odel that consistently over estimates streamflow could still have a small entropy value due to a peaked uni modal error distribution. Also, if the error distribution is not uni modal, analyzing a secondorder moment based metric like NSC without consideri ng entropy values can lead to misinterpretation of model behavior. Taking these into consideration, it can be inferred from Table 3 4 that N1 and N2 perform equally well in the confined regions, achieving high NSC and low mean absolute error and entropy values. In general, the performance degrades in the unconfined region with lower NSC and larger mean and entropy values with N2 outperforming N1 primarily due to the inclusion of WAM estimates. The estimated and observed streamflow values over the test p eriod for the three stations are shown in Fig ures 34, 35, and 36 for N1 and N2. Maximum a posteriori probability (MAP) values were used to analyze performance. Since the distributions were non Gaussian, investigating error as a multiple of standard d eviation would not be meaningful. T he MAP values obtained over time is very similar for the two networks although higher values are achieved throughout the estimation in case of N2 ( supports the fact that WAM ranked higher than runoff in Section 32 in t erms of information contribution). The following analyses apply to the performance of both networks. MAP values are moderately high at stations 2321000 and 2321500 during heavy rain/drought events. Since flow in confined regions is predominantly charact erized by rainfall and surface runoff, inclusion PAGE 54 54 of these nodes in the network leads to good performance. However, lower MAP values are achieved during intermittent rain events. This suggests that a component that efficiently captures complex short term sub surfa ce behaviors is required for improving the reliability of flow estimates during such periods. In the unconfined region, s ub surface interactions contribute to a persistent baseflow. Since baseflow is neither captured by WAM (due to simplified gr ound water representation ) nor nearest neighbor flow (i.e., flow at 2321500 located in the confined region), MAP values are low during baseflow dominant periods Relatively higher MAP values are achieved while predicting flow during rain events when basef low is augmented by surface runoff. The performance drop for N2 was also analyzed under reduced percentage of training data (i.e. a percentage of days starting from 2001 instead of all days from 20012004) and was found to be largely negligible ( Table s 35 and 3 6), particularly in the confined region The reduction in entropy values at station 2321000 is counter intuitive as this is accompanied by a reduction in NSC values and an increase in mean absolute error. The steeper performance drop in the unc onfined region is naturally understandable because of the use of fewer pertinent features in the network. This supports the notion that when more relevant features are used in the prediction, a shorter temporal record of data can still give satisfactory r esults. This study is important since many watersheds have limited historical data, in some cases because in situ sensors were deployed only recently and in others because existing sensors were not maintained continuously. Time series of actual and pred icted errors are plotted for all three training stations over the test period in Figures 37, 38, and 39 along with the PDFs of the error time series. It can be PAGE 55 55 observed that the error predictions are significantly close to the actual errors, particular ly at the confined regions, indicating reliability of model performance. 34 Model Predictions at Intermediate Stations One of the primary motivations for developing data fusion estimation methodologies for watersheds is to eventually be able to estimate streamflow, and subsequently transport rates for sediment and pollution, in large river basins that can only be sparsely instrumented with in situ sensors. Viewed another way, it is desirable for the estimators to be spatially explicit such that they can be applied to any arbitrary location along the reach of a stream or river It is also highly desired that, in a sparsely distributed sensor problem, all available information be exploited for achieving best performance. To estimate flow and uncertainty at any intermediate test point along the river, information from all available training stations are fused (Fig ure 310) Similarities in feature space are calculated between training stations and intermediate test points (given WAMk, TNNk,GWLk) which are then used to weigh individual flow values estimated from each training station. T he flow and posterior probabilities obtained on October 4th, 2006 at points all along the river is shown in Figure 3 11. Locations where estimates are accompanied by low posterior probabilities (i.e. high uncertainty values) can be considered as future instrumentation sites. The model indirectly refers to heterogeneity in the feature vector (i.e. the feature vector is not effectively captured in the probability space desc ribing the BN) at these locations. The formulae used to estimate expected flow and maximum a posteriori probabilities at intermediate points along the river are given in Equation 3 7. The error analysis, explained in Section 3 23 is applied to validate performance of the models at intermediate points. Availability of sparse streamflow measurements imposes considerable constraints in performing such an analysis at every intermediate point along the river. The methodology was hence validated at USGS stat ion 2321975 (located between stations PAGE 56 56 2321500 and 2322500: Figure 21) that had measured streamflow for the test period. The observed and predicted streamflow values at 2321975 are plotted in Figure 312. The estimates were predicted with an NSC value of 0.29, mean absolute error of 530ft3/sec, and an entropy value of 2.96nats. The actual and predicted errors along with their PDFs, shown in Figure 313, suggest reliable model performance. (|)(|,)(|) (|)(|,) (|)() max (|)()(|)(|,) (|)(|) (|,)(|) (|)(|,) (|kjk jkiik i jikji kii j kll jikji i lj kk jkjk j jjkiik ji ikjjki ij kpflowxpxspsx pspxs pxsps pxspspspxs Eflowx pflowx pxspsx psxpxs px ,(|)(|,) )() (|)()(|)(|,) Meanstreamflowinclass;[1,15];featurevect oratpoint ,[2321000,2321500,2322500];()1/3,jikji ii j kll jikji ij lj jk il ilpspxs sps pxspspspxs jjx k ss psil (3 7) PAGE 57 57 Table 3 1. Comparison of computational complexity between traditional Bayesian method, RDBE and STBN Method Operation Complexity Description Bayes PDF estimation O ( Nn d ) d No. of reduced dimensions or features N No. of training samples n No. of discrete levels alon g a dimension RDBE PDF estimation O ( Nn t ) t No. of reduced dimensions or features N No. of training samples n No. of discrete levels along a dimension Entropy est imation O ( mNn t ) t No. of dimensions in reduced set N No. of training s amples n No. of discrete levels along a dimension m No. of class pairs BN PDF estimation Max i O ( Nn i ) i No. of incoming arcs to a node in the tree N No. of training samples n No. of discrete levels along a dimension Table 3 2. Feature ranking based on conditional entropy (CH) Stations Rank Rank 1 Rank 2 Rank 3 Rank 4 2321000 First feature TNN :1.5768 WAM:1.8315 GWL:1.8764 Q net :1.9576 Second feature GWL :1.5480 WAM:1.5587 Q net :1.5759 Third feature WAM:1.5397 Q net :1 .5477 2321500 First feature TNN :1.8980 WAM:1.9033 GWL:1.9243 Q net :2.0952 Second feature GWL :1.7872 WAM:1.8443 Q net :1.8881 Third feature WAM:1.7735 Q net :1.7848 2322500 First feature WAM:2.0097 GWL :2.0718 TNN :2.1713 Q net :2.4283 Secon d feature GWL :1.9934 TNN :1.9935 Q net :2.0074 Third feature WAM:1.9318 Q net :1.9844 The set of available features reduces by one wi th each column, represented by e ntries Table 3 3. Performance metrics for cases 1 4 validating information gain analysis Station Case no. Feature NSC E H 2321000 1 TNN 0.6434 94 2.24 2 WAM 0.3312 88 1.98 3 TNN, GWL 0.6434 92 2.29 4 GWL, Q net 4.4210 523 2.14 2321500 1 TNN 0.7727 215 3.01 2 WAM 0.4330 206 2.12 3 TNN, GWL 0.7717 202 2.41 4 GWL, Q net 1.8580 926 2.58 2322500 1 WAM 0.0440 327 2.44 2 TNN 0.0240 345 2.93 3 WAM, TNN 0.2581 272 2.47 4 GWL, Q net 1.9340 840 3.67 PAGE 58 58 Table 3 4. Performance metrics for networks N1 ( Qnet) and N2 (WAM) Metric Network 2321000 2321500 23 22500 NSC N1 0.6865 0.7664 0.0371 N2 0.6486 0.7183 0.2586 Mean abs. error (cu.ft/sec) N1 83.0000 207.0000 317.0000 N2 64.0000 174.0000 272.0000 Std.dev of error (cu.ft/sec) N1 103.0000 182.0000 435.0000 N2 125.0000 248. 0000 387.0000 Entropy of error PDF (nats) N1 2.0400 2.4100 2.7900 N2 2.2000 2.1500 2.5300 NSC : Nash Su tcliffe coefficient ranges from Table 3 5. Performance metrics for N2 under 80% Training Data Metric 2321000 2321500 2322500 NSC 0.5704 0.7068 0.1655 Mean abs. error (cu.ft/sec) 68.0000 171.0000 385.0000 Std.dev of error (cu.ft/sec) 139.0000 257.0000 320.0000 Entropy of error PDF (nats) 1.2700 2.3700 2.9200 NSC : Nash Sutcliffe coefficient ranges from Table 3 6. Performance metrics for N2 under 50% Training Data Metric 2321000 2321500 2322500 NSC 0.5482 0.4939 1.0850 Mean abs. error (cu.ft/sec) 60.0000 183.0000 676.0000 Std.dev of error (cu.ft/sec) 147.0000 363.0000 414.0000 Entropy of error PDF (nats) 1.0800 2.4800 3.4600 NSC : Nash Sutcliffe coefficient ranges fr om A B Figure 31. Ba sic node structures in BNs. A) C onvergent connection B ) S erial connection PAGE 59 59 Figure 32. Sample network representation of a BN for embedding spatiotemporal data Figure 33. Bayesian networks for streamflow e stimation given spatio temporal data sources PAGE 60 60 A 100 200 300 400 500 600 700 0 500 1000 1500 Days (1/7/2005 12/31/2006)Flow in cu.ft/secPredicted versus Observed flow at 2321000 Predicted Observed 100 200 300 400 500 600 700 0 0.5 1 Days (1/7/2005 12/31/2006)P(f) Posterior Probability B 100 200 300 400 500 600 700 0 500 1000 1500 2000 Days (1/7/2005 12/31/2006)Flow in cu.ft/secPredicted versus Observed flow at 2321000 Predicted Observed 100 200 300 400 500 600 700 0 0.5 1 Days (1/7/2005 12/31/2006)P(f) Posterior Probability Figure 34. Streamflow estimates at station 2321000. A) N1 estimates B ) N2 estimates PAGE 61 61 A 100 200 300 400 500 600 700 0 1000 2000 3000 Days (1/7/2005 12/31/2006)Flow in cu.ft/secPredicted versus Observed flow at 2321500 Predicted Observed 100 200 300 400 500 600 700 0 0.5 1 Days (1/7/2005 12/31/2006)P(f) Posterior Probability B 100 200 300 400 500 600 700 0 1000 2000 3000 4000 Days (1/7/2005 12/31/2006)Flow in cu.ft/secPredicted versus Observed flow at 2321500 Predicted Observed 100 200 300 400 500 600 700 0 0.5 1 Days (1/7/2005 12/31/2006)P(f) Posterior Probability Figure 35. Streamflow estimates at station 2321500. A) N1 estimates. B) N2 estimates PAGE 62 62 A 100 200 300 400 500 600 700 2000 4000 6000 Days (1/7/2005 12/31/2006)Flow in cu.ft/secPredicted versus Observed flow at 2322500 Predicted Observed 100 200 300 400 500 600 700 0 0.5 1 Days (1/7/2005 12/31/2006)P(f) Posterior Probability B 100 200 300 400 500 600 700 2000 4000 6000 Days (1/7/2005 12/31/2006)Flow in cu.ft/secPredicted versus Observed flow at 2322500 Predicted Observed 100 200 300 400 500 600 700 0 0.5 1 Days (1/7/2005 12/31/2006)P(f) Posterior Probability Figure 36. St reamflow estimates at station 2322500. A) N1 estimates. B) N2 estimates PAGE 63 63 A 0 100 200 300 400 500 600 700 800 0 200 400 600 800 1000 1200 1400 1600 Actual error vs Expected error: 2321000 DaysError in cu.ft/sec Actual Expected B 0 200 400 600 800 1000 1200 1400 1600 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2 Actual error pdf vs Expected error pdf: 2321000 Error in cu.ft/secP(error) Actual Expected Figure 37. Model validation at 2321000. A) E stimation error s. B) PDF of errors PAGE 64 64 A 0 100 200 300 400 500 600 700 800 0 200 400 600 800 1000 1200 1400 1600 1800 Actual error vs Expected error: 2321500 DaysError in cu.ft/sec Actual Expected B 0 200 400 600 800 1000 1200 1400 1600 1800 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 Actual error pdf vs Expected error pdf: 2321500 Error in cu.ft/secP(error) Actual Expected Figure 38. Model validation at 2321500. A) E stimation error s. B) PDF of errors PAGE 65 65 A 0 100 200 300 400 500 600 700 800 0 1000 2000 3000 4000 5000 6000 Actual error vs Expected error: 2322500 DaysError in cu.ft/sec Actual Expected B 0 1000 2000 3000 4000 5000 6000 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 Actual error pdf vs Expected error pdf: 2322500 Error in cu.ft/secP(error) Actual Expected F igure 39. Model validation at 2322500. A) E stimation error s. B) PDF of errors PAGE 66 66 Figure 310. Streamflow estimation at intermediate points Figure 311. Streamflow and MAP estimates along the Santa Fe river on October 4th, 2006. 