Comparison of a Hyperspectral Classification Method Implemented in Different Remote Sensing Software Packages

Transcription

1 Comparison of a Hyperspectral Classification Method Implemented in Different Remote Sensing Software Packages A Study Based on a CHRIS/PROBA Dataset in Switzerland Benedicte Odden Diploma Thesis submitted to the Department of Geography University of Zurich Faculty representation: Prof. Dr. K. I. Itten Supervision: Dr. Mathias Kneubühler Zurich, February 2008

2

3 Preface Preface At this point I would like to thank all of the people who have helped and supported me in any way throughout this thesis. Fist of all I would like to thank Prof. Dr. Itten for giving me the opportunity to do this thesis at the Remote Sensing Laboratories. I would also like to give a special thanks to Mathias Kneubühler for helping me choose the topic of this thesis and for helping and supporting me through every step of the work process. I also thank him for the corrections of this thesis. I thank all the people at RSL who have helped me in any way, among these, especially Tobias Kellenberger and Yves Bühler. Further, I would like to thank the girls for all the help they have given me and for funny coffee breaks. Tusen takk til Mauro, mamma, pappa og Cecilie for all støtten dere har gitt meg på alle mulige måter.

4

5 Content Content Content...I Figure Index...III Table Index...IV Abbreviation Index... V Summary... VII 1 Introduction Problem Definition Objective RSL-Test Site Outline Background Imaging Spectroscopy CHRIS/PROBA Spectral Angle Mapper (SAM) Algorithm Statistical Measures Overall Accuracy Producer s Accuracy User s Accuracy Kappa Coefficient Data and Software Data CHRIS/PROBA Dataset Ground truth Data ADS40 Dataset Software ENVI (Environment for Visualizing Images) Geomatica Data Preparation CHRIS/PROBA ENVI Metadata MNF-Transformation Geomatica Metadata MNF-Transformation Endmember Selection Background Introduction Convex Geometry Pixel Purity Index Applied Endmember Selection Ground Truth I

6 Content ENVI Spectral Hourglass Wizard (SHW) Regions of Interest (ROI) Pixel Purity Index (PPI) Geomatica Bitmaps ROI-Import Pixel Purity Index (PPI) Endmember Valuation Spectral Angle Mapper (SAM) Classification ENVI ROI-Classes Spectral Hourglass Wizard (SHW) SAM Classification Geomatica ROI-Classes from ENVI Spectra Extraction Spectra Plot Regions of Interest (ROI) SAM Classification Results ENVI Geomatica Accuracy Assessment Conclusions Concluding Observations Outlook Glossary Bibliography Appendix CHRIS/PROBA nadir data set acquired on August 17 th Complete table of original ground truth data ENVI Header File Abbreviation of the required format of an image metadata XML document CHRIS/PROBA XML File ENVI Pixel Purity Index Image for CHRIS/PROBA Dataset II

7 Figure Index Figure Index Fig. 1.1: Subset of CHRIS/PROBA Nadir Scene from the village St.Urban... 2 Fig. 2.1: Electromagnetic spectrum (modified Albertz, J., 2001.)... 5 Fig. 2.2: The imaging spectroscopy concept (modified Vane, G., 1993) Fig. 2.3: Plot of a reference spectrum and a test spectrum for a two-band image... 7 Fig. 3.1: Spectral Hourglass Wizard Flowchart (ENVI Help) Fig. 3.2: Hyperspectral Analysis Package Flowchart Fig. 4.1: Eigenvalue plot of CHRIS/PROBA data set from ENVI for all 37 bands Fig. 4.2: Spatial Coherence Value of CHRIS/PROBA dataset. Threshold: 0.02, Number of bands over threshold: Fig. 5.1: Spectral mixing analysis. (a) Simplex. (b) Endmembers Fig. 5.2: Pixel Purity Index Plot of CHRIS/PROBA dataset Fig. 5.3: Spectral Profiles of Endmembers valuated as very good Fig. 5.4: Spectral Profiles of Endmembers valuated as good Fig. 5.5: Spectral Profiles of Endmembers valuated as acceptable Fig. 6.1: Endmember spectra from SPL-file Fig. 7.1: ENVI SAM classification result Fig. 7.2: Geomatica SAM classification result III

