The MATLAB Hyperspectral Image Analysis Toolbox

Transcription

1 The MATLAB Hyperspectral Image Analysis Toolbox Samuel Rosario-Torres, Miguel Vélez-Reyes Phd. and Luis O. Jiménez Ph.D. Laboratory of Applied Remote Sensing and Imaging System University of Puerto Rico Mayagüez Campus Abstract 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 the HIAT Toolbox is to provide information extraction algorithms to users of hyperspectral and multispectral imagery in environmental and biomedical applications. 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 developed at Laboratory of Remote Sensing and Image Processing (LARSIP) over the last 10 years. New features added this year include incorporations of Spatial Feature Extraction, Abundance Estimation with manual endmember acquisition, Image Enhancement using oversampling and principal component filtering. In this tutorial we present the image processing workflow using the AVIRIS cuprite image. Keywords: Hyperspectral, Image Analysis, Supervised Classification, Unsupervised Classification, Toolbox I. Introduction Hyperspectral Image analysis is supported by a variety of available software packages. The best known commercial product is the Environment for Visualizing Images (ENVI) [1] of ITT Visual Information Solutions, a ITT Industries, Inc. subsidiary. ENVI provides code extensibility through the Interactive Data Language (IDL), allowing the possibility for The toolbox development work and algorithms development work was partially supported by the NSF Engineering Research Centers Program under grant EEC , by NASA University Research Centers Program under grant NCC5-518, the Department of Defense under DEPSCoR Grant DAAG and NIMA grant NMA (formally call NGA.)

2 BOOK CHAPTER 2 routine and features expandability. Among the educational non-commercial products, the best known is MultiSpec [2] developed at Purdue University by Dr. David Landgrebe and the Remote Sensing research group in Purdue s LARS. Multispec provides similar features to ENVI but does not provide extensibility. At UPRM, researchers at the Laboratory for Applied Remote Sensing and Image Processing (LARSIP) have been working on multi and hyperspectral image processing for over 8 years. To support researchers in environmental and biomedical applications using multi/hyperspectral imagery at UPRM LARSIP and at the Center for Subsurface Sensing and Imaging Systems (CenSSIS), a toolbox that incorporated the algorithms developed at LARSIP along with standard algorithms for classification similar to those included in ENVI and MultiSpec in the MATLAB platform widely used in engineering and science was necessary. HIAT includes original work in the areas of feature extraction/selection [3], [4] and contextual information classification enhancement [5] developed at UPRM. This paper presents the second release of HIAT. The algorithms implemented in the toolbox were developed during the last years in research projects sponsored by NASA-TCESS, DoD, ARMY TEC, and NSF. Different teams of students developed these algorithms. The key objective of the original work was primarily proof of concept demonstrations. In order to reach a wider audience in environmental and biomedical applications and to be able to use the algorithms in the processing of larger data sets within the MATLAB environment, we worked on the integration, optimization, robustness and user interfaces of the original prototypes. HIAT is aimed to be the common application for researchers in CenSSIS and LAR- SIP that use hyperspectral and multispectral imagery in different applications. The following sections will describe the previous HIAT prototypes developed at LARSIP, the modified toolbox, and an example how to use HIAT. HIAT is available for download at II. Hyperspectral Image Analysis Toolbox The Hyperspectral Toolbox is currently being developed as an element of the CenSSIS Solutionware framework [11]. The objective of the CenSSIS Solutionware team is to develop a set of catalogued tools and toolsets that will provide for the rapid construction of a range of subsurface algorithms and applications. Solutionware tools will span toolboxes, visualization toolsets, database systems and application-specific software systems that have been developed in the Center.

3 BOOK CHAPTER 3 Fig. 1. Data Processing Scheme To facilitate programming of this toolbox, languages such as MATLAB are most appropriate. By programming at a high level of abstraction, the programmer can quickly construct a set of algorithms to solve a problem. Also, MATLAB is capable of providing a framework for proper software engineering practices to be followed. In addition, MATLAB provides portability of the code in the different platforms in which MATLAB works Windows family, Mac OS and UNIX systems. HIAT is used within an optimized MATLAB environment. It provides useful image analysis techniques for educational and research purposes, allowing the interaction and development of new algorithms, data management, results comparisons and post-processing. It is easy-to-use and powerful software for the CenSSIS community involved with HIAT. This toolbox implemented in MATLAB serves as an excellent educational resource for student laboratories, hence improving the classroom experience with graphical examples. A similar approach has been taken in the development of Multi-View Tomography Toolbox, another CenSSIS Solutionware application [7]. We made this toolbox available to the public for research and educative purpose in Table I shows a list of some of the HIAT users. A. HIAT Functionalities Figure 1 shows the processing schema of HIAT (V 2.0). The processing phases of HIAT were divided into 4 groups Image Enhancement (Pre-processing), Feature Selection/Extraction, Classification/Unmixing and Post-processing. As Figure 1 shows, data could be processed with feature selection/extraction algorithms (or not) before the classification and to enhance the classification map we could use post processing algorithms.