100 200 300 400 500 600 700 0 2000 4000 Observed vs Predicted Flow: 2321975Days (1/7/2005 12/31/2006)Flow in cu.ft/sec Observed Predicted Figure 312. Obs erved versus predicted streamflow at USGS station 2321975. PAGE 67 67 A 0 100 200 300 400 500 600 700 800 0 500 1000 1500 2000 2500 3000 3500 4000 4500 Actual error vs Expected error: 2321975 Days (1/7/2005 12/31/2006) Error in cu.ft/sec Actual Error Expected Error B 0 20 40 60 80 100 120 140 0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1 Actual error pdf vs Expected error pdf: 2321975 Error in cu.ft/sec P(error) Actual Error Expected Error Figure 313. Model validation at 2321975. A) E stimation error s. B) PDF of errors PAGE 68 68 CHAPTER 4 HARDWARE ACCELERATIO N ON FPGAS 41 Pattern based Algorithm Decomposition An emerging method for accelerating algorithm execution involves the use of RC based on FPGA technology. Designers who have achieved success in attaining orders of magnitude increase in performance (e.g., speed and power) through hardware acceleration of algorithms are in a sel ect group and are typically highly skilled at exploiting FPGA based systems and related tools (e.g., hardware description languages (HDLs), logic design, and performance prediction tools). To significantly increase the productivity of FPGA design and deve lopment in general, we must improve the productivity of designers who are not FPGA experts. High level language tools, popularly known as application mappers (e.g., Impulse C, AccelDSP, Carte C, C2H), translate high level language code to HDL versions, th ereby improving productivity. Unfortunately, performance in such scenarios is highly dependent on the mapper's efficiency in extracting parallelism from the algorithm structure. The mapper also dictates algorithm decomposition for a parallel implementati on, restricting the designers freedom to explore different decomposition strategies Another way of improving design productivity is to reuse and leverage existing hardware designs. One of the primary motivations behind this work is to represent FPGA designs at a higher level of abstraction, thus making them more readable and therefore reusable. Design patterns [ 62] were introduced in the software engineering domain as common solutions to recurring problems and have successfully formed the basis for cr eating and exploring designs, reusing solutions for similar software applications. More recently, there have been efforts to apply pattern based design to FPGA based reconfigurable computing [ 63, 64]. In [ 55], the authors identified and cataloged an exte nsive list of design patterns that can be used to solve a PAGE 69 69 broad range of problems spanning diverse fields. By decomposing algorithms in terms of simple design patterns, a developer is able to understand the dataflow and structure behind the parallel imple mentation without delving into the details of a hardware circuit. In this work, a list of primitive design patterns have been identified and categorized in such a manner that is most appropriate for the design of machine learning algorithms used in many e stimation problems and for efficiently exploring the designs using performance prediction tools like RAT [ 53]. Two of the most important categorizations of patterns and their examples are shown in Table 41 : computation patterns and communication patterns While computation patterns deal with extracting and implementing parallelism in the algorithm, communication patterns provide different means of regulating data flow to keep the computation patterns busy. Note that the patterns listed in Table 41 are primitive patterns. A variety of composite patterns can then be constructed from these primitive patterns to solve potentially any complex problem. For example, a wavefront pattern [ 55] can be constructed using a datapath replication pattern with each pa rallel datapath comprising of pipeline patterns of different lengths. 411 Pattern Description The pipeline and datapath replication patterns are described in this section using a standard format suggested in [ 55]. These two primitive patterns are highly important in that any composite pattern for the purpose of expressing and implementing parallelism can be constructed from a combination of pipeline and datapath replication patterns. 4111 Pipeline p attern INTENT. Used for implementing programmin g structures wherein successive/intermediate instructions overlap in execution. The instruction throughput is increased by executing a number of instructions per unit of time resulting in a lesser number of net clock cycles spent during processing. MOTIVA TION. The primary motivation is to achieve faster processing speeds. When a program runs on a processor, the basic assumption is that each instruction in the program is PAGE 70 70 executed before the execution of a subsequent instruction is begun. However, in many cases the instruction throughput could be increased if the program allows the possibility to execute more instructions per unit of time. Each instruction is computed and linked into a chain so each instructions output is an input to a subsequent instru ction. APPLICABILITY. This pattern is intended for program structures where different instructions in the program execution could operate over different data concurrently. PARTICIPANTS. A communication pattern regulates dataflow; partial/intermediate resul ts may need to be accumulated or buffered for subsequent usage. COLLABORATIONS. A communication pattern regulates dataflow and keeps data fed to the pipeline. CONSEQUENCES. o When all instructions in the program execution are not independent, the pipeline co ntrol logic (i.e. the controller) must insert a stall or wasted clock cycle into the pipeline until the dependency is resolved. o While pipelining can theoretically increase performance over a non pipelined core by a factor of the number of stages (assuming the clock frequency also scales with the number of stages), in reality, design limitation will not always lead to ideal execution. o The instruction latency in a non pipelined design is slightly lower than in a pipelined equivalent. This is due to the fact that extra flip flops must be added to the data path of a pipelined design. KNOWN USES. Any algorithm/program that has independent instructions in its structure (e.g., PDF estimation, Correlation). 4112 Datapath r eplication p attern INTENT. Used for ex ploiting computational parallelism observed in sequential programming structures such as computational loops. MOTIVATION. The primary motivation is to achieve faster processing speeds. Parallelism observed in computations can be efficiently exploited usin g custom logic. Computational structures such as loops iterate over the same set of processing instructions multiple times. In case of zero data dependencies between iterations (or limited data dependency), the processing instructions can be replicated i n hardware to perform multiple iterations of the loop in parallel. APPLICABILITY. This pattern is intended for exploiting parallelism in computational structures such as loops. An important consideration for implementing the pattern is data dependencies b etween loop iterations. Depending on the individual case, the PAGE 71 71 implementation may need intermediate buffers or communication links between parallel implementations of computational kernels. PARTICIPANTS. Apart from the replicated kernels, a communication p attern regulates dataflow with possible requirements of additional buffers. COLLABORATIONS. Communication patterns regulating dataflow and the total number of iterations required on the kernel instances implemented in hardware. CONSEQUENCES. o Since the patt ern implements an area time tradeoff, higher processing speeds achieved via parallel instances of the kernel come at the cost of increased implementation footprint in hardware. o Additional overhead in terms of parallel data communication and control logic a re present. Depending on the FPGA platform and implementation, the problem may be communication bound, limiting the number of parallel kernels that can be fed with data in parallel. KNOWN USES. PDF estimation, Molecular dynamics, K Means clustering, Sorti ng algorithms. 4113 Loop f usion p attern INTENT. Used to alleviate potential pipeline stalls for nested loops in hardware designs particularly for automated pipelining HLL compilers. It a lso encompasses the combining of separate loops into a single loop for potential optimization MOTIVATION. The primary motivation is to remove pipeline stalls from computational blocks. Nest ed loops are a common feature of computations but can create pipeline stalls in between iterations of the inner most loop. Automa te d compilers can quickly create optimized designs for inner most loop computations. Consequently, fusing loops will allow the automated compiler to pipeline the entire computational block APPLICABILITY. This pattern is primarily intended for nested loops. It also applies to combining separate loops to reduce the overall number of interactions, potentially at the expense of a larger memory footprint. Reducing overall cycle count (through alleviating pipeline stalls) is the goal of loop fusion PARTICIPAN TS. A nested loop is fused into a single loop. Alternatively, multiple independent loops are combined into a single loop. Additional logic is potentially added to the loop to retain n dimensional indexing in a flat, single dimensional loop COLLABORATION S. A c ontrol ler manages the dataflow to keep feeding data in parallel to the implemented kernels. It also manages the total number of iterations required on the number of kernel instances implemented in hardware PAGE 72 72 CONSEQUENCES. o Additional logic for indexing is required to support nested loop fusion. Nave implementations of loop fusion can require computationally expensive operations such as division and modulo arithmetic. Other implementations require counters whose values persist between loops. Without extra compiler support, such implementations may not be optimal. o Independent loop fusion may not have substantial benefit if loop overheads can be overlapped with computation KNOWN USES. Molecular Dynamics, N body problem 412 Quantifying Design Patter ns Since design patterns provide a formalized way to represent algorithm decomposition and the associated communication for parallel implementation, they can be quantified and used in analytical models like RAT to predict the algorithms amenability to har dware acceleration (to be described in Section 421), before undertaking a lengthy (and possibly fruitless) development process. T he impact of each design pattern can be analyzed in terms of throughput and/or overhead contribution to design performance. In this work, throughput of a design pattern is defined as the maximum number of operations that it can perform per clock cycle. Latency is defined as the number of cycles spent in data transfers before any useful computation is performed. For example, a pipeline pattern has a throughput equal to the number of pipeline stages and contributes an equal cycle count on latency (i.e. clock cycles spent in filling up the pipeline stages). The throughput latency and graphical notation of patterns relevant t o this work are given in Table 42. Having decomposed the algorithm as a composite of primitive patterns, the net throughput and latency can be inferred from the structure and behavior of the constructed composite pattern. PAGE 73 73 42 Scalable and Portable Arch itecture for PDF Estimation Having investigated and analyzed the commonalities shared by the three case study algorithms in this work (Section 244) we suggest that an efficient FPGA design developed for one of the algorithms, say PDF estimation, can be e ffectively reused for computing the others. For this reason, an in depth explanation of the design and performance evaluation of a suitable architecture for only the PDF estimation algorithm will be given in the following sub sections. Then, in Section 43, we will discuss how the architecture developed for PDF estimation can be effectively adapted by reusing most of its design for developing other similar algorithms. The general architecture of the 1 D PDF estimation algorithm based on pattern based de composition is shown in Figure 4 1. A computational kernel updates the PDF value at a particular point x (i.e. a bin) based on the value of a particular data sample xi (where xi 1 and x 2). The Parzen window technique is an embarrassingly parallel algorithm where the PDF values at multiple points can be evaluated at the same time by replicating the computational kernel (datapath replication pattern). The kernels in the pa rallel datapath are seeded with different values in I2 via a scatter communication pattern. Thus, data samples can be broadcasted across kernels that can then process data independently in parallel. Each kernel can also benefit from a pipelined implementation while performing all the operations (Eq uation 2 6) required on every data sample (pipeline pattern). The PDF values are computed in ( I1,I2), accumulated in p( I2) and stored in output memory O Load balancing and reduction of communication and synchronization requirements are necessary to ensure that hardware outperforms its sequential software counterpart. In estimating multi dimens ional PDFs, compu tation of the kernel functions in each dimension are independent of the others and hence performed in a parallel fashion within each pipeline. Internal registering for each bin keeps a running total of all processed data; these cumulati ve totals comprise the final estimation of the PDF function. PAGE 74 74 The development stages in the syste m design are highlighted in Figure 42 and explained briefly in the following bullets for a 1 D PDF design. The same can be extended for higher dimensional P DF estimation with suitable modifications made to the computational kernel. PERFORMANCE PREDICTION. Based on a preliminary design of the algorithm (in the form of design patterns shown in Figure 41A ) and the basic resources available for a selected FPGA based computing platform (in the form of parameter values to quantify the design patterns), the attainable speedups are predicted using RAT. Numerical analysis i s conducted and a suitable fixed point implementation is chosen because probability values lie between 0 and 1, negating the need for the dynamic range features of a floating point format. Details of RATs performance prediction are provided in Section 421. KERNEL AND CORE DESIG N. The basic task of a kernel in the 1 D case is the computation of the kernel function = 1 ( xix )2. A core contains a number of kernels ( k ) with each kernel in the design performing the aforementioned computation for a different x (i.e. different seed ). The computation increases in complexity while estimating higher dimensional PDFs (Section 423). The parameter k is chosen based upon the preliminary resource analyses performed earlier. Details concerning the kernel and core designs are given in Section 422. TEST BENCH AND SIMULA