8 Table Index Table Index Tab. 3.1: CHRIS specification for Mode 5 (Kneubühler, M., et al., 2006) Tab. 3.2: CHRIS image acquisition geometry [ ] for the 17th of August 2005 scene (negative values for backscatter angles), *sensor to target direction; agricultural test sites not covered (Modified table from Kneubühler, M., et al., 2006) Tab. 4.1: Required and optional Global Metadata (Geomatica Focus User Guide-Geomatica 10) Tab. 4.2: Required and optional Band-Specific Metadata (Geomatica Focus User Guide- Geomatica 10) Tab. 5.1: Modified ground truth with land use classes qualified for CHRIS/PROBA data set Tab. 5.2: Applied endmembers for SAM classification of CHRIS/PROBA dataset from August 17 th Tab. 5.3: Applied endmembers with valuation flags Tab. 7.1: ENVI Random Sample Listing Tab. 7.2: Geomatica Random Sample Listing Tab. 7.3: ENVI Error (Confusion) Matrix Tab. 7.4: Geomatica Error (Confusion) Matrix Tab. 7.5: ENVI Accuracy Statistics Report, Overall Accuracy: 71.80% Tab. 7.6: Geomatica Accuracy Statistics Report, Overall Accuracy: 71.80% IV

9 Abbreviation Index Abbreviation Index ENVI RSL CHRIS PROBA SAM MNF VIS NIR SWIR ESA CSES CIRES FZA GIS SHW MTMF GDB DEM HAP RMSE PC BSQ HDR Environment for Visualizing Images Remote Sensing Laboratories Compact High Resolution Imaging Spectrometer Project for On-board Autonomy Spectral Angle Mapper Minimum Noise Fraction (ENVI) Visible light Near Infrared light Short-Wave Infrared light European Space Agency Center for the Study of Earth from Space Cooperative Institute for Research in Environmental Studies Fly-by Zenith Angle Geographical Information System Spectral Hourglass Wizard Mixture Tuned Matched Filtering Generic Database Digital Elevation Model Hyperspectral Analysis Package Root Mean Square Error Principal Components Band Sequential format Header file format V

10 Abbreviation Index XML PIX MNFNR MNFLT MNF PPI ROI IEA Extensible Mark-up Language PCI Geomatica file format Maximum Noise Fraction Noise Removal Maximum Noise Fraction Linear Transformation Maximum Noise Fraction (Geomatica) Pixel Purity Index Regions of Interest Iterative Error Analysis VI

11 Summary Summary The two software packages ENVI and Geomatica, which are used for digital image processing at the Remote Sensing Laboratories (RSL), contain several classification methods. Recently a great deal of advances in the Geomatica Hyperspectral Analysis Package (HAP) has been made. The implemented method for land use classification Spectral Angle Mapper (SAM) was applied, compared and characterized in both programs, using a hyperspectral nadir image data set from the European sensor CHRIS on the satellite PROBA-1. The data set was acquired over the RSL-test site Vordemwald on August 17 th SAM is a tool designed to classify hyperspectral image data, on the basis of a set of reference spectra that define the classes. These classes were defined with the help of ground truth data which were collected in parallel to the image data set acquisition. The ground truth data contain different types of land use classes, among which not all were qualified for the CHRIS/PROBA data set. The choice of endmembers (reference spectra) was of great importance to obtain a good classification result. The land use fields which are contained in the ground truth data were selected by considering whether they were located inside the image frame and whether they were large enough to be used as samples. Several classes were also defined by optical means. The process of endmember selection was not easy considering that the CHRIS data set has a spatial resolution of only 18 meters which means that the image data are very heterogeneous. The endmembers which were defined for the classification in ENVI were used in Geomatica as well. In that way possible differences in the implementation of the two programs could be found easier. The classification workflow in the two software packages shows many differences. The hyperspectral classification tool Spectral Hourglass Wizard (SHW), being integrated in ENVI, contains all the work steps for hyperspectal classification methods and also contains dialog boxes where every step of the classification is explained and hints are given. This fact turned out to be very helpful. The Hyperspectral Analysis Package (HAP) which is contained in Geomatica is also a very useful tool for a hyperspectral classification, but here the work steps are only listed in the Geomatica help-function and not integrated into a workflow, like in ENVI. The classification procedure in Geomatica was therefore not as fluent as it was in ENVI. Not only the implementation of the two classification tools differed from each other, but also some of the work steps were either different or they were non existent in one or the other tool. Most work steps, however, were the same in both classification tools. Despite of all these differences, comparing the classification results, there were no differences visible at first sight. Both result images show some overlapping between the different land use classes, mostly those with spectra similar to each other, but these effects were identical in both classifications. To be able to statistically differentiate between the two classification results an Accuracy Assessment was preformed. This task was found in Geomatica and was used for both classification images. The Accuracy Assessment determines the correctness of classified images. The measure of accuracy is the correlation between a standard that is assumed to be correct and an image classification of unknown quality. This standard was defined by a VII