4 BOOK CHAPTER 4 TABLE I HIAT Users Florida Environmental Research Institute Lawrence Livermore National Lab National Coral Reef Institute/NSU Surrey Space Centre Air Force Institute of Technology Air Force Research Laboratory The Forestry Research Institute Of Sweden Raytheon US Army Jet Propulsion Laboratory Pacific Northwest National Labs Marine Corps. Intelligence Activity Canada Border Services Agency HIAT have capabilities to load images in the common remote sensing format (bip, bil, bsq), jpg, tiff, ASTER and ours.mat files format. The visualization of the images could be in gray, color composite (manual or automatic selection) or true color (wavelength needed). Image Enhancement pre-processing algorithms enhance the hyperspectral image in the spatial domain or spectral domain. In HIAT we include image enhancement algorithms in the spectral domain, some from the literature and others develop in LARSIP. The algorithms include in HIAT are: Resolution Enhancement PCA Filtering Enhancement Hyperspectral images contains hundreds of contiguous narrows bands making the classification process expensive computationally. Feature Selection/Extraction algorithms provide users the ability to reduce the dimensionality of the HSI data. The available algorithms are: Principal Components Analysis [8] [9] Discriminant Analysis [8] [9] Singular Value Decomposition Band Selection [3] [4] Information Divergence Band Subset Selection [4] Information Divergence Projection Pursuit [10] Optimized Information Divergence Projection Pursuit [4] In the toolbox, standards classifiers are included. In addition, supervised and unsupervised versions of each classifier are included. Supervised classification process allows the users to select the training and testing samples for the spectral classes represented in the image or the users can load the training and testing samples stored previously. In unsupervised classification, the users can select the threshold and stopping criteria to stop the

5 BOOK CHAPTER 5 algorithms. The classifiers are []: Euclidean Distance Fisher s Linear Discriminant Mahalanobis Distance Maximum Likelihood Angle Detection Fuzzy Maximum Likelihood Fuzzy Euclidean Distance Post Processing techniques integrate contextual information of the scene into the resulting classification map, this integration results in an increase on the classification accuracy [5]. Integrated to the toolbox are: Supervised and Unsupervised Extraction and Classification of Homogeneous Objects (ECHO)[5] Window size 2x2, 3x3 and 4x4. Other options to classification is spectral unmixing. The spectral unmixing is the procedure of decompose the measure spectrum of mixed pixels into a set of originating spectra, endmember, and a set of corresponding abundance fractions, abundances [14]. In the toolbox we have integrated some algorithms for abundance estimation these are [14]: Non Negative Sum to One Non Negative Sum Less or Equal to One Non Negative Least Square Sum to One Constrained Unconstrained The GUI of the new version of the toolbox is shown in Figure 2. As we can see, the GUI was composed of a main window in which the image being analyzed is displayed. In the toolbox, when no image is loaded (when the application start) the application toolbar is inactive only the File and Help menu are active. One of the advantages of the toolbox has to offer is that the user can apply supervised or unsupervised classification algorithms and routines to the same image in the same application in order of doing a more extensive analysis of the data they wish to analyze. Also can do post-processing algorithms based on the different results obtained. In addition if the user have other algorithms that which desire to apply to the processed image it can be done through the MATLAB command prompt.