TION. Memory instantiations for x and xi are made, test bench files are generated, and functional simulations are performed to validate the design. OVERALL SYSTEM DESIGN, VERIFICATION, AND VISUALIZATION. Integration of the core with the host processor (middleware design) is developed foll owed by verification of the computed PDF. 421 Performance Prediction Although FPGAs have much to offer in terms of flexibility and performance, they are not amenable for all algorithms. RAT [ 53] is a simple methodology developed for predicting the perfo rmance of a specific algorithm design on a specific platform. By parameterizing a particular design strategy and a platform selection into RAT, the developer can analyze and predict likely speedups attainable. For RAT, speedup is defined as the ratio of execution time on a relevant general purpose processor ( tGPP) to the execution time on an FPGA ( tRC). RC GPPt t speedup (4 1) PAGE 75 75 For ease of predicting speedup, the RAT analytic models take the form of a worksheet, reproduced in Table 4 3, and feature two important steps. The first step deals with estimating the communication burden ( tcomm) involved in transferring data in and out of the FPGA. The entries in the worksheet related to this step are the communication parameters ( throughputideal, write, read) and the dataset parameters. The second step involves the estimation of time spent in performing computation ( tcomp) over the data transferred ( Nelements) on the FPGA. In the original formulation of RAT, parameters Nops/element, fclock, thr oughputproc and Nelements determine computation time. The total time spent on the FPGA ( tRC) to execute the entire algorithm is calculated (see [23] for details) as given in Equation 4 2. proc clock element ops elements comp ideal write read element bytes elements write read write read comm comp comm iter RCthroughput f N N t throughput N N t t t t t t N t / / / /) ( (4 2) The memory available on the FPGA or the board hosting the FPGA is often much smaller than what is required for storing all of the application data. Niter denotes the number of iterations of communication and computation required to process all available data. The authors in [ 53] suggest t hat the parameter throughputproc be approximately chosen based upon the number of data samples that can be processed in parallel. This is not trivial to estimate in many cases and might lead to suboptimal predictions if carelessly chosen. A pattern based design embeds the parallelism ( throughput ) and the accompanying overhead ( latency ) during algorithm decomposition and would help better parameterize RAT while estimating tcomp. Latency effects PAGE 76 76 in pipelines and buffers are easily extracted from the patter ns as against manual interpretation. For example, in the PDF design (Figure 41), tcomp is computed as given in Equation 4 3. d p g p s net clock element ops elements net compT T throughput L L L Latency f throughput N N Latency t 1/ (4 3) The worksheet in Table 4 3 shows the input parameters for the RAT analysis for the 1 D and 2 D PDF estimation algorithms. The communication parameters model a Nallatech H101 PCIXM card containing a Xilinx V4LX100 user FPGA connected to a Xeon host processor over a 133MHz PCI X bus. Although the entire application involves 204,800 data samples, each i teration of the 1 D PDF estimation on the FPGA will involve only a portion of that data (512 data samples, or 1/400 of the total set) because of memory constraints on the FPGA. A corresponding input buffer size of 512 is chosen for entry in the worksheet. Since the computation is performed in a two dimensional space for a 2 D PDF, twice the number of data samples (in blocks of 512 words for each dimension) is sent to the FPGA. The number of output elements in the RAT table corresponds to the number of ou tputs for every call to the FPGA. In the 1 D case, the estimated PDF values at 256 points are returned from the FPGA after all calls to the FPGA are complete (i.e. after Niter calls). There is sufficient block RAM on the FPGA to hold the results of the 1D PDF algorithm. Since the PDF output elements need not be sent for each call, the communication time is accounted for by distributing the total number of output elements over Niter calls. From Table 4 3, it is underst ood that the FPGA is called 400 times resulting in an average value of <1 output element per call. This is round ed to the next largest integer (1 in this case). In the 2 D case, the PDF values are computed over 256 256 (i.e. PAGE 77 77 65536 in the worksheet) points and are sent back to the host for every call made to the FPGA due to insufficient block RAM to store results between calls. The computational density of the algorithm is defined in terms of Nops/element. Each data sample in the 1 D PDF case requires 3 operations at each of the 256 points at which the PDF is estimated, resulting in 768 operations. In the 2D PDF algorithm, each data sample requires 6 operations at each of the 256 256 points leading to 393216 operations. The computational throughputs for the 1 D and 2 D designs are estimated using FPGA clock frequencies of 150MHz and 100MHz, respectively. The throughput of the design is determined based on the constructed composite pattern ( Figure 41A ) 8 parallel datapaths, each comprising of a 3 stage pipeline, leads to a net t hroughput of 24 ops/cycle. Ls and Lg equal 256 (i.e., size of I2 used to seed the parallel datapaths) and Lp equals 3 leading to a net latency of 515 cycles. For the 2D PDF design, the latency Ls and Lg increase to 65536 (256 256 seeds) while the number of replicated datapaths equals 16 (i.e. 8 for each dimension). The software baseline ( tGPP) for computing speedup values was measured from an optimized C program (optimization flag was set to O3) executed on a 3.2GHz single core Xeon processor using s ingle precision floating point. The truncated Taylor series expansion was used in place of the exponential in the C program as well. RAT predictions are then estimated to determine an attainable execution time and later compared against experimentally me asured software results to compute speedup. The predicted speedup for the 1 D and 2 D PDF algorithms was 11.5 and 9.0, respectively 422 1D PDF Design A design of a 1 D PDF relies heavily on the availability of dedicated arithmetic units and memory b locks. Not only should the available resources be efficiently used but also the design should scale well with application complexity. Taking these points into account, a multi core design with a key design parameter k (the number of kernels or parallel datapaths in a core) is PAGE 78 78 proposed. The support size of the PDF is 256 along one dimension ( I2) and the number of data samples processed for every call to the FPGA is 512 ( I1). The multi core aspect of the design is developed with consideration toward futur e extensions to multi FPGA systems. Larger FPGAs would be able to house more kernels and hence process more data in parallel leading to faster computations. In the following sections, single and dual core designs are described, along with a discussion o f scalability with respect to the number of cores. 4221 Single core d esign The FPGA receives data ( I1 and I2) over an interconnect (e.g., PCI X, PCI Express, RapidIO) and the core accesses data from the FPGA memory. The basic blocks in the design a re pictorially represented in Figure 44A and the host centric execution flow is illustrated in Fig ure 45A A set of data samples ( I1) is first loaded onto the FPGA that is then signaled to start the computation. As the FPGA processes data, the host pro cessor polls for a process complete signal from the FPGA. The FPGA sends the process complete signal once the computation is finished and the host sends in the next set of data samples ( I1) until all data are processed. Data and control flows to the parallel kernels housed in the core are regulated by a finite state machine explained as follows (Fig ure 43). The set of seeds x scattered to the k parallel kernels in the core are represented by I2, the data samples xi by I1, and the PDF values by p in the state machine description. IDLE. Remain in idle state until activation signal GO=0; transition to scatter state if GO=1 SCATTER. Distribute k values of I2, p( I2) to the parallel kernels; transition to compute state. COMPUTE. Broadcast data samples I1 sequentially, compute ( I1,I2), and update p ( I2); transition to scatter state to load subsequent k values of I2, else move to g ather state. GATHER. Update p in FPGA memory; transition to done state. DONE. Send process complete signal done=1 to host; tr ansition to idle state when GO=0. PAGE 79 79 4222 Dual core d esign The architectural details and the execution flow of the dual core design are illustrated in Fig ures 44B and 45B respectively. The state machine for a single core design and the corresponding core are replicated with changes made to the data flow in the software control program at the host end. Since multiple cores share the same interconnect, arbitration for bus access is performed while loading data onto the on chip memory of the FPGA and th e subsequent polling for the process complete signal. The FPGA systems used in this work have a host processor for explicitly managing data movement in and out of the FPGA. Data is written to the first core and, while the first core processes the data, the next set of data samples is loaded to the second core to process. The host then polls the first core for the process complete signal as the second core processes its data. There could be a situation where the host cannot poll one of the cores whil e it is loading data to the other core. The core that is not being polled would sit idle during this contention period. 423 2D PDF Design In the 2 D case, the PDF is evaluated at a matrix of points or bins ( n1 rows of x values n2 columns of y value s). The number of computation s grows from ( N n)2 + c to ( N n1)2 + ( N n2)2 + c and the basic kernel computation expands to 1 (( xix )2 + ( yiy )2) where xi, yi represent one data sample. Despite the added complexity of the 2 D PDF algorithm, the incre ased quantity of parallelizable operations makes this algorithm still amenable to the RC paradigm, assuming sufficient quantities of hardware resources are available. To perform the 2 D PDF estimation, the basic architecture designed for the 1 D case is e xtended with modifications made primarily to the kernel design (Fig ure 46) and the control flow. Due to the limited amount of FPGA memory, the PDF matrix is computed one column at a time (i.e. for all x and one y represented PAGE 80 80 by p( I2 x,y )). This partial PDF matrix is returned to the host before a subsequent call to the FPGA is made. The kernels in the 2 D PDF design are initialized with two sets of seeds ( x and y ). The finite state machine regulating the data and cont rol flow is illustrated in Figure 47 and explained as follows. The set of seed values x and y scattered to the parallel kernels ( k in number) in the core are represented by I2 x and I2 y, the data samples xi and yi by I1 x and I1 y, and the PDF column values by p ( I2 x,y ) in the state machine description IDLE. Remain in idle state until activation signal GO=0; transition to scatter state if GO=1 SCATTER. Distribute y k values of I2 x, and p ( I2 x,y ) to the parallel kernels; transition to compute state. COMPUTE. Broadcast data samples I1 x and I1 y concurrently and update p( I2 x,y ); transition to scatter state to load subsequent k values of I2 x, else move to g ather state. GATHER. Update p( I2 x,y ) (i.e. column of PDF matrix) in FPGA memory; transition to done state. DONE. Set the process complete signal done=1; when GO=0 transition to scatter state if CNT( I2 y) y ), else move to idle state. 424 Experimental Results In this section, experimental results with the PDF FPGA designs are presented and analyzed. The experim ental platform used throughout these experiments contains the Nallatech H101 board with V4LX100 FPGA. This Xilinx chip has dedicated DSP48 arithmetic blocks capable of performing rapid 18bit multiplies and multiply accumulates compared to logic based MAC cores. The board communicates with the host processor over a PCI X interconnect. A 32 bit wide communication channel is used of which 9 bits were allocated for the fixed point fractional segment. Function level simulation was performed in ActiveHDL fro m Aldec, and H101 implementation was rendered using DIMEtalk from Nallatech. PAGE 81 81 4241 Speedup, p ower, and r esource u tilization One of the primary objectives in this work is to obtain speedups in execution by exploiting parallelism at the hardware level when compared to the sequential version execution on a high end CPU. The single core 1 D and 2 D PDF designs operated at FPGA clock speeds of 150MHz and 100MHz, respectively. The number of kernels, k in the core was set to 8. The basic criterion was to develop an efficient design where the resources (e.g., DSP48s, block RAMs, slices) are uniformly consumed. It should be noted that, since the algorithm is embarrassingly parallel, improved performance can be achieved by scaling the design, that is proces sing more data in parallel by increasing k until on chip resources are exhausted. Actual speedup power and resource utilization factors for the 1 D and 2 D PDF algorithms are presented in Table 44. Power was computed using the XPower analysis tool pro vided by Xilinx under a 12% activity rate. It can be seen from the comparisons made in Table 45 that the actual speedups obtained are reasonably close to the predicted speedups using RAT (originally shown in Table 4 3). The deviation between the values is primarily due to inaccuracies in the estimate of tcomm values. Although the channel efficiency values ( write, read) chosen for the RAT analysis are a reasonable assumption for large data transfers, they tend to be lower for smaller data transfers as in this scenario. The tcomp predictions are relatively close to the experimental values validating the adv antage of employing design patterns for algorithm decomposition and performance prediction. The number of DSP48s scales linearly with the number of kernels k in the core. In the 2 D PDF scenario, each kernel uses twice the number of DSP48s as was consume d in the 1 D case because the computation is conducted along two dimensions. Although the total amount of FPGA memory consumed by the PDF core is moderate, a significant percentage of it is consumed by DIMEtalks interface to buffer and transfer data to a nd from the host and FPGA. In addition to low net power consumption, two - PAGE 82 82 thirds of the total consumption was affiliated to static power (~1020mW) in both the 1D and 2 D PDF designs. This suggests that scaling the design would potentially result in small power increments while still maintaining low total power. 