12 Summary number of verification samples contained in the ground truth data. The Accuracy Assessment produces several different reports which contain statistical measures of the differences between the classification images and the reference data (verification samples). All of the Accuracy Assessment measures contained in these reports were identical for both classifications. Based on these obtained results one can assume that the implementation of the SAM algorithm in ENVI and Geomatica is equal. The Accuracy Assessment measures were in most cases satisfying, considering the difficulties of endmember selection due to the low spatial resolution and heterogeneity of the dataset. VIII

13 Introduction 1 Introduction 1.1 Problem Definition Remote sensing is practiced by the examination of features as observed in several regions of the electromagnetic spectrum. Hyperspectral remote sensing, also known as imaging spectroscopy, provides more detailed information than multispectral imaging; in this way we are able to identify and differentiate spectrally unique materials. Recent advances in remote sensing and developments in hyperspectral sensors have allowed these types of data to be used by people other than spectral remote sensing experts. Improvements were not only made in technology for acquiring such data, advances have also been made in analysis techniques. The range of disciplines that can use this information has grown from, once mainly geological applications, to include ecology, atmospheric science, forestry and agriculture to name a few (PCI Geomatics Hyperspectral Data Primer). The two software packages ENVI and Geomatica, which are used for digital image processing at the Remote Sensing Laboratories (RSL), contain several classification methods. Recently a great deal of advances in the Geomatica Hyperspectral Analysis Package has been made. In this thesis the implemented methods for land use classification in both programs are applied, compared and characterized, using a hyperspectral image data set from the European sensor CHRIS on the satellite PROBA-1. The classification method Spectral Angle Mapper (SAM) which is implemented in both Geomatica and ENVI was chosen to compare the two programs. SAM is a tool designed to classify hyperspectral image data, on the basis of a set of reference spectra that define the classes. These classes are set with the help of ground truth data that were collected in parallel to the image data set acquisition. Another possible classification method would have been the Spectral Unmixing approach. 1.2 Objective The objective of this thesis is to compare an implemented method for land use classification in the two programs ENVI and Geomatica, using a hyperspectral nadir CHRIS/PROBA image data set from August 17 th 2005, acquired over the RSL test site Vordemwald. For the land use classification in both programs, the Spectral Angle Mapper method was applied. The SAM algorithm permits a rapid classification and mapping of the spectral similarity of image spectra to reference spectra (Kruse and Boardman 1993), but its functionality is not the main focus of this thesis. The choice of endmembers (reference spectra) is of great importance for a reliable classification. In this thesis they are set by means of ground truth data that were collected in parallel to the image acquisition. The same endmembers are used for ENVI as well as Geomatica. In this way possible differences in the implementation of the two programs can 1

14 Introduction be found easier. Differences in results, if they occur, are to be discussed and possible solutions presented. 1.3 RSL-Test Site The test site Vordemwald is located in central Switzerland (7 53 E, N). The lower parts of the area are dominated by agricultural fields and the hilltops mainly consist of mixed forests (elevations up to 700 m a.s.l). The agriculture concentrates above all on barley, wheat, maize, sugar beet and pasture land. CHRIS multiangular data sets were acquired over the test site on eight different dates between the 26 th of May 2005 and the 22nd of September 2005 (Kneubühler, M. et al., 2006). Out of these data sets, only one date was used for this thesis. The selected date is the 17 th of August 2005 and only the nadir image was used. Ground truth data were collected in parallel to the CHRIS data takes on most dates. On the 17 th of August 2005 the ground truth data were collected in the area around the village of St. Urban. This village and the area around it is therefore the investigation area in this thesis. Fig. 1.1: Subset of CHRIS/PROBA Nadir Scene from the village St.Urban. 1.4 Outline Following the objectives mentioned above, the outline of this thesis is arranged as follows: In chapter 2 the basics of imaging spectroscopy, the satellite and the sensor used for the image acquisition, and the Spectral Angle Mapper algorithm are explained. 2