6 BOOK CHAPTER 6 Fig. 2. HIAT GUI MATLAB version 6.5 was used for the implementation of the HIAT. We are currently conducting tests using MATLAB version 7.2 to ensure the toolbox is fully upward compatible with different platforms. In addition, we create an standard alone version for Windows. III. Using the Toolbox: An Example A graphical user interface (GUI) in MATLAB has been developed in order to facilitate the use of the HIAT routines. The main window displays the image once it is loaded. All the interaction between HIAT and the user is done through the GUI. Advanced users have the alternative of calling the functions from the MATLAB command window. As an example, we will show how to use the toolbox to analyze a AVIRIS image. This image was acquired from AVIRIS, a high resolution hyperspectral imager with 220 spectral bands (.4 to 2.5 m) and a 2-20m resolution (depending of the altitude). The area of interest is shown in Figure 3. This image covers some area of using a portion of an AVIRIS data set taken over NW Indiana s Indian Pine test site in June 1992 [13]. With this image, we will give an example of how to use the HIAT to identify different targets characteristics. The first step is to load an image file. Once loaded, the image can be viewed as shown in Figure 4. The visualization of the

7 BOOK CHAPTER 7 Fig. 3. AVIRIS Image of Indiana s Indian Pine and Ground Truth Map (taken from [13]) image can be done band by band in a grayscale color map. There is a scroll bar for the user to browse the bands. If the user wants to visualize an RGB composite of the image, users can set the desired RGB bands for visualization. We mark the Figure 4 with different labels to guide trough this tutorial. The label 1 indicates the loaded file name, label 2 shows number (or number and wavelength) of the band displayed, the label 3 indicates the image loaded in HIAT, label 4 shows the scrolling bar where the users can move trough to look the desired band and the label 5 the users can select the visualization mode for the loaded image (greyscale, RGB composite or true color). By default the images visualization in HIAT are in greyscale as shown Figure 5(a), the user can change the visualization selecting the RGB Composite and True Color button. The result RGB composite are shown in Figure 5 (b). In addition, the RGB composite GUI the users can select two automatic band subset selection algorithms for the RGB composition. Figure 6 shows the image enhancement process with Resolution Enhancement. Figure 6 (a) shows the original image with the GUI to select one of the algorithms for resolution enhancement, also shows the cutoff frequency parameter for users selection to image enhancement spectrally as shown in Figure 6 (b). Once we have enhance the image, if the image have high dimensionality we can use one of the features extraction/selection algorithms to reduce the dimensionality of the HSI data. In this example we use SVDSS algorithm to reduce the data from 220 to 15 bands, notice in Figure 7 the scroll bar is bigger compare to Figure 6. Figure 7 shows the supervised classification GUI, the areas selected as testing and training data for the different classes are shown. In this example four classes where selected, corn,

8 BOOK CHAPTER 8 Fig. 4. Loaded Image grass, soybean-notill and soybean-mintill as shown in Figure 7. Once testing and training areas are selected for each class, users can save the areas in a file for future reference. Finally, the users select the desired classifier to analyze the results obtained from the input data. Figure 8 (a) shows the results obtained using the Maximum Likelihood distance classifier. The training/testing accuracy are shown in Table II. HIAT provides post processing techniques to enhance the results of the obtained classification map. In the toolbox we have integrated Supervised and Unsupervised Extraction and Classification of Homogeneous Objects (ECHO) to smooth the resulting classifica-

9 BOOK CHAPTER 9 Fig. 5. (a) Greyscale view (b) RGB view tion maps. Users have the option of selecting the different window size for ECHO. Figure 8 (b) shows the results obtained using ECHO with a window size of 3x3. Clearly you can notice the smoothing effect removing possible misclassified pixels and creating more clear regions. TABLE II Supervised Classification Training/Testing Accuracy Training Samples Accuracy Testing Samples Accuracy Soybean-notill Corn-notill Grass Soybean-mintill Total Once the user select the unmixing algorithm from the classification menu a GUI similar to Figure 9 (a) will open, in this GUI the users can supplied the endmembers to the algorithm. For this example we load a hyperspectral image Enrique Reef Lajas, Puerto Rico. This image was acquire with Hyperion this image is supplied with the toolbox. The classes of

10 BOOK CHAPTER 10 Fig. 6. (a) Original Image (b) Image Enhance (using Resolution Enhancement) Fig. 7. Supervised Classification GUI

11 BOOK CHAPTER 11 Fig. 8. (a)maximum Likelihood Classmap (b) ECHO 3x3 Post Classification Fig. 9. (a) Unmixing GUI (b) Unmixing Results the area are sand, reef head, deep water, sea grass and mangrove. The users can select the endmembers manually from the image (as this example) or loaded from a file, the endmembers selected are shown in Figure 9 (a). Once the unmixing process is done a GUI similar to Figure 9 (b), it show the different abundance maps. The users can click over the endmembers labels to study the abundance map results or can save the results using the file menu. Figure 10 shows the abundance map results, where red indicate more concentration of the endmember in the pixel and blue no concentration.