4242 Data verification Even though significant speedup was shown in the previous section, speedup is meaningless if the results are not accurate. In this section, we verify the accu racy of th e data provided by the FPGA designs. Data samples from a bi modal mixture of Gaussian distribution and from a two dimensional uni modal Gaussian distribution were generated in MATLAB to verify the functionality of the 1 D and 2 D PDF designs. PDF values were computed by generating 800KB ( N =204,800) of data samples ( n=256 for 1D PDF; n1= 256 and n2= 256 for 2D PDF). The computed PDF values were read and error in the solution was computed as the difference in FPGA and GPP estimates. Maximum error E of 3.8 % and 0.57% were obtained for the 1 D and 2 D algorithms respectively (Eq uation 44), which are reasonable for most real time applications that involve decision making based on probabilistic reasoning. The errors in the estimates are due to the Taylor ser ies truncation of the exponential function rather than the fixed point effects (the exponential function was not truncated in the GPP estimates for this study). The resulting PDFs shown in Figures 48 and 49 were plotted in MATLAB for verification. GPP GPP RCx p x p x p E ) ( ) ( ) ( max (4 4) 4243 Scaling to m ultiple c ores Since the PDF algorithm is embarrassingly parallel, significant speedup can be achieved by scaling the design to multiple cores. In this work, dual core architectures were developed and evalua ted to study inherent characteristics for scalability on single or multiple device systems. The dual core results for the 1 D and 2 D cases are summarized and compared to the single core PAGE 83 83 implementation in Table 46. Design frequencies of 150MHz and 70MH z were obtained for the 1D and 2 D algorithm designs, respectively. The reduction in frequency was primarily because of increased routing delays attributed to a larger design. Consumption of a greater percentage of DSP48 slices is also bound to increase routing delays as the DSP48 slices are located at specific areas across the IC. An important point to note here is that speedup does not double in a dual core design (as compared to the single core design) due to interconnect contention as discussed in S ection 4222 and a reduced frequency in the 2 D PDF case. These parameters would potentially differ from one platform to another depending upon the system interconnect bandwidth and its architecture. Also, the dual core 1 D PDF design offered a higher increase factor in speedup when compared to the dual core 2 D PDF design because time spent on data communication could be effectively hidden under time spent on computation in the 1 D PDF case. This outcome was not achievable in the 2 D PDF designs due t o longer communication times (columns of the PDF matrix are sent back to the host after every call to the FPGA) and more importantly a reduced frequency of operation. 4244 Portability c onsiderations In addition to having scalability impacts, applic ation characteristics also contribute greatly to platform selection and performance. In this work, optimum choice of the number of kernels per core, k should be made based upon the interconnect bandwidth and FPGA resources available. An important byprod uct of the RAT methodology, not mentioned earlier, is the computation and communication utilization factors ( Eq uation 45). comp comm comp comp comp comm comm commt t t util t t t util ; (4 5) PAGE 84 84 These metrics indicate the nature of a particular design in terms of whether it is communication bound or computation bound. If there are a variety of platforms to choose from, the designer could make use of these metrics to modify architecture selection for achieving optimum results. The primary resource limiting factor in increasing the value of k in this algorithm design is the number of dedicated multiplier blocks in the FPGA. While platforms with FPGAs having more multipliers (e.g., the Cray XD1 with Virtex 2 Pro FPGAs) and faster theoretical interconnects (e.g., RapidArray in XD1) might intuitive ly suggest improvements in speedup by performing more computations in the PDF algorithm in parallel, these could be negated by poor read and write efficiencies ( readwrite) during data communication. This case was particularly true with the Cray XD1 system. Due to extremely low read speeds on CPU initiated transfers from FPGA host memory (~4MB/s for small data transfers), RAT predicted a smaller speedup numbe r (Table 47) in migrating the 2 D PDF algorithm to the XD1 system even though it housed a theoretically faster interconnect. In comparison to the Nallatech platform, the utilization factors in Table 47 illustrate the fact that in the XD1 more time is sp ent in data communication as compared to algorithm computation for the 2 D PDF design. On the contrary, since the computed PDF values are sent to the host after all FPGA iterations are complete, fewer read operations are required in the 1 D PDF design res ulting in the XD1 offering more than double the speedup achieved on the Nallatech platform. 4245 Scaling to multi FPGA systems With emerging multiFPGA systems, it is also desirable to analyze the potential for scaling the PDF designs across multipl e FPGAs and potentially achieve higher speedup. The 2D PDF single core design was replicated across a cluster of Nallatech cards connected over infiniband (Figure 410). Unified Parallel C (UPC) was used to develop the system level code wherein the enti re application data is distributed equally ( i.e., shared) among nodes in the cluster. The PAGE 85 85 additional communication times involved in the distribution and collection of data along with the attained speedup for different cluster sizes (number of Nallatech c ards) is shown in Table 48. As the number of FPGAs increase, the number of iterations required to process all application data decrease. Speedup eventually drops with increasing number of nodes as time spent on gathering PDF results from nodes increases 43 Composite Patterns for Design Reuse As discussed earlier, it is often the case that most of the estimation frameworks (inclusive of the BN framework presented in this work) employ different algorithms at different stages of data analysis /processing To be highly productive in creating efficient hardware designs for each of those algorithms, it is essential that we reuse existing hardware designs. One of the goals of design patterns is to provide the framework to realize this concept. In this sectio n, the commonalities between the algorithms described in Section 244 and the similarities in the composite patterns describing their decompositions are exploited by reusing the structural and communication architecture developed for PDF estimation. Also state transition diagrams (Fig ures 43 and 47) primarily deal with managing data flow and activate the computations to execute at appropriate times. This task is particularly cumbersome to design and debug, and any means of reducing this burden would be helpful to designers. Due to similarities in algorithm decomposition, the state transition diagram developed for PDF estimation can be effectively reused for the other two algorithms 431 K M eans Clustering K means clustering can be decomposed in a similar fashion to that of PDF estimation. Computation of the Euclidean distance between every data sample and the k cluster centers can be done in parallel. The minimum distance indicating the cluster closest to the data sample can then be obtained hier archically by comparing the Euclidean distances pair wise in a parallel PAGE 86 86 fashion ( Figure 411). The basic update rule in K means clustering for the cluster to which the (samplenew centerold), where is th e learning rate [ 65 ]. In this work, the cluster centers are updated according to Equation 46, as and when a new data sample is assigned to a particular cluster 2new old newsample center center (4 6) This approximate update rule (where = 1/2) is often used for on line clustering and is applicable to real time scenarios in which fast reaction to changing statistics is often desirable. The update method reduces the computational requirements to addition and right shift operations and eliminates the more resource consuming division operation. While this offers an efficient hardware implementation, it can lead to slower convergence for broadly distributed clusters due to its relatively high sensitivity to each new sample. In such cases, the traditional approach proposed by MacQueen [ 65 ] can be used where = 1/ nj (here, nj is the number of data points in class Sj). This choice for achieves a more gradual updating of cluster centers as the number of data samples presented to algorithm grows. The design develope d in this work can be extended to employ MacQueens approach with minimal resource usage by the inclusion of a multiplier and lookup table (for storing values) to implement the update rule. All the computations involved in the algorithm can further be implemented as a pipeline. Analyzing the composite pattern for decomposing K means in Figure 4 11, it can be inferred that the architecture developed for PDF estimation can be effectively reused for K means clustering with minimal modifications as explain ed in the following section. The cluster centers ( I2) are used to seed the parallel datapaths in Figure 41A (for the PDF estimation algorithm) and the data samples ( I1) are fed into the pipelines sequentially to determine their cluster association. The number of parallel datapaths and the length of I2 are both increased from 8 (in the architecture of PDF PAGE 87 87 estimation) to 64, corresponding to the number of clusters. Upon convergence, the updated cluster centers are read back by the host. Figure 411 shows the architecture for performing K means clustering on a 1 D dataset ( N =102,400 and number of iterations for convergence of the algorithm T =100) over 64 clusters. In the first pipeline stage, the Euclidean distance between the cluster centers and a data s ample is computed. In the second stage, the cluster closest to the sample is inferred and in the third stage, the inferred cluster ce nter is updated according to Equation 46. It can be inferred that only the pipeline c omputations are modified in Figure 41B to implement the clustering algorithm while retaining most of the communication fabric. More importantly, the connection to the middleware design, which is platform specific and often a time consuming process, remains unchanged. F or the 1 D case (i. e. d=1) the architecture reduces the computational complexity of the algorithm from O(NmT) to O(NT) where m is the number of clusters. This reduction in complexity is attainable as long as there are enough multiplier resources ( m multipliers) on the FPG A. The experimental speedup and utilization for K means clustering algorithm on the Nallatech platform is shown in Table 49. The speedup obtained is considerably low for the extent of parallelism extracted in the algorithm (i.e. all operations required on a data sample are performed in one cycle). The speedup is primarily affected by the overhead time spent in transferring data from the CPU to the FPGA over a relatively small bandwidth low efficient interconnect. While only 400 iterations ( Niter in th e RAT worksheet) of FPGA communication and computation were required for PDF estimation, a total of 100 200 iterations are required for K means clustering (200 for processing all available data and 100 for algorithm convergence) leading to a considerable increase in total RC time ( tRC = Niter( tcomm+ tcomp)). The utilcomm factor for K means clustering is slightly higher in comparison to that obtained for 1D PDF estimation indicating a longer time spent in data communication PAGE 88 88 432 Correlation The archi tecture developed for PDF estimation is also a good fit for computing correlation on an FPGA. The computations involved in calculating the cross correlation value at different lags can be decomposed and mapped in a similar way as that of PDF estimation for a pa rallel implementation ( Figures 412B and 412C ) The similarity in dataflow between the two algorithms cannot easily be deciphered until we analyze a decomposition strategy for Correlation via design patterns. Consider two vectors X =[ x1, x2, x3, ,x512] and Y =[ y1,y2,y3, ,y512] whose cross correlation needs to be computed. A number of product terms involved in the computation can be performed in parallel. In particular, the xi values ( I2) seed the parallel datapaths in Fig ure 41A while the yi values ( I1) are fed sequentially in a pipeline flow. Due to resource limitations (multipliers) on the FPGA, correlation values are computed over 64 parallel datapaths (i.e., in blocks o f 64 values of xi) and accumulated until all xi values are accounted for. While accumulating values every cycle, the write address of output memory O is incremented by one to match the raster like pattern observed in computing correlation at various lags (Fig ure 412A ). Since the underlying architecture is scalable and portable (Sections 4243 and 424 4), the designs can be extended to solve larger scale problems (i.e. clustering data points into more clusters/computing correlation over longer datase ts) and across various platforms having bigger FPGAs with minimal efforts. The experimental speedup and utilization for the correlation algorithm on the Nallatech platform is shown in Table 49. PAGE 89 89 Table 4 1. List of primitive patterns and their descri ption Pattern name Description Computation Pipeline Increases throughput by overlapping execution of ind computations Datapath replication Exploits parallelism by duplicating computational kernels Loop fusion Alleviates potential pipeline stalls in ne sted loops by fusing them Tree Provides computational structure for implementing tree based algs Mesh Provides templates for implementing image based or 2 D problems Look up tables Reduces run time comp of complex operators by simpler lookup ops Commu nication Scatter Distributes data among many computational kernels Gather Collects data (results) from many computational kernels Broadcast Replicates data to many computational kernels Round robin Scatters equally sized data blocks in sequential orde r among all kernels Memory resolution Resolves memory contention by inserting pipeline stages & buffers Table 4 2. Parameters of primitive patterns Pattern Throughput (ops/cycle) Latency (cycles) Notation Pipeline Tp = no. of pipeline stages Lp = no. of pipeline stages Datapath replication Td = no. of replications N/A Broadcast N/A N/A Scatter N/A Ls = buffer length Gather N/A Lg = buffer length PAGE 90 90 Table 4 3. RAT input parameters for analysis of 1 D and 2 D PDF estimation algorithms Dataset parameter Units 1 D PDF 2 D PDF N elements input E lements 512.00 1024.00 N elements output E lements 1.00 65536.00 N bytes/element B ytes/element 4.00 4.00 Communication parameters throughput ideal MB/s 1000.00 1000.00 write 0.37 0.37 read 0.16 0.16 t comm S ec 6.0E 6 1.6E 3 Computation parameters N ops/element ops/element 768.00 393216.00 T d ops/cycle 8.00 16.00 Latency net C ycles 515.0 0 131072.00 T p ops/cycle 3.00 3.00 f clock MHz 150.00 100.00 t comp S ec 1.2E 4 4.3E 2 t GPP S ec 0.58 158.80 N iter I terations 400.00 400.00 Predicted speedup 11.50 9.00 Table 4 4. Speedup power, and utilization for single core designs Description 1 D PDF 2 D PDF DSP48s (%) 8.0 16.0 BRAMs (%) 12.0 15.0 Slices (%) 11.0 13.0 Power ( mW ) 1552.0 1595.0 Actual speedup 7.8 7.1 Table 4 5. Performance factors for single core designs Description Predicted Predicted Actual Actual Algorithm 1 D PDF 2 D PDF 1 D PDF 2 D PDF t comm (sec) 6.0E 6 1.6E 3 2.5E 5 1.1E 2 t comp (sec) 1.2E 4 4.3E 2 1.4E 4 4.4E 2 t RC (sec) 5.1E 2 1.7E+1 7.4E 2 2.2 E+1 Actual speedup 11.5 9.0 7.8 7.1 PAGE 91 91 Table 4 6. Speedup and utilization for dual core designs Description Single core Single core Dual core Dual core Algorithm 1 D PDF 2 D PDF 1 D PDF 2 D PDF DSP48s (%) 8 16.0 16.0 33.0 BRAMs (%) 12.0 15.0 15.0 21 .0 Slices (%) 11.0 13.0 16.0 22.0 Actual speedup 7.8 7.1 13.4 8.3 Table 4 7. Speedup and utilization for single core designs across platforms Description Nallatech Nallatech Cray XD1 Cray XD1 Algorithm 1 D PDF 2 D PDF 1 D PDF 2 D PDF util comm. 0.14 0.20 0.03 0.38 util comp 0.86 0.80 0.97 0.62 RAT predicted speedup 11.50 9.00 13.00 3.70 Actual speedup 7.80 7.10 13.40 8.30 Table 4 8. Speedup for single core 2D PDF design across multiple FPGAs No. of nodes 2 4 6 8 10 t scatter (sec ) 0.176 0.177 0.183 0.174 0.175 t gather (sec) 0.950 2.800 4.667 6.521 8.383 N iter 200.000 100.000 67.000 50.000 40.000 t RC (sec) 16.140 10.400 9.750 10.340 11.420 Speedup 9.900 15.300 16.300 15.400 13.900 Table 4 9. Speedup and utilization for PDF estimation, K means clustering, and Correlation (single core designs) Description 1 D PDF 2 D PDF K means Correlation DSP48s (%) 8.00 16.00 66.00 66.00 BRAMs (%) 12.00 15.00 9.00 11.00 Slices (%) 11.00 13.00 9.00 11.00 util comm. 0.14 0.20 0.18 0.21 RAT predicted speedup 11.50 9.00 4.30 9.60 Actual speedup 7.80 7.10 3.20 8.10 PAGE 92 92 A B Figure 41. Pattern based analysis of 1D PDF estimation algorithm. A ) D ecomposition B ) D esign Figure 42. Development stages in system design PAGE 93 93 Figure 43. State transition diagram for a single core 1 D PDF design A B Figure 44. Design details of 1 D PDF. A) S ingle core design. B) D ual core design PAGE 94 94 A B Figure 45. Host ce ntric execution flows. A) S ingle core design. B) Dual core design Figure 46. Architecture design for 2D PDF estimation algorithm Figure 47. State transition diagram for a single core 2 D PDF design PAGE 95 95 Figure 48. 1D PDF estimates on GPP and FPGA. Figure 49. 2D PDF estimates on GPP and FPGA. Figure 410. Multi FPGA system configuration PAGE 96 96 Figure 411. Architecture for K means clustering PAGE 97 97 A B C Figure 412. Decomposition and design of correlation. A) C omputation structure B) P attern based decomposition C) D esign of correlation PAGE 98 98 CHAPTER 5 MULTISCALE ESTIMATIO N FRAMEWORK While not strictly part of the BN or RC accelerated work discussed so far, we take a brief look at the problem of observations at multiple resolutions. T he problem of streamflow estimation is further exacerbated when measurements (image data in particular) are made at a wide range of scales. In such cases, one would want to get all the features to a common spatial resolution (through a downsample/upsample operation). Most often, this is a scale at which the spatio temporal features are measured (e.g. the NexRad rainfall estimates) Surface morphology (elevation, slope, and curvature) and land cover (vegetation type and density) strongly influence the dist ribution of precipitated water among the major mechanisms of infiltration, increased soil moisture, evaporation, and runoff ( Figure 51). Precipitation, topography and surface morphology measurements could be embedded into a state space model for estimati ng various hydrologic boundary conditions across scales. By understanding multiscale variations, fine scale features can be estimated from sparse multisource measurements observed at multiple scales This could prove beneficial for the integrated study of streamflow from small sub basins (catchments) at urban scales all the way up to the larger watersheds in which the small catchments reside. Erosion rates and information regarding flooding/land slide risk are affected by shapes of stream banks and hill slopes. These features are often small in spatial scale (meter to sub meter) and so require fine scale measurements over large areas. Yet, when estimating the long term flow behavior at the outlet of an entire watershed, the larger scale topography that defines contributing area is more important As a proof of concept, estimation of topography at multiple scales is chosen as a case study in this work. The elevations of many natural surfaces, including the topography of the Earth, have long been recog nized to approximate self affine fractals [6 6, 67]. A statistically self affine fractal PAGE 99 99 surface can be thought of as a fractal surface that is not isotropic. In a two dimensional xy space, a self affine fractal f(x,y) is statistically similar to rHf(rx,r y) where r is a scaling factor and H is the Hausdorff measure. Such surfaces can be modeled as fractional Brownian motions (fBm) since their power spectra exhibit power law dependence on the wave number of spatial frequency [ 68, 69, 1, 70]. Multiscale K alman Smoothers (MKS) with embedded fractional Brownian motion stochastic models have been used previously to estimate topography [1] from multi resolution topographic data However, the standard MKS algorithm with a single embedded stochastic model is no t sufficient to incorporate spatial variations in the elevation statistics; for example, a rough undulating terrain yields an elevation surface with a shorter correlation length than a flat smooth terrain and hence a different power spectrum as well. In [ 68], the author emphasizes the need for different terrain descriptors for different terrain types and an efficient multiscale representation. Although the 1/ f assumption tends to be a reasonable approximation over large areas (>10 km2), it is often a poor approximation over smaller areas (<10 km2), especially if vegetation is present and modulates the topographic measurements The inclusion of multiple models in the MKS estimation framework has been employed in this work and is found to give better estim ates of topography over non stationary terrain than the standard MKS algorithm [ 71]. A mixture of experts (MOE) network developed in [ 72] is employed to weigh the individual model outputs based on the Kalman filter residuals (known as the innovations) F ractal theory concepts are employed for estimating the parameters of the models by finding the fractal dimension of the image over different regions. Fractal based measures have been previously used in classification and segmentation algorithms [ 69, 70, 73, 74]. In particular, the relation between fractal dimension (related to Hurst exponent) and roughness has been verified in [ 70, 75] through psychophysical studies. Authors in [ 69, 76] PAGE 100 100 propose the use of multiple hurst parameters for representing topogr aphic like datasets deviating from the 1/f property. This work not only acknowledges the need for multiple models to effectively represent topography but also incorporates such a model in an MKS framework that further handles uncertainties in the observat ions and hence aids in better segmentation of terrain at multiple scales. Although the estimates of topography are improved, the primary motivation for employing a multiple model MKS (MMMKS) is for segmentation of terrain images based solely on spatial v ariability of the elevation surface. This is important when no other information (e.g., multispectral) is available. The importance of multiscale terrain analysis in remote sensing applications is discussed in [ 77 ] 51 Test Area The study area analyze d for this work is a mix of coniferous and deciduous forest in North central Florida, USA. The forest is part of the Hogtown Greenway within the city of Gainesville, so urban development (roads and houses) appears along the forest edge, as shown in Figure 52. The study site was imaged with an ALSM sensor owned by the University of Florida (UF) from an above ground altitude of 600m. The last stop returns represent penetration on targets that have sparse structure at the scale of a few tens of centimeters and are used for this analysis. The 3D point cloud obtained is gridded into different spatial resolutions to produce images of varying resolutions. 52 Multiscale Estimation Framework The multiscale models of focus in this work were proposed in [1] and p rovide a scale recursive framework for estimating topographies at multiple resolutions. These models are defined on index sets and organized as multi level quadtrees. M ultiscale estimation is initiated with a fine to coarse sweep up the quadtree that is analogous to Kalman filtering with an added merge step. The fine to coarse sweep up the quadtree is followed by a coarse to fine sweep PAGE 101 101 down the quadtree that corresponds to Kalman smoothing. T his algorithm is referred to as MKS The linear coarse to fin e model in [1] is given in Equation 5 1 where s is the node index on the tree, and B is the backshift operator in scale such that Bs is one scale coarser than s ) ( ) ( ) ( ) ( ) ( ) ( ) ( )( ) ( s v s x s H s y s w s B Bs x s A s x (5 1) Here, x ( s ) is the state variable (elevations), and y ( s ) represen ts the ALSM observations. Random process w ( s ) is assumed to be a Gaussian white noise process with identity variance, and the measurement error v ( s ) is a Gaussian white noise process with variance R ( s ). A ( s ) is the coarse to fine state transition operato r, B ( s ) is the coarse to fine process noise standard deviation and H ( s ) is the measurement state relation. The scale is represented by the level in the quadtree, and is denoted by m The support of the image at level m is 2m 2m. The standard Kalman formu lation provides optimal estimates (in the mean squared sense) when there is perfect a priori knowledge of the state and measurement model parameters A ( s ), B ( s ), and R ( s ). The state process is assumed to follow a 1/ f model. Using this model, the power spe ctrum of the state variable x ( s ) can be represented by the multiscale model in Equation 5 1 by using A ( s ) = 1 and B ( s )= B02(1 ) m /2 [1]. The MKS approach provides reasonable estimates for the evolution of the state process in scale. However, the resulting coarse to fine process noise variance B2( s ) is constant at each scale; hence, it cannot accommodate non stationarities in th e imagery at a particular scale. The fine to coarse process noise variance Q ( s ) used in the Kalman filter is a function of B2( s ) and is therefore also spatially uniform. An image that contains more than one type of topography is an example of a non stati onary 2 D process. Therefore, using a data model that is variable in scale but uniform in space will lead to suboptimal estimates in general. PAGE 102 102 In this work, fractal dimension ( D ) wa s used to discriminate different classes of terrain based on topographic var iation. Spectral matching methods have been widely employed to determine the parameters for a single model but do not give accurate results when applied to multiple models. Fractals, however, provide a good description of natural objects exhibiting statistical self similarity. In this work, the fractal dimension was calculated using a 9m 9m window by the Triangular Prism Surface Area (TPSA) method [ 78 ]. In order to determine the natural portioning of the ALSM height data, the point cloud data were gri dded into an image array with a spatial resolution of 2m 2m. Small fractal dimension implies smooth areas as can be seen over the road and roof tops. A large fractal dimension characterizes rough areas and is seen over forest regions. The number of fr actal models and their respective fractal dimension are calculated by employing the Expectation Maximization (EM) algorithm on the histogram of the fractal dimension image. The mean value of the Gaussians would then give the fractal dimension of the under lying terrain. The algorithm, when applied to the data resulted in three different Gaussians (clearly following the tri modal distribution of D ). The histogram plot (Fig ure 53) of a dataset generated from the resulting Mixture of Gaussians is very simila r to that of the fractal dimension plot. The fractal dimension D is related to the Hurst exponent H through H =3 D for two dimensional data. H in turn can be related to the slope parameter of the power spectrum density through =2 H + 1 [ 79]. By virtue of using the Multiscale structure and fBm process, the incremental variances between scales are expressed as given in Equation 5 2 where W represents the windowed data used in the sample varia nce calculation 2 4 16 2111 1 ... 