15 Introduction In chapter 3 the data and software used in this thesis are described. Hereby the acquisition of the image data set as a part of an RSL research activity is further explained, and the software origin and work flows are described. In chapter 4 the data preparations that needed to be made are presented. Among these preparations the importance of metadata for hyperspectral image datasets is brought out, and the approach used in both software packages is described. The interest of running an MNF- Transformation and the necessity of such an operation for the image data set used in this thesis are discussed for both of the software packages. In chapter 5 the hard and crucial procedure of endmember selection is explained. Each step of the work flow and its importance is analysed, and important basics on different functions relevant to these work steps are pointed out. The work flow for both software packages is described in detail, and the results of these work steps are valuated. Chapter 6 concentrates on the classifications performed by the two software packages. Since the classification work flow was already described in detail in chapter 3, chapter 6 describes only the last step; the SAM classification. Hereby the applied endmembers are mentioned once more along with the functions needed to prepare them for the classification. The results of the classifications are displayed in chapter 7. Among these results, the classification images from both software packages are demonstrated and also compared, using an accuracy assessment function. Conclusions on the results are made in the end of the chapter. Chapter 8 sums up the findings of this thesis as a brief repetition. In chapter 9 a small outlook discusses possible improvements that could be made to achieve better results. 3

16 Background 2 Background 2.1 Imaging Spectroscopy To define the concept of imaging spectroscopy one must first define some of the basic principles of physics. Among these principles, above all, the electromagnetic spectrum is of importance. The electromagnetic spectrum is not only important for imaging spectroscopy, but for all remote sensing tasks. The atmospheric transmission of electromagnetic radiation is particularly high in the areas of visible 1 (VIS), near infrared 2 (NIR), and short-wave infrared (SWIR) 3 light (see figure 2.1). Areas other than these are frequently used in remote sensing, but for most applications in imaging spectroscopy only these three areas are of special interest (Goetz, 1995 ). Electromagnetic radiation is a self propagating wave in space with electrical and magnetic components. The electrical components, or the electrical field, vary in magnitude in a direction perpendicular to the direction of propagation. The magnetic field is oriented at perpendicular angles to the electrical field and is propagated in phase with the electrical field. Both fields propagate with the speed of light c 4. The intensity of the electrical and magnetic fields change sinusoidal during this phase, and defines the wavelength λ 5. The frequency ν describes the number of oscillations made in the sinusoidal propagation during one second (entity: Hz). The product of the wavelength λ and the frequency ν in a vacuum is the speed of light c: c= ν λ A large wavelength thus means that the frequency is low. A small wavelength on the other hand shows a high frequency. 1 Visible light is the range of the electromagnetic spectrum that is visible to the human eye (400 nm to 700 nm). 2 Near infrared light equals the wavelength from 700 nm to 1300nm. 3 Short wavelength infrared light range between 1300 nm to 3000 nm. 4 c = m/s 5 λ is the distance from one wave crest to the next. 4

17 Background Fig. 2.1: Electromagnetic spectrum (modified Albertz, J., 2001.). The main source of electromagnetic radiance is the sun. The solar radiation maximum is found between 400 nm and 500 nm. The intensity of solar radiation is higher outside the Earth s atmosphere than it is on the Earth s surface. The reasons for these differences are several physical interactions between the solar radiation and the Earth s atmosphere and surface. To be able to acquire remote sensing data, it is important that the land surface and the objects which are present on this surface react differently with the incoming radiation. The characteristic reflections made by different land surface objects depend on the type of material and of the physical condition (e.g. humidity) of the object. The surface finish and the geometrical settings (e.g. angle of solar incidence, view angle) are also important factors that make an instant rendering of the land surface objects in image data possible (Albertz, J., 2001). Fig. 2.2: The imaging spectroscopy concept (modified Vane, G., 1993 ). 5

