High Performance Computing in Remote Sensing - Chapter 13

Chapter 13 A Solutionware for Hyperspectral Image Processing and Analysis Miguel Ve´lez-Reyes, University of Puerto Rico at Mayaguez Wilson Rivera-Gallego, University of Puerto Rico at Mayaguez Luis O. Jime´nez-Rodriguez, University of Puerto Rico at Mayaguez Contents 13.1 Introduction ........................................................... 310 13.2 Hyperspectral Image Analysis Toolbox ................................. 311 13.2.1 HIAT Functionality ............................................. 312 13.2.1.1 Pre-Processing ........................................ 312 13.2.1.2 Feature Extraction and Band Subset Selection .......... 316 13.2.1.3 Classification .......................................... 317 13.2.1.4 Unsupervised Spatial-Spectral Post-Processing ......... 317 13.2.1.5 Covariance Estimators ................................. 318 13.2.1.6 Unmixing ............................................. 319 13.2.2 Other Components .............................................. 320 13.2.3 Toolbox Availability ............................................ 321 13.3 Implementing Components of HIAT in Parallel Computing Using the Itanium Architecture ............................................... 321 13.3.1 Principal Components Algorithm ................................ 323 13.3.2 Euclidean Distance Classifier ................................... 324 13.3.3 Maximum Likelihood Classifier ................................. 325 13.4 A Grid Service-Based Tool for Hyperspectral Image Analysis ........... 326 13.4.1 Grid-HSI Architecture .......................................... 327 13.4.2 Functional Process of Grid-HSI ................................. 329 13.4.3 Experimental Results ........................................... 329 13.5 Conclusions ........................................................... 331 13.6 Acknowledgment ...................................................... 332 References .................................................................. 332 309 © 2008 by Taylor & Francis Group, LLC 310 High-Performance Computing in Remote Sensing This chapter describes the concept of a solutionware for hyperspectral image anal-ysis. Solutionware is a set of catalogued tools and toolsets that will provide for the rapidconstructionofarangeofhyperspectralimageprocessingalgorithmsandappli-cations. The proposed hyperspectral solutionware will span toolboxes, visualization toolsets, and application-specific software systems at different computational resolu-tion levels. A MATLAB hyperspectral image analysis toolbox (HIAT) provides the lowest resolution level but the friendliest interface where scientists and engineers in hyperspectral image processing can try different combinations of hyperspectral image processing algorithms in a simple fashion and add their own algorithms via theMATLABprogramminglanguage.Asapplicationsrequiretheprocessingoflarge datasetsinatimelyfashion,thesolutionwarewillprovidegrid,parallel,andhardware computational platforms to provide the user with computational alternatives that can be used to optimize performance and take full advantage of the data. In this chapter, theMATLABhyperspectraltoolboxispresentedandparallelprocessingimplementa-tions of some of its components in the Itaniun architecture are described. A prototype versionofthehyperspectralimageprocessingtoolboxoverthegrid,Grid-HSI,which extends the hyperspectral image processing environment developed in HIAT to take advantage of computational resources that can be distributed over the network, is depicted. 13.1 Introduction Hyperspectral image analysis usually consists of performing a series of highly com-putational intensive operations on large data sets. The analysis extracts information of interest from the data contained in a region for the application scientist. This in-formation of interest may include the extraction of features, the classification of a region in an image, or simply the detection of some specific object. However, two of the main constraints in obtaining analysis results in a timely manner are the efficient computation of the operation itself and the efficient management of large volumes of data. The massive volumes of data involved in hyperspectral imagery is the main limitation in testing different varieties of algorithms on the data as well as in the extraction of features in a timely fashion. When describing the methods to solve a computational problem in hyperspectral imaging, the levels of abstractions in the architecture are of primary importance on the performance observed in the computation. Figure 13.1 illustrates the different levels of abstractions where a computing problem may be solved. It is interesting to note that the higher the level of abstraction, the design description syntax the to the ‘language’ spoken by the application closer is the programmer. By programming at a high level of abstraction in an environment such as MATLABT M [1] or IDLT M [2], the programmer can quickly construct a set of algorithms to solve a problem. Also, these environments are capable of providing a framework for proper software engineeringpracticestobefollowed.However,thisimpliesthattheapplication’sper-formance might decrease as the abstraction level increases. In theory, working at a © 2008 by Taylor & Francis Group, LLC A Solutionware for Hyperspectral Image Processing and Analysis 311 Problem Solving Level (High-level Languages) Assembly Language Level OS Machine Level ISA Level Microarchitecture Level Digital Logic Level Figure 13.1 Levels in solving a computing problem. lower level of abstraction can result in better system performance since the developer has more control over the computational structures. However, as the level of abstrac-tion decreases, the