222HHHVarW (5 2) PAGE 103 103 The total variance of the models at the 512 512 support level was set equal to the sample variance computed for different regions in the ALSM elevation data. The variance scale factor is solved from E quation 5 2. This can then be used to find B ( s ) at different scales from [1,(1/2)H,(1/2)2 H,,(1/2)7 H] and the value of B0 can be interpolated by fitting a linear line in log space. 53 Mixture of Experts (MOE) The MOE used in this work is a single layer network that consists of multiple experts and a gating network that arbitrates among the expert estimators. The MOE assigns weights to the experts in an unsupervised fashion as they compete for the desired response. In this work, each expert system is a multiscale Kalman filter embedded with its own 1/ f model for a particular class of topography (parameter vector i) as shown in Fig ure 54. The most likely model among these filters for a particular input is assigned a higher weight gi by the gating network. The input to such a network is the observation data set and the desired outputs are the elevation estimates of the topography. The weights can be interpreted as a priori probabilities for the corresponding experts for the current input. The softmax transformation (Equation 5 3) forces the condition on the weights to sum to one 1 0 N j u u ij ie e g (5 3) H ere, ui=zTai and ai is the weight matrix of the ith filter in the modular network. The filter residuals (innovations) rk are used to calculate the Gaussian conditional probability f of the measured observations at time k given the filter realization and the measurements up to time k 1 (Equation 5 4) where Wk = HkPk -Hk T + Rk and rk= zkHkxk [1] k k T kr W r k k i ke W z z f15 0 5 0 1| | 2 ) | ( (5 4) PAGE 104 104 The probability distribution of the filter bank is given as the weighted sum of the individual conditional distributions of the filte rs (Equation 5 5) N i i ik kg z f z f1) | ( ) ( (5 5) Equation 55 can be considered as a likelihood function, which upon maximization would give the best possible estimates. The a posteriori probabilities hi are defined in Equation 5 6. 1 1) | ( ) | ( N j j j k i i k ig z f g z f h (5 6) The weight matrix ai is updated by maximizing Equation 5 6 using a gradient ascent procedure, yielding an update ai=ai ( higi) zk, where is a learning rate parameter. In the case of multiscale image fusion, the filter recursion steps represent di fferent scales and therefore cannot exceed the base 2 logarithm of the finestscale image support. In traditional data fusion applications, observations are present generally only at two or three scales, which limit the number of iterations that the gatin g network has to estimate weights. The limit ed number of iterations requires an adaptive estimation algorithm that can react quickly. The MOE is able to react quickly due to the geometrically decreasing impact of previous measurements. The effect is fur ther alleviated by gridding the ALSM point data to different resolutions hence providing observation sets at more scales. 54 Results and Analyses Simulated data was first used to illustrate the performance benefits of employing multiple models in the MKS framework. Fractals are often used for simulating 1/ f processes and natural terrain. Figure 55 shows one such simulated fractal image embedded with two different terrains. The mean square d error (MSE) obtained from a multiple model MKS embedded with two models is lower than that obtained by either of the two models alone (Table 51). Figure 56 PAGE 105 105 shows the weights assigned to each of the two models. The elevation values over the smooth terrain are significantly small and lead to very low residuals in both filter estimates. The rough filter hence still receives moderately high weights over the smooth terrain In this work, real ALSM data were gridded to populate the quadtree at scales m ={10,9,8,7,6,5,4,3,2,1}. Interestingly, it was observed that real data does not follow a 1/ f process as much as the simulated fractal terrain. Incremental variances and empirical F ( s ) values calculated from the dataset across scales illustrate this fact (Table 52). Canopy penetration and presence of tall vegetation in the dataset causes the data collected over large areas (>100s of km2) to deviate from the 1/f property. Embedded 1/ f models are no longer well matched to the state process and hence there is scope for further refinements to the proposed MM MKS approach. Though the measurement noise covariance R ( s ), by definition, embeds only the uncertainty in the measuring device, the interception of laser light by the vegetation can be thought of as a source of measurement noise and is embedded into the model parameter R ( s ). The covariance value is calculated on a per pixel basis by computing the variance of the points that fall under each grid cell forming the pixel. In this way, the estimators are still modeled under a 1/ f assumption with the deviations taken care o f by a spatially varying R ( s ). Ideally, MKS filters should embed non 1/ f models to account for the deviation from 1/ f behavior, but following the above procedure gives acceptable solutions. Proposing a non 1/ f modeling approach requires more elaborate te sting and is considered for future research work. The state transition operator F ( s ) in the upward sweep of the MKS, defined as the ratio of the state covariance at two consecutive scales (with A ( s )=1) [1], regulates the estimation of the prior value of th e state at each scale based on the embedded model. The variations in the prior PAGE 106 106 estimates of the filters are hence significantly affected by its value and indirectly by the initialization of the root node covariance P0. ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) (2s B s A s P s A s P where s P s P s FT s s s s (5 7) For la rge values of P0, the effects of stochastic details B ( s ) embedded in the models are diminished and the filters give similar estimates and are assigned similar weights at almost all the scales. Limited recursions in scale imply that the weights have little chance of convergence. For small values of P0, the estimates of the smooth filter deviate significantly from the observation (due to smaller Kalman gain K ) and the rough filter is preferred by the MOE at finer scales (Figure 57). Knowledge of the increm ental variances across scales gives an intuitive way of initializing the root node state covariance. By virtue of the multiscale structure, the extrapolated sample variance at scale 1 gives the expected root node state covariance ( P0= B0 2). The Kalman mode l parameters were specified before the multiple model MKS estimator was implemented. The fractional Brownian stochastic model dictates that A ( s )=1. The value for the mapping matrix H ( s ) is unity in grid cells containing ALSM data and zero in empty grid c ells. For quadtree levels containing observations, the measurement noise variance R ( s ) was specified empirically for each scale. Corresponding to the tri modal nature of the fractal dimension values, a three model filter bank was chosen. One model repre sents low variance (smooth) terrain, another represents high variance terrain and land cover, and the third is tuned to track the intermediate terrain. Parts of the forest encompassing sudden changes from tall trees to bare ground (forest gaps) exemplifie d the highly varying region (variance=52.56). Road surfaces represented the other extreme (variance=3.5). Forest edges represent the third category (variance=25). These variances are then used to find the value of (Equation 5 2) using the corresponding Hurst exponents calculated using segmentation for road, forest and edges. The PAGE 107 107 stochastic detail B ( s ) terms were determined as detailed in Section 52. The estimation algorithm starts with the upward sweep wherein the estimates calculated by the individual filters are employed to compute the weights using MOE. The filter and smoothers then continue to run independently of the weighing network. The weights are then used to fuse the Kalman filter estimates in the K alman smoothing downward sweep. The presence of observations at multiple scales provides several recursion steps for the model weights to adapt. Average weights assigned to each model for different terrain and land cover regimes are listed in Table 53. Fig ure 58 shows the weights assigned to each of the three models. Notice that Model 1 receives high weights over the smooth road and relatively low weights over the forest and edges with higher fractal dimension. The converse is true for Model 2. Mode l 3 reacts to forest edges and the less rough areas. The relative weighting remains consistent across scales. All a priori (scale 10) model weights were 0.33 Table 5 1. MSE obtained over the simulated image Estimates MSE (sq.m) %MM MKS improvement MM MKS 7.1999E 4 N/A Rough filter 7.6675E 4 6.1 Smooth filter 8.3828E 4 14.1 Table 5 2. Incremental variance across scales and empirical F ( s ) values Variance between scales Variance Empirical F(s) 11 10 9.5621 0.7661 10 9 8.2091 0.8133 9 8 4.2015 0 .8825 8 7 4.8402 0.8466 7 6 4.9543 0.8146 6 5 5.0405 0.7684 PAGE 108 108 Table 5 3. Average model weights obtained using MM MKS for real data Description Scale 9 Scale 8 Scale 7 Scale 6 Road Model 1 0.4064 0.5619 0.5791 0.6055 Model 2 0.2423 0.1660 0.157 8 0.1210 Model 3 0.3513 0.2721 0.2631 0.2735 Highest fractal dimension forest Model 1 0.0977 0.0517 0.0165 0.0045 Model 2 0.4555 0.5304 0.8123 0.9135 Model 3 0.4468 0.4099 0.1712 0.0820 Intermediate fractal dimension forest Model 1 0.3430 0. 3559 0.2649 0.1658 Model 2 0.2891 0.2580 0.3046 0.3175 Model 3 0.3679 0.3861 0.4306 0.5167 Model 1, 2, and 3 represent smooth, rough, and intermediate terrains respectively Figure 51. Forest hydrologic cycle processes [ P ike 2003] PAGE 109 109 Figure 52. Elevation image created by gridding last stop elevation points using an inverse distance weighting of study area. Note: Axis and colorbar values in meters. Figure 53. Mixture of Gaussians fitted over the histogram plot of fractal dimension Figure 54. Fusion of Kalman filter bank estimates using Mixture of experts network PAGE 110 110 Figure 55. Fine scale simulated image with two terrain roughness Figure 56. Weights obtained for smooth and rough filter s at scales m=7 and 6 respectively PAGE 111 111 A B C D Figur e 5 7. Filter weights under large and small P0. A) S mooth filter under large P0. B) R ough filter under large P0. C) S mooth filter under small P0. D) R ough filter under small P0. A B Figure 58. Weights assigned to Model 1, 2, and 3 at various scales. A ) S cale 9 B) S cale 7 PAGE 112 112 CHAPTER 6 CONCLUSIONS AND FUTU RE WORK The primary motivation for this work was to develop a practical framework for estimating streamflow in a com plex system (i.e. a watershed) monitored by an ad hoc collection of measurement typ es. A highly scalable BN framework was proposed that allows the fusion of many disparate forms of data that are typically used to characterize hydrologic processes, including remotely sensed images from multiple sensors, GIS layers, point source time seri es, and spatiotemporal data, such as meteorological data acquired from multiple groundbased sensors. An information theoretic methodology, based on conditional entropy was also presented to help quantify the impact of adding nodes to the network in term s of information gain. Posterior probabilities of streamflow estimates and the associated entropy values provide d valuable information in quantifying network performance and also offer ed directions for future watershed instrumentation. The proposed methodology can also quantitatively suggest the most informative features for estimation at any location in the watershed for which there exists feature data and potentially reduce the number of features needed to achieve a specific performance goal. While the network achieves improved performance by exploiting calibrated physical model estimates like WAM, it also offers the possibility to account for inherent non stationarities in space and time by incorporating both spatial and temporal input features in the absence of WAM and achieve comparable performances. The framework is spatially explicit such that it can be applied to any arbitrary location along the reach of a stream or river The networks produce d accurate predictions in the confined region (stations 2321000 and 2321500). Subsurface interactions in the unconfined region (station 2322500) necessitate a more complex network to sufficiently capture streamflow variability. The most important improvement to be made in the near future will be to further expand the network graph to allow PAGE 113 113 additional features to be considered. In particular, data related to