18 Background The objects can be detected and quantified with a remote sensor. Depending on what is trying to be achieved with the data detected, there are different types of sensors that can be used. Optical sensors are used in the reflective part of the electromagnetic spectrum. Multispectral systems do not cover the spectrum contiguously and their bands are generally wide (70 to 400 nm). Usually they have a dozen or less bands and are sensitive in the range of 400 nm to 2500 nm depending on the sensor. Hyperspectral systems differ from multispectral sensors because they collect information in many contiguous narrow bands (5 to 10 nm). Hyperspectral images in the reflective part of the spectrum normally contain dozens to hundreds of bands between 400 nm and 2500 nm, depending on the sensor. Hyperspectral imaging provides more detailed spectral information than multispectral imaging. In that way spectrally unique materials can be differentiated and identified. Depending on the application discipline, certain ranges of the electromagnetic spectrum are more helpful than others. Plants for example have distinct spectral features in the visible and near infrared range. Minerals on the other hand show such features mainly in the SWIR range (Albertz, J., 2001). 2.2 CHRIS/PROBA CHRIS/PROBA is a spaceborne mission operated by the European Space Agency (ESA). The Project for On-Board Autonomy; PROBA-1, was launched on October 22nd, 2001 from Shriharikota in India. The platform carries a number of scientific instruments of which the most advanced one is CHRIS (Compact High-Resolution Imaging Spectrometer). The CHRIS instrument provides hyperspectral and multidirectional data of the Earth s surface. While the spectral information contained in the CHRIS data is able to assess the biochemistry of vegetation canopies, the directional information can describe the structure of the observed canopy (Kneubühler, M., et al., 2006). CHRIS operates in a push-broom mode and is capable of delivering highresolution spectra throughout the visible and NIR wavelengths (400 nm to 1050 nm). It acquires high spatial resolution (17-20 or m) images of the Earth s surface in up to 62 narrow spectral channels 1, depending on which operation mode the sensor is acquiring the image data. Mode 1: In this mode, CHRIS acquires image data across full swath (nominal 13 km) at reduced spatial resolution (nominal 34 m), in 62 spectral channels. Mode 2-4: This are the most common modes of operation, in which CHRIS is used to acquire image data at the highest spatial resolution (nominal 17 m) over the full swath and in 18 spectral channels. Mode 5: CHRIS can also be used to acquire image data across half of the nominal swath at the highest nominal spatial resolution. In this mode, which was used for this thesis, the instrument is able to record image data in 37 spectral channels with a spatial resolution of 18 meters (Barnsley, J., et al., 2004). 1 Located in the visible and NIR wavelengths. 6

19 Background PROBA-1 is highly manoeuvrable so that CHRIS can image a given site five times during a single overpass pointing along-track (+55, +36, 0, -36, -55 ), while across-track pointing ensures that the revisit time for a site of interest is less than a week. 2.3 Spectral Angle Mapper (SAM) Algorithm The Spectral Angle Mapper algorithm was developed by J.W. Boardman of the Center for the Study of Earth from Space (CSES), Cooperative Institute for Research in Environmental Studies (CIRES) and the Department of Geological Sciences, University of Colorado, Boulder. The algorithm is the base of the SAM programs in both ENVI and Geomatica. The SAM is a tool that permits rapid mapping of the spectral similarity of image spectra to reference spectra (Kruse and Boardman, 1993). The reference spectra can either be a spectrum measured in a laboratory, a field spectrum or it can be extracted directly from the image. The SAM algorithm determines the spectral similarity between two spectra, treating them as vectors in a space with dimensionality equal to the number of bands (Kruse, F. A., et al., 1993). This can be easier explained in figure 2.3, where a reference spectrum and a test spectrum are represented as two-band data in a two-dimensional plot. Fig. 2.3: Plot of a reference spectrum and a test spectrum for a two-band image (Kruse, F. A., et al., 1993). As mentioned above, the algorithm assumes that the reference spectrum represents the same wavelength sampling as the image (i.e., the number of measurement values in the reference spectrum equals the number of bands in the image, and the band centre wavelength values match the wavelength coordinates of the reference spectrum measurement values). The angle (α) between the test spectrum (T) and the reference spectrum (R) is computed as follows (Geomatica v10 Help ): 7