12 BOOK CHAPTER 12 Fig. 10. NNSLO Abundance Estimates (a) Original image (b) Deep water (c) Sand (d) Sea grass (e) Reef head (f) Mangrove IV. Other Functionalities The toolbox includes others functionalities useful in multi/hyperspectral image processing such as class statistics, 3D visualization, pixel spectral response and image cropping. The statistic GUI gives the users the option of view the means and the correlation matrix for each of the classes of a previous classification. Figure 11 (a) shows and example of the statistics GUI. Figure 11 (b) shows the visualization cube of the loaded data, as it shows users can rotate the cube to view a different angle of the data cube. In addition, users have the option of viewing the spectral response of a single pixel showing also the spectral response of the neighborhood and the average. Figure shows an example of the pixel response tool. Finally if the users need some kind of guide or example, the toolbox has an online help to introduce the users to the use and functionalities of HIAT. Figure 12 shows the online help. V. Final Remarks A MATLAB Toolbox for Hyperspectral Image Analysis was presented. This toolbox is intended for researchers doing multi/hyperspectral image analysis in various fields. HIAT provides a unified framework that gives support to these different disciplines. This framework facilitates the dissemination of research and development working new multi/hyperspectral image analysis algorithms.

13 BOOK CHAPTER 13 Fig. 11. (a) Classes Statistics (b) 3D Image Visualization References [1] Research Systems Inc., ENVI, The environment for visualizing images, url: [2] Landgrebe, D., Biehl, L., MultiSpec, image spectral analysis url: biehl/multispec/description.html. [3] Vlez-Reyez M. Jimnez L. Subset Selection Analysis for the Reduction of Hyperspectral Imagery, Geoscience and Remote Sensing Symposium Proceedings, IGARSS 98. IEEE International Volume 3, pp , [4] Arzuaga-Cruz, E., Jimenez-Rodriguez, L. O., and Velez-Reyes, M. Unsupervised Feature Extraction and Band Subset Selection techniques based on Relative Entropy Criteria for Hyperspectral data Analysis, Algorithms and Technologies for Multispectral, Hyperspectral, and Ultraspectral Imagery IX. SPIE Proceedings Volume 5093, pp , [5] Rivera-Medina, J. L., Development of an Unsupervised Extraction and Classification of Homogeneous Objects for Hyperspectral Images, CpE Masters Thesis, University of Puerto Rico at Mayagez, PR, [6] I. Sommerville, Software Engineering, 5th edition, Addison-Wesley, [7] Kaeli, D., et al., The Multi-view Tomography Toolbox, url: Center for Subsurface Sensing and Imaging Systems, Northeastern University, Boston, MA, [8] Fukunaga, K., Introduction to Statistical Pattern Recognition, 2nd edition, San Diego, CA, [9] Richards, J.A., Remote Sensing Digital Image Analysis, An Introduction, 3rd Edition, New York: Springer [10] Ifarraguerri, A. and CHIATng, C., Unsupervised Hyperspectral Image Analysis with Projection Pursuit, IEEE Transactions on Geoscience and Remote Sensing, vol 38, no. 6, pp , [11] Arzuaga-Cruz, E. Jimnez-Rodrguez, L., Vlez-Reyes, M. and et. al. A MATLAB Toolbox for Hyperspectral Image Analysis. In Proceedings of International Geoscience and Remote Sensing Symposium, September [12] Laracuente, J., Hunt, S. Determining noise in hyperspectral imagery for the application of oversampling to supervised classification. Master in Science Thesis, University of Puerto Rico at Mayagez, PR, 2005.

14 BOOK CHAPTER 14 Fig. 12. HIAT help window [13] Tadjudin S. and Landgrebe, D. A. Robust Parameter Estimation for Mixture Model. IEEE Transactions on Geoscience and Remote Sensing, Vol. 38, No. 1, pp , January [14] Samuel Rosario-Torres, Miguel Vélez-Reyes An algorithm for fully constrained abundance estimation in hyperspectral unmixing. Algorithms and Technologies for Multispectral, Hyperspectral, and Ultraspectral Imagery XI, June 2005