subsurface interactions and evapotranspiration (e.g. daily temperatures) are expected to significantly improve estimator performance. A graphical mod el constructed by representing spatial stations as a chain of nodes along the river would potentially offer further performance improvements by fusing and propagating the impact of new evidence (streamflow measurements) and beliefs efficiently via message passing 61 Message Passing Message passing deals with fusing and propagating the impact of new evidence and beliefs through B N s. It does so via a self activated propagation mechanism wherein the network is considered as an array of simple and autonomous processors which communicate locally via the links provided by the network [80] The impact of any new evidence is viewed as a perturbation that propagates through the network via message passing between neighboring variables. Consider the network X Y If evidence e ={ Y = y } is observed, then from Baye s rule, the belief distribution of X is given by Equation 6 1 where = P ( e )1, P ( x ) is the prior probability of X and ( x ) is the likelihood vector ( x )= P ( e | x )= P ( Y = y | x )= My | x. ) ( ) ( ) |( ) ( x x P e x P x BEL (6 1) As it can be inferred, the parameter ( x ) can be computed at node Y (from the knowledge of the probability density function P ( Y =y | X =x )) and transmitted to X enabling X to compute its belief BEL ( x ). To understand the propagation scheme, consider the network X Y Z with the evidence e ={ Z = z } The likelihood vector ( x ) can no longer be directly obtained from the matrix My| x, however, but must reflect the matrix Mz|y as well (Equation 6 2). ) ( ) ( ) ( ) ( ) ( ) ( ) ( y M x y P y e P x y P x y e P x e P xx y y y (6 2) PAGE 114 114 Generalizing the chain of networks, every node can calculate the correct current value of its vector if it learns the correct vector of its successor. Since the chain ends with an observed variable whose value is determined externally, the vector of all variables can be determined recursively. If the chain ends with an unobserved variable Z ( z ) is set to one and all belief distributions will coincid e with the prior distribution. Assuming that each node constantly inspects the of its child and updates its own accordingly, it is guaranteed that every variable along the chain will obtain its correct properly reflecting any changes that might hav e occurred in e A dual parameter ( and ) communication/message passing is required when new evidence emerges from both a descendent of a node and its ancestor (i.e. evidence at both the head and tail of the chain). Consider the network in Figure 6 1A where e+ and erepresent evidences available at nodes T and Z (shaded) nodes in X s parent and child subsets In this case, the two evidences are handled by two separate vectors ( and ) and propagated along the chain independent of the other (Equatio n 6 3) where ( x ) remains as defined before (Equation 6 2) and ( x )= ( u) Mx | u. Belief propagation using the bi directed message passing is represented in F igure 61B. ) ( ) ( ) ( ) ( ) ( ) ( ) ( x x x BEL x e P x e x P x (6 3) 61 Streamflow Estimation via Message Passing The test po ints along the river ( T1, T2, etc ) at which we estimate streamflow form the chain network. I n message passing probability vectors ( and ) are transmitted between nodes. If nodes in the network represent flow, the process would involve passing around f low probabilities. This would imply discretising flow that has a relatively large dynamic range. PAGE 115 115 Instead, we formulate the chain wherein nodes represent the difference in flow between consecutive test points (nodes N1, N2, etc in Figure 6 2). The diffe rences in flow estimates have a relatively smaller dynamic range and potentially lead to fine grained results (reduce s blockiness) Preliminary results and performance improvements in using message passing are illustrated in Figure 6 3 and Table 6 1 respe ctively. The following bullets describe the procedure adopted to update streamflow estimates using message passing. Twenty four test points along the river and the 3 USGS stations ( 2321000, 2321500, 2322500) form the chain network. Streamflow estimated us ing the BN model are employed for computing difference nodes ( N1, N2, etc) and used as priors to estimate the conditional densities relating adjacent nodes in the chain ( N1 N2 ) namely My|x. o The three node BN model comprising of WAM, t rue nearest neighb or, and ground water level was used Measured flow at 2321000 and 2322500 for the test period (Feb. 2005 Dec. 2006) was used to compute values for N1 and Nn (tail node) and used as evidence In the current formulation, each Ni node has 64 levels (i.e., t he PN 1( D ) has 64 bins). If evidence on a particular day is d, P ( D = d)=1, that is in the ve c tor P ( D ), only one entry is 1, the rest of the 63 levels are zero ( i.e. P ( D ) is a delta vector). Accordingly at the head node and at the tail node are delta vectors Message passing is utilized to propagate evidence available at N1 and Nn. After estimating difference flow at intermediate nodes, validation is performed at 2321500. Measured flow at 2321500 is compared to three model BN estimated flow and belief propagated flow Solving estimation problems over large data sets in high dimensional spaces leads to computational demands that are often impractical to run on General Purpose Processors. FPGAs were pr oposed as suitable candidates for fast and efficient execution of computationally intensive algorithms Key algorithms that would benefit from hardware acceleration were identified in the BN framework. In proposing FPGA based reconfigurable computing as a suitable technology for hardware acceleration, various challenges faced while designing and developing algorithms PAGE 116 116 on FPGAs were analyzed Algorithm decomposition, performance prediction, and design reuse for parallel implementation all need to be perfor med in an efficient and structured manner for developing successful FPGA designs in a productive manner. To address these challenges, an approach for pattern based decomposition of algorithms for FPGA design and development was proposed. Significant perfo rmance improvements in terms of speedup were obtained by using FPGA accelerated implementation of the Parzen window based PDF estimation algorithm decomposed using primitive RC design patterns. Further improvements can be obtained by scaling the design on FPGA device families that have larger number of dedicated DSP units and a high bandwidth interconnect. Dual core architectures were developed and evaluated on a single device system by exploiting the scalability in the algorithm. Key design parameters w ere identified for tuning the architecture to suit different platforms as well. Precision effects were investigated and data verification along with error statistics suggested a sufficient fixed point configuration for the algorithm. T he benefit of compo site patterns for design reuse was also validated With the successful design of a scalable and portable architecture for the PDF algorithm, rapid hardware development for the K means clustering and correlation algorithms was possible due to the similarit y of their underlying design patterns. Further, the work also showcased the benefit of quantifying design patterns in efficiently exploiting a performance prediction tool, RAT, to predict an algorithms amenability to a hardware platform before undertakin g a lengthy development process. The architecture developed in this work was designed with consideration toward future extensions to multi FPGA systems. Further research is warranted in understanding the potential improvements that can be achieved and bot tlenecks that might have to be addressed while PAGE 117 117 migrating designs to multi FPGA systems. Directions for future work also include investigating and developing design patterns for solving other problem sets that have underlying similarity in their algorithms Case study algorithms for such an analysis can be sampled from popular toolboxes in various fields and develop composite patterns for classes of algorithms within them. For example, Figures 64 and 65 illustrate patterns for tree based (e.g., neural n etworks, many to one mapping problems, multiscale algorithms, Bayesian networks) and mesh based algorithms (e.g., image erosion/dilation, edge detectors, color conversion, spatial filters) param eterized by tree depth and mesh size respectively. This would be a critical step in enabling FPGA based RC for solving a general class of computationally intensive problems. This work has shown that one such general architecture exists for a class of machine learning algorithms. In summary, the primary contribution s of this work are listed below. Proposed a scalable framework that allows efficient fusion of multi modal (domain based diversity) datasets Developed a Spatiot emporal Bayesian Network topology for estimating streamflow along a river by fusing spatio tem poral datasets Predicted and validated streamflow estimates with uncertainties as a function of location (space), condition (time) and particular mix of sensors Developed a scalable and portable architecture for rapid estimating of multi dimensional PDFs on multiple platforms Identified and catalogued primitive design patterns for representing key algorithm decompositions in the proposed BN framework for parallel implementations Quantified design patterns to efficiently exploit and parameterize analytic models in a performance prediction tool Effected design reuse aided by pattern based algorithm decomposition and achieved speedups in the development of K means and Correlation. A few future directi ons to the current work are provided below. Integrating MKS and the proposed BN estimator. This could be accomplished by converting MKS to a BN formulation on a multiscale tree (e.g. quadtree) PAGE 118 118 Implementing a Dynamic Bayesian Network (DBN) to model streamflow over days (or any time resolution) as a sequence o f variables thereby extracting knowledge of past streamflow in estimating current or future streamflow Applying the proposed data fusion framework to non hydrologic application s such as forestry, bio medical engineering and other remote sensing applicati ons Exploiting design patterns for generating code templates for FPGA design and development. Table 6 1. Performance metrics at 2321500 after message passing Metric 3 node BN (N2) Message passing NSC 0.7183 0.7346 Mean abs. error (cu.ft/sec) 174.0000 58.0000 Std.dev of error (cu.ft/sec) 248.0000 288.0000 Entropy of error PDF (nats) 2.1500 2.3400 A B Figure 61. Illustration of m essage passing. A) E vidence originating at both ends of chain B) D ual parameter message passing Figure 62. Message passing framework for improving streamflow estimates PAGE 119 119 Figure 63. Prediction at USGS 2321500 after message passing Figure 64. Composite pattern for tree based algorithms Figure 65. Composite pattern for mesh based algorithms PAGE 120 120 LIST OF REFERENCES 1. P. W. Fieguth, W. C. Karl, A. S. Willsky and C. Wunsch, Multiresolution Optimal Interpolation and Statistical Analysis of TOPE X/POSEIDON Satellite Altimetry, IEEE Trans actions on Geoscience and Remote Sensing vol. 33( 2) 1995, pp. 280292. 2. M. M. Daniel, and A. S. Willsky, A Multiresolution Methodology for Signal level Fusion and Data Assimilation with Applications to Remote Sensing, Proceedings of the IEEE, vol. 85( 1 ), 1997, pp. 164180. 3. F. Camastra and A. Vinciarelli, Machine Learning for Audio, Image and Video Analysis Theory and Applications, Springer Verlag London Limited, London, UK, 2008. 4. N. Dimitrova, H. Zhang, B. Shahraray, I. Sezan, T. Huang, and A. Zakhor, Applications of Videocontent Analysis and Retriev al, Journal of Multimedia, vol. 9(3), 2002, pp. 4255. 5. F Lillo, S Basile, and R N. Mantegna, Comparative Genomics Study of Inverted Repeats in Bacteria, Bioinformatics, vol. 18(7), 2002, pp. 971979. 6. F. Pitie, A.C. Kokaram, and R. Dahyot, N Dimens ional Probability Density Function Transfer and its Application to Color Transfer, Proc eedings of the 10th International Conference on Computer Vision, Beijing, China, 17th October 2005 21st October 2005, pp. 1434 1439. 7. S M. Kay, A H. Nuttall, and P M. Baggenstoss, Multidimensional Probability Density Function Approximations for Detection, Classification, and Mod el Order Selection, IEEE Transactions on Signal Processing, vol. 49(10), 2003, pp. 2240 2252. 8. Y. Chang, W. Zeng, I. Kamel, and R. Alo nso, Integrated Image and Speech Analysis for Content Based Video Indexing, Proc eedings of 3rd International Conference on Multimedia Computing and Systems, Hiroshima, Japan, 17th June 1996 23rd June 1996, pp. 306313. 9. H. Zhang, Y. Zhuang, and F. Wu, Cross modal Correlation Learning for Clustering on Image Audio Dataset, Proc eedings of the 15th International Conference on Multimedia, Augsburg, Germany, 2007, pp. 273276. 10. R. Goecke, J. B. Millar, A. Zelinsky, and J. Robert Ribes, Analysis of Audio Vi deo Correlation in Vowels in Australian English, International Conference on Audio Visual Speech Processing, Aalborg, Denmark, 2001, pp. 115120. 11. L. Greengard, and J. Strain, The Fast Gauss Transform, SIAM Journal on Scientific and Statist ical Computin g vol. 12(1), 1991, pp. 7994. 12. D.E. Culler and J.P. Singh, Parallel Computer Architecture: A Hardware/Software Approach, Morgan Kaufmann Publishers Inc orporation San Francisco, California, 1999. PAGE 121 121 13. WATERS NETWORK, WATERS Test Bed Sites, 2005, http://www.watersnet.org/wtbs/index.html, 22nd April 2009. 14. D. E. Holmes, L. C. Jain, Innovations in Bayesian Networks: Theory and Applications, Springer, New York, New York, 2009. 15. J. Zhang, and B. Whalen, Estimated Phosphorus Load Reductions Under Various Water Management Alternatives, American Society of Agricultural and Biological Engineers Annual International Meeting, St.Joseph, Michigan, July 2005. 16. A. B. Bottcher, B. M. Jacobson, and J. G. Hiscock, C haracterizing Flow and Nutrient Loads for TMDL Development in Florida using WAM, Proceedings of the Conference on Total Maximum Daily Load (TMDL) Environmental Regulations II, Albuquerque, New Mexico, 8th November 2003 12th November 2003, pp. 2734. 17. M. Yadav, T. Wagener, and H. Gupta, Regionalization of Constraints on Expected Watershed Response Behavior for Improved Predictions in Ungauged Basins, Advances in Water Resources, vol. 30 (8) 2007, pp. 17561774. 18. D. P. Boyle, H. V. Gupta, and S. Sorooshian Toward Improved Calibration of Hydrologic Models: Combining the Strengths of Manual and Automatic M ethods, Water Resources Res earch vol. 36(12), 2000, pp. 36633674. 19. M. E. Borsuk, D. Higdon, C. A. Stow, and K. H. Rechhow, A Bayesian Hierarchical Model to Predict Benthic Oxygen Demand from Organic Matter Loading in Estuaries and Coastal Zones, Ecological Modeling, vol. 143(3), 2001, pp. 165181. 20. K. C. Slatton, S. Cheung, and H. Jhee, ReducedComplexity Fusi on of Multiscale Topography and Bathymetry Data over the Florida Coast, IEEE Geosci ence and Remote Sensing Letters, vol. 2(4), October 2005, pp. 389393. 21. L. Alparone, S. Baronti, A. Garzelli, and F. Nencini, A Global Quality Measurement of Pan sharpened Multispectral Imagery, IEEE Geoscience and Remote Sensing Letters, vol. 1(4) October 2004, pp. 313317. 22. V. Tsagaris, V. Anastassopoulos, and G. A. Lampropoulos, Fusion of Hyperspectral Data Using Segmented PCT for Color Representation and C lassificatio n, IEEE Transactions on Geoscience and Remote Sensing, vol. 43(10) October 2005, pp. 2365 2375. 23. W. T. Crow, Correcting Land Surface Model Predictions for the Impact of Temporally Sparse Rainfall Rate Measurements Using an Ensemble Kalman Filter and Surf ace Brightness Temperature Observations, Journal of Hydrometeorology, vol. 4(5) Ap ril 2003, pp. 960973. 24. J. W. Jones, W. D. Graham, D. Wallach, M. Bostick, and J. Koo, Estimating Soil Carbon Levels using an Ensemble Kalman Filter, Transactions of the A merican S ociety of A gricultural E ngineers vol. 47(1), 2004, pp 331339. PAGE 122 122 25. L. Bruzzone, D. F. Prieto, and S. B. Serpico, A Neural Statistical Approach to Multitemporal and Multisource Remote Sensing Image Classification, IEEE Trans actions on Geoscience and Remote Sensing, vol. 37(3), May 1999, pp. 13501359. 26. M. Datcu, F. Melgani, A. Piardi, and S. B. Serpico, Multisource Data Classification with Dependence T rees, IEEE Transactions on Geoscience and Remote Sensing, vol. 40(3), March 2002, pp. 609617. 27. C. K Wikle, R. F. Milliff, D. Nychka, and L. M. Berliner, Spatiotemporal Hierarchical Bayesian Modeling: Tropical Ocean Surface Winds, Journal of the American Statistical Association, vol. 96(454), June 2001, pp. 382397. 28. M. Koeppen, The Curse of Dimension ality, 5th Online World Conference on Soft Computing in Industrial Applications (WSC5), held on the internet, 4th September 2000 18th September 2000. 29. C. Krekeler, K. C. Slatton, and M. Cohen, Multi variate Bayesian Classification of Soil Drainage Using Feature level Fusion of Topographic and Hydrologic Data, Proceedings of IEEE International Geoscience and Remote Sensing Symposium Denver, Colorado, 31st July 2006 4th August 2006, pp. 25222525. 30. N. Mount, and T. Scott, A Discrete Bayesian Network to Invesigate Suspended Sediment Concentrations in an Alpine Proglacial Zone, Journal of Hydrological Processes, vol. 22(18) February 2008, pp. 37723784. 31. M. Molina, R. Fuentetaja, and L. Garrote, Symbolic and Quantitative Approaches to Reasoning with Unc ertainty, SpringerLink, Heidelberg, Berlin, 2005, pp. 8899. 32. S. Lanini, Water Management Impact Assessment Using a Bayesian Network Model, International Conference on Hydroinformatics, Nice, France, 4th September 2006 8th September 2006, pp. 105112. 33. W. A. Pike, Modeling Drinking Water Quality Violations with Bayesian Networks, Journal of the American Water Resources Association, vol. 40(6), 2004, pp. 15631578. 34. M. Park, and M. K. Stenstrom, Using Satellite Imagery for Stormwater Pollution Managem ent with Bayesian Networks, Waters Research, vol. 40(18), 2006, pp. 3429 3438. 35. B. G. Katz, R. S. DeHan, J. J. Hirten, and J.S Catches, Interactions Between Ground Water and Surface Water in the Suwannee River Basin, Florida, Journal of the American Wat er Resources Association, vol. 33(6), 1997, pp. 12371254. 36. C. A. Crandall, B. G. Katz, and J. J. Hirten, Hydrochemical Evidence for Mixing of River Water and Groundwater During High flow Conditions, Lower Suwannee River Basin, Florida, USA, Hydrogeology Journal, vol. 7 (5) October 1999, pp. 454467. PAGE 123 123 37. US Geological Society, National Elevation Dataset, 2008, http://seamless.usgs.gov/, 22nd April 2009. 38. National Engineering Handbook, Section 4 Hydrology, Washington, D.C. USDA SCS, 1985. 39. W. Viessman, G. L. Lewis, and J. W. Knapp, Introduction to Hydrology, Harper Collins 3rd edition New York NY, 1989. 40. R. H. McCuen, Hydrologic Analysis and Design, Prentice Hall, 3rd edition, Upper Saddle River, NJ, 2005. 41. Sant a Fe River Watershed T est bed, a NSF WATERS Project, 2007, http://suwanneeho.ifas.ufl.edu/SantaFe.htm 22nd April 2009. 42. M. Llyas, General Probability Density Function of Packet Service Times for Computer Networks, Electronics Letters, vol. 23(1) 1987, pp. 3132. 43. R R. Bliss and N Panigirtzoglou, Testing the S tability of I mplied P robability D ensity Functions, Journal of Banking and Fin ance, vol. 26(2), 2002, pp. 381422. 44. H. Jiang, T. Lin, and H. Zhang, Video Segmentation with the Assistance of Audio Content Analysis, International Conference on Multimedia and Expo, New York, NY, 2000, pp. 15071510. 45. K.S. Hemmert and K.D. Underwood, An A nalysis of the D ouble precision F loating point FFT on FPGAs, IEEE Symposium on Field Programmable Custom Computi ng Machines, Washinton, DC, April 2005, pp. 171180. 46. G. Govindu, S. Choi, V. Prasanna, V. Daga, S. Gangadharpalli, and V. Sridhar, A H igh performance and E nergy efficient A rchitecture for Floating point B ased LU D ec omposition on FPGAs, IEEE Symposium on Parallel and Distributed Processing, Santa Fe, N ew M exico, April 2004, pp. 149 47. J. S. Kim, P. Mangalagiri, K. Irick, N. Vijaykrishnan, M. Kandemir, L. Deng, K. Sobti, C. Chakrabarti, N. Pitsianis, a nd X. Sun, TANOR: A Tool for Accelerating N body Simulations on Reconfigurable Platform, Proc eedings of the 17th International Conference on Field Programmable Logic a nd Applications, Amsterdam, August 2007, pp. 68 73. 48. M. Leeser, J. Theiler, M. Estlick and J. J. Szymanski, Design Tradeoffs in a Hardware Implementation of th e K means Clustering Algorithm, Proceedings o f Sensor Array and Multichannel Signal Processing Workshop, Cambridge, Massachusetts, 16th March 2000 17th March 2000, pp.520524. PAGE 124 124 49. I Frohlich, A. Gabriel, D. Kirschner, J. Lehert, E. Lins, M. Petri, T. Perez Cavalcanti, J. Ritman, D. Schafer, A. Toia, M. Traxler, and W.Kuehn, Pattern Recognition in the HADES Spectrometer: An Application of FPGA Technology in Nuclear and Particle Ph ysics, Proc eedings of the International Conference on Field Programmable T echnology (FPT), Singapore, December 2002 pp. 443 444. 50. S. Neo, H. Goh, W. Y. Ng, J. Ong, and W. Pang, Real time Online Multimedia Content Processing: Mobile Video Optical Chara cter Recognition and Speech Synthesizer for the Visual Impaired, Proc eedings of the International Convention on Rehabilitation Engineering and Assistive Technology, Singapore, 2007, pp. 201206. 51. H. Schmit and D. Thomas, Hidden Markov M odeling and F uzzy C ontrollers in FPGAs, Proc eedings of Symposium on FPGAs for Custom Co mputing Machines, Napa, CA, April 1995, pp. 214 221. 52. T. VanCourt and M. Herbordt, Three Dimensional Template Correlation: Object Recognition in 3D Voxel Data, Proc eedings of Computer Architecture for Machine Perception, Washinton, DC, 2005, pp. 153158. 53. B Holland, K Nagarajan, C Conger, A Jacobs, and A D. George, "RAT: A Methodology for Predicting Performance in Application D esign Migration to FPGAs," Proceedings of High Performan ce Reconfigurable Computing Technologies and Applications Workshop, Reno, Nevada, 11th November 2007. 54. C. P. Steffen, Parametrization of Algorithms and FPGA Accelerators to Predict Performance, Proceedings of Reconfigurable System Summer Institute (RSSI), Urbana Illinois 2007, pp. 1720. 55. A. DeHon, J. Adams, M. DeLorimier, N. Kapre, Y. Matsuda, H. Naeimi, M. Vanier, and M. Wrighton, Design Patterns for Reconfigurable Computing, IEEE Symposium on Field Programmable Custom Computing Machines, Napa Valley, C alifornia 2004, pp. 1323. 56. J. W. Fisher and J. C. Principe, A Methodology for Information Theoretic Feature Extraction, Neural Networks Proceedings on IEEE World Congres s on Computational Intelligence, vol. 3(1) 4th May 1998 9th May 1998, pp. 17121716. 57. J. Pearl, Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference, Morgan Kaufmann, San Francisco, CA, 1988. 58. F. Jensen, Bayesian Networks and Decision Graphs, Springer Verlag, New York, New York, 2001. 59. R. O. Duda, P. E. Ha rt, and D. G. Stork, Pattern Classification, John Wiley & Sons Inc. 2nd edition New York, New York, 2001. PAGE 125 125 60. D. Erdogmus, Y. N. Rao, and J. C. Principe, Nonlinear Independent Component Analysis by Homomorphic Transformation of the M ixtures, Proceedings of the IEEE International Joint Conference on Neural Networks, vol. 1, 25th July 2004 29th July 2004, pp. 4752. 61. D. R. Legates and G. J. McCabe, Evaluating the Use of Goodness of fit Measures in Hydrologic and Hydroclimatic Model Validation, Water R esources Research, vol.35(1), 1999, pp. 233242. 62. E. Gamma, R. Johnson, H. Helm, J. M. Vlissides, and G. Booch, Design Patterns: Elements of Reusable Object Oriented Software, Addison Wesley Professional, 1994, pp. 416. 63. J. Anvik, S. MacDonald, D. Szafron, J. Schaeffer, S. Bromling and K. Tan, Generating Parallel Programs from the Wavefront Design Pattern", Proc eedings of the Int ernational Workshop on High Level Parallel Programming Models and Supportive Environments, Fort Lauderdale, FL, 2002, pp. 104111. 64. K. T. Gribbon, D. G. Bailey, and C. T. Johnston, Design Patterns for Image Processing Algorithm Development on FPGAs, IEEE TENCON, 21st November 2005 24th November 2005, pp. 16. 65. M. Y. Mashor, Improving the Performance of K means Clustering Algorith m to Position the Centres of RBF Networks, Int ernational Journal of the Computer, the Internet and Management, vol. 6(2), 1998, pp. 119. 66. D. L. Turcotte, Fractals and Chaos in Geology and Geophysics Cambridge University Press 2nd edition, New York, Ne w York, 1997. 67. H. O. Peitgen and D. Saupe, The Science of Fractal Images Springer, New York, New York, 1988. 68. Z. Li, Q. Zhu and C. Gold, Digital Terrain Modeling: Principles and Methodology CRC Publishers, Boca Raton, Florida 2004. 69. L. M. Kaplan, Exte nded Fractal Analysis for Texture Classification and S egmentation, IEEE Transactions On Image Processing, vol. 8 ( 11 ) November 1999, pp. 1572 1585. 70. A. P. Pentland, Fractal based Description of Natural S cenes, IEEE Transactions on P attern A nalysis and M achine I ntelligence vol. 6, November 1984, pp. 661674. 71. V. Aggarwal, K. Nagarajan, and K. C. Slatton, Multiple Model Multiscale Data Fusion Regulated by a Mixture of Experts Network, Proceedings of IEEE International Geoscience and Remote Sensing Symposium vol. 1, September 2004, pp. 364 367. 72. W. S. Chaer, R. H. Bishop, and J. Ghosh, A Mixture of Experts Framework for Adaptive Kalman F iltering, IEEE Transactions on Systems, Man and Cybernetics, vol. 27( 3) June 1997, pp. 452464. PAGE 126 126 73. C. V. Stewart, B. M oghaddam, K. J. Hintz, and L. M. Novak, Fractional Brownian Motion Models for Synthetic Aperture Radar Imagery Scene S egmentation, Proceedings of IEEE, vol. 81, October 1993, pp. 15111522. 74. J. M. Keller, S. Chen, and R. M. Crownover, Texture Description and Segmentation Through Fractal G eometry, Computer Vision Gra phics and Image Processing, vol. 45(2) 1989, pp. 150166. 75. T. Kumar, P. Zhou, and D. A. Glaser, Comparison of Human Performance with Algorithms for Estimating Fractal Dimension of Fraction al Brownian S tatistics, Journal of the Optical Society of America vol. 10(6) June 1993, pp. 11361146. 76. L. M. Kaplan and C. J. Kuo, Texture Roughness Analysis and Synthesis via Extended Self similar (ESS) M odel, IEEE Transactions on Pattern Analysis an d Machine Intelligence vol. 17(11) November 1995, pp. 10431056. 77. Y. Deng, New Trends in Digital Terrain Analysis: Landform Definition, Representation, and Classification Progress in Physical Geography, vol. 31(4), 2007, pp. 405419. 78. K. C. Clarke, Com putation of the Fractal Dimension of Topographic Surfaces Using the Triangular Prism Surface Area Method, Co mputers and Geoscience, vol. 12( 5) May 1986, pp. 713 722. 79. M.Schroeder, Fractals,Chaos,Power Laws W. H. Freeman, Gordonsville, Virginia 8th print, 1992. 80. J. Pearl, Probabilistic reasoning in intelligent systems: Network of plausible inference, Morgan Kaufman, San Mateo, California, 1988. PAGE 127 127 BIOGRAPHICAL SKETCH K arthik N agarajan received his Ph.D. and M.S. degrees in electrical engineering from the University of Florida in 2005 and 2009 respectively while simultaneously working for the Adaptive Signal Processing Laboratory (ASPL) and the NSF Center for High performance Reconfigurable Computing (CHREC) He received the B.E. degree in e lectronics and c ommunication e ngineering from the University of Madras in 2003. His research interests include pattern recognition, information theory, graphical models and multiscale estimation concepts. He is also an active researcher in the field of high perfo rmance and reconfigurable computing with focus in FPGA design and development, performance analysis and productivity studies |