20 Background ()= i T test spectrum values (i = to n) R (i) = reference spectrum values (i = 1 to n) T R Angle ( ( T() i T() )) ( ( R() i R() )) = i = i ( ) ( T R )) ( α ) ar cos ( T( i) R( i) ) = This measure of similarity is insensitive to gain factors because the angle between two vectors is invariant with respect to the lengths of the vectors. Due to this fact laboratorymeasured material spectra can be directly compared to reflectance spectra from hyperspectral images, which have an unknown gain factor related to topographic illumination effects. The angle between each reference spectrum and each test spectrum is computed. The SAM program, in both ENVI and Geomatica, assigns the angles to output channels (rule images), then each pixel is assigned to the class defined by the reference spectrum. The class assigned to each pixel is saved in the output channel. 2.4 Statistical Measures The statistical measures presented below are a part of the Accuracy Assessment which is explained in chapter 7.3. The measures are used to statistically differentiate between the classification results achieved in ENVI and Geomatica Overall Accuracy The statistical measure Overall Accuracy (Ao) is the percentage of correctly classified pixels in a classification image. n χaa a= Ao = 1 N, where χ aa = correctly classified pixels N = all pixels This measure is simple, but it is not very meaningful because the commission and omission errors are not taken into consideration. It is therefore important that a high Overall Accuracy is combined with a high accuracy of each class (Congalton, R.G., et al., 1998). 8

21 Background Producer s Accuracy The Producer s Accuracy (Ap) is a measure of the accuracy of a particular classification scheme. It shows what percentage of a particular reference class (b) was correctly classified. It is calculated by dividing the number of correct pixels for a class by the actual number of reference pixels for that class (ab). Ap = χbb 100 i χab a= 1, where χbb = correctly classified pixels in class (b) χab = actual number of reference pixels for class (b) In this process the measure does not consider the pixels which are falsely assigned to class b (the commission errors) (Congalton, R.G., et al., 1998) User s Accuracy Another measure of accuracy is the User s Accuracy (Au) which shows what percentage of a class (b) was correctly classified. It is calculated by dividing the number of correct pixels for a class by the number of classified pixels for that class (ba). Au = χbb 100 i χba a= 1, where χ bb = correctly classified pixels in class (b) χba = Actual numbers of classified pixels for class (b) In this case the measure does not take the errors of omission (pixels are not identified by the class) into consideration (Congalton, R.G., et al., 1998) Kappa Coefficient The Kappa Coefficient is a statistical measure of the agreement, beyond chance, between two maps (e.g. output map of classification and ground-truth map). It is represented by the symbol (kappa hat). Correctly assigned pixels may have been assigned by chance and not based on the classification decision rule (SAM). The kappa value ( ) indicates how accurate the classification output is after this chance, or random, portion has been accounted for. The equation expresses kappa in terms of Overall Accuracy (Ao) and chance agreement (θ). 9

22 Background ˆ Ao θ = 1 θ K, where n χai χia a= 1 θ = 2 is the chance agreement, N and χ ai = Number of observations made in columns of error matrix χ ia = Number of observations made in rows of error matrix While calculating the Kappa Coefficient the chance agreement is eliminated and then, by given chance agreement, normalized through its maximal value. In this way the highest possible kappa value will always be equal to 1 (Congalton, R.G., et al., 1998). 10