complexity in the design process increases, making it harder for the developer to have a complete grasp of the whole design process. The objective of our work is to develop a hyperspectral solutionware or a set of catalogued tools and toolsets that will provide for the rapid construction of a range of hyperspectral image processingalgorithmsandapplications.Solutionwaretoolswillspantoolboxes,visu-alizationtoolsets,andapplication-specificsoftwaresystemsthathavebeendeveloped at the Center for Subsurface Sensing and Imaging Systems1 (CenSSIS). The chap-ter is organized as follows. First, an overview of the MATLAB hyperspectral image analysis toolbox is given. Second, parallel and grid implementations of some of the algorithmsinthetoolboxaredescribed.Futuredirectionsoftheworkaresummarized at the end. 13.2 Hyperspectral Image Analysis Toolbox The Hyperspectral Image Analysis Toolbox (HIAT) is a collection of algorithms that extend the capability of the MATLAB numerical computing environment for the processing of hyperspectral and multispectral imagery. The purpose of HIAT is to provide information extraction algorithms to users of hyperspectral and multispectral imagery in different application domains. HIAT has been developed as part of the NSF Center for Subsurface Sensing and Imaging (CenSSIS) Solutionware that seeks to develop a repository of reliable and reusable software tools that can be shared by researchers across research domains. HIAT provides easy access to supervised and unsupervised classification algorithms, unmixing algorithms, and visualization tools 1http://www.censsis.neu.edu. © 2008 by Taylor & Francis Group, LLC 312 High-Performance Computing in Remote Sensing Unmixing Methods Unsupervised Classification Abundance Estimates Class Map Figure 13.2 HIAT graphical user interface. developed at UPRM Laboratory for Applied Remote Sensing and Image Processing (LARSIP) over the last 8 years. HIATisimplementedwithinanoptimizedMATLABenvironment.Itprovidesuse-ful image analysis techniques for educational and research purposes, allowing the in-teractionanddevelopmentofnewalgorithms,datamanagement,resultscomparisons, andpost-processing.Itisaneasy-to-useandpowerfultoolforresearchersinvolvedin hyperspectral/multispectral image processing. In addition, MATLAB provides porta-bilityofthecodetothedifferentplatformsinwhichMATLABruns:Windowsfamily, Mac OS, and UNIX systems. 13.2.1 HIAT Functionality The GUI of the toolbox is shown in Figure 13.2. MATLAB version 6.5 was used for theimplementationoftheHIAT.TestsarecurrentlybeingconductedusingMATLAB version7.2(MATLAB2006a)toensurethetoolboxisupwardcompatible.Figure15.3 shows the processing schema implemented in HIAT. The processing phases of HIAT are divided into four groups: Feature Selection/Extraction, Classification, Unmixing, andPost-Processing.AsFigure13.3shows,HSIdatacouldbeprocessedwithfeature selection/extraction algorithms (or not) before the classification or unmixing and to enhance the classification map post-processing algorithms that are used. Users can combine different processing algorithms to generate different data products. 13.2.1.1 Pre-Processing In the toolbox, it is assumed that the image has undergone any sensor specific pre-processing, de-glinting, or geometric and atmospheric correction before processing. © 2008 by Taylor & Francis Group, LLC A Solutionware for Hyperspectral Image Processing and Analysis 313 Multi/ Hyperspectral Data Image Enhancement Feature Selection/ Extraction Classification/ Unmixing Classification or Abundance Map HIAT Functionality Post Processing Figure 13.3 Data processing schema for hyperspectral image analysis toolbox. Pre-processing in the toolbox is limited to image enhancement including noise reduction. One of the most widely used noise reduction algorithms for hyperspectral imagery is Reduced Rank Filtering (RRF). In this type of filtering, a principal component de-composition is performed using the singular value decomposition (SVD); the small singular values are set to zero, which produces a reduced noise (low rank) approxi-mation of the original image. In addition to this algorithm, a noise reduction method based on oversampling filtering, developed by the authors [3], is available in the toolbox. The oversampling filtering technique takes advantage that hyperspectral imagers typically collect 100-300 contiguous spectral bands at a high spectral resolution ( 10nm in most sensors), which results in more samples than are needed to repre-sent the spectra of many objects. Having more samples than are needed is known as oversampling. Oversampling is defined as sampling a signal higher than its Nyquist rate. Specifically, the oversampling rate can be written as OSR = 2 fm (13.1) where fm is the maximum frequency in the signal and fs is the sampling frequency. The maximum frequency of the sampled signal power spectral density (PSD) is directly proportional to the maximum frequency of the original signal and inversely proportionaltothesamplingrate.Thismeansthatforafixedmaximumfrequencyina signal,thehigherthesamplingrate,thelowerthemaximumfrequencyofthesampled signalPSD.ThisisillustratedinFigure13.4.Figure13.4(a)showsthesampledsignal PSD when the signal is sampled at the Nyquist rate while Figure 13.4(b) shows the sampled signal spectra when the signal is sampled at twice the Nyquist rate. The usefulness of oversampling for noise reduction is that if the signal has been oversampled and there is noise (anything other than the signal of interest) in the frequencyrangenotoccupiedbythesignal,itcanbelowpassfilteredwithoutchanging thesignal.Thereductioninnoisemeansanincreaseinthesignal-to-noiseratio.Ithas beenshownthattypicalreflectancespectraareoversampledbyafactorof4whenthey © 2008 by Taylor & Francis Group, LLC ... - tailieumienphi.vn