23 Data and Software 3 Data and Software 3.1 Data The data set used for the classification in this thesis was acquired by the sensor CHRIS on the platform PROBA CHRIS/PROBA Dataset Between the 26 th of May and September 22nd, 2005, numerous multiangular data sets were acquired over the RSL-test site Vordemwald with the sensor CHRIS in Mode 5 (Tab.3.1). Spatial Sampling km altitude Image area View angles Spectral bands 6.5x13 km 5 nominal 37 bands (372x748 with 6-33 pixels) +55, +36, nm width 0, -36, -55 Tab. 3.1: CHRIS specification for Mode 5 (Kneubühler, M., et al., 2006). Spectral range nm Out of the data sets acquired, four dates were selected, representing major steps in phenology of the agricultural fields in the test site. The selected dates are May 26 th, June 20 th, August 17 th and September 22nd. These image data were acquired for another study performed at RSL, on vegetation growth using multitemporal data (Kneubühler, M., et al., 2006), which is, however, not part of this thesis. As explained in chapter 1.2, the objective of this thesis is to compare two classification programs contained in the software packages PCI Geomatica and ENVI using CHRIS-data. The information on phenology acquired in these multitemporal data sets is therefore not needed to achieve this goal. For this reason only the nadir data set from the 17 th of August was chosen. Information concerning the viewing geometries of this selected data set is given in Tab /08/05 FZA FZA FZA FZA FZA Solar Zen./az Obs. Zen. Obs Az.* Tab. 3.2: CHRIS image acquisition geometry [ ] for the 17th of August 2005 scene (negative values for backscatter angles), *sensor to target direction; agricultural test sites not covered (Modified table from Kneubühler, M., et al., 2006). Geometric and atmospheric corrections of the multiangular CHRIS data set were preformed previously to this study (Kneubühler, M. et al, 2005). The geocorrection is based on a 3D physical model developed by Toutin which is implemented in the software Geomatica from PCI (Toutin, T., 2004). Atmospheric correction of the CHRIS radiance data products was 11

24 Data and Software preformed using ATCOR-3 (Richter, R., 1998), which is based on MODTRAN-4 (Berk, A., et al., 1989). The remaining necessary data preparations of the CHRIS nadir scene used in this thesis are described in chapter Ground truth Data Ground truth data were collected in the field parallel to the CHRIS data takes on most dates, including the 17 th of August The ground truth consists of several types of land use classes present within the RSL-test site Vordemwald (a table with complete ground truth information can be found in appendix p.58). These classes were used to set samples for the classification in the CHRIS nadir scene. Not all of the land use classes could be utilized for this work, due to different reasons; since the image spatial resolution of the CHRIS scene is not higher than 18 meters, many of the data collected came from fields that were relatively small and could therefore not be separated from other fields (they only contained impure pixels). They could not be used for the classification. Other reasons and details will be further explained in chapter ADS40 Dataset An ADS40 data set acquired on the 17 th of August 2005 over the RSL-test site in Vordemwald parallel to the CHRIS data takes was used as an additional ground truth. ADS40 data have a very high spatial resolution (25cm), and were used to locate fields with different land use classes visually. In that way more training samples could be defined to make the classification as accurate as possible. Further descriptions on the necessity of these samples can be found in chapter Software At RSL the two software packages ENVI and Geomatica are available for digital image processing. The implemented method, SAM, for land use classification in both programs is to be applied, compared and characterized ENVI (Environment for Visualizing Images) The Environment for Visualizing Images (ENVI) program was developed by the American company ITT Visual Information Solutions. ENVI's complete image-processing package includes advanced spectral tools, geometric correction, terrain analysis, radar analysis, raster and vector GIS capabilities and extensive support for images from a wide variety of sources. The program s approach to image processing combines file-based and band-based techniques with interactive functions. Opening a data input file, its bands are stored in a list where they can be accessed from all system functions. Its interface is complemented by a comprehensive 12

25 Data and Software library of processing algorithms. ENVI includes all the basic image processing functions, and it does not impose limitations on the number of spectral bands that can be processed, so both multispectral and hyperspectral data sets can be handled (ENVI user s Guide ). The Spectral Hourglass Wizard (SHW) is an application integrated in ENVI to guide the user step-by-step through the ENVI hourglass processing flow to find and map image spectral endmembers from hyperspectral data. The wizard displays detailed instructions and useful information for each function (see figure 3.1). Fig. 3.1: Spectral Hourglass Wizard Flowchart (ENVI Help). The hourglass processing flow uses the spectrally over-determined nature of hyperspectral data to find the most spectrally pure or spectrally unique pixels (endmembers) within the data set and to map their locations and sub-pixel abundances. This processing flow begins with reflectance or radiance input data. The data can be spectrally and spatially subset, visualized in n-d space and the purest pixels can be clustered into endmembers. Optionally, the endmembers can be supplied by the user, which is the way it was done in this thesis (ground 13

26 Data and Software truth). When the endmembers are located, the Spectral Analyst can be used to identify them and to review the mapping results. To perform a classification, one of three methods in the Mapping Method panel must be chosen. The methods integrated in the wizard are: SAM, the Mixture Tuned Matched Filtering (MTMF) and the Linear Spectral Unmixing. The SAM method, which was used for this thesis, produces a classified image based on the value the user specifies for a SAM Maximum Angle. Decreasing this threshold usually results in fewer matching pixels. Increasing this threshold may result in a more spatially coherent image; however, the overall pixel matches will not be as good as for the lower threshold (ENVI Help). This fact will be further explained in chapter 6.2. Each step in the Spectral Hourglass Wizard executes a standalone ENVI function available from the Spectral menu. The name of the corresponding ENVI function appears at the top of the panel during each step (see figure 3.1) Geomatica Geomatica is a software package developed by the Canadian corporation PCI Geomatics. It includes a series of integrated work environments; Focus, OrthoEngine, Modeler and EASI. Each environment is linked by a data translation feature. Underlying all PCI Geomatics technologies is the Generic Database Access layer (GDB). It acts as a transparent interface between PCI applications and different data sources, enabling reading, writing, and transforming of more than one hundred raster and vector file formats (PCI Geomatics - Geomatica 10 Brochure, 2006). For this thesis the working environment Focus in Geomatica version 10 was used. Focus is Geomatica s main data visualization environment. It integrates technologies for Remote Sensing, Image Processing, GIS/Spatial Analysis, and Map Publishing into a single environment. Geomatica Focus can read various geospatial data formats 1, and an Algorithm Librarian, with over 300 algorithms, is also available. This Algorithm Librarian contains all of the data processing functions for image filtering, data interpolation, image classification, spatial analysis, and DEM analysis 2. The Algorithm Librarian in Geomatica Focus lists the algorithms either by licensing or in application groupings. The algorithm used in this thesis (SAM) can either be found in the licensing group Hyperspectral Add-on or in the application grouping Analysis under Hyperspectral Analysis. All algorithms in both of these groupings are designed for processing and analysing images acquired with airborne and satellite imaging spectrometers. Under Technical references in the Geomatica Help, the Hyperspectral Analysis Package (HAP) can be found. The Package is designed for processing and analysing images acquired with airborne and satellite imaging spectrometers. However, here each step that needs to be made in this process is only described, not implemented like in the Spectral Hourglass Wizard in ENVI. This fact made the classification process in Geomatica more complicated than the one in ENVI. The classification workflow in Geomatica, which is based on HAP, is listed in the flowchart in figure 3.2. The work steps are described below. 1 Including vector, raster and ASCII data. 2 Digital Elevation Model. 14

27 Data and Software Fig. 3.2: Hyperspectral Analysis Package Flowchart. The Hyperspectral Analysis Package accommodates images with up to 1024 bands, and consists of the following components: PACE application programs A set of visualization operations accessible through Focus Spectral library files (.spl) The PACE application programs consist of different functions like Preprocessing, Atmospheric Correction, Local Analysis, Spectral Handling and Metadata I/O (In/Out). The functions needed for the data set used in this thesis were the Preprocessing, Local Analysis, and the Metadata I/O. Under the function Preprocessing the tasks sensor-related calibration, geometric correction, and noise removal are found. Since the CHRIS data set already had preprocessing done to it (c.f. chapter 2.2), a noise removal was the only task which was used. The noise removal task is further explained in chapter The Local Analysis function consists of three different programs: ENDMEMB, SAM, and SPUNMIX. The ENDMEMB program helps to locate endmembers in the dataset, and was not needed since the endmembers were already defined (ground truth). The two programs SAM and SPUNMIX (Spectral Unmixing) are based on 15

28 Data and Software different algorithms which classify hyperspectral image data. For this thesis the SAM classification method was chosen. It is further described in the chapters 2.3 and 6. To support the image data, various metadata about the mission and the sensor must be attached to the image file. With the function Metadata I/O, these mission and sensor data can be imported or exported. This function is described in detail in chapter Since the endmembers were already defined with the help of ground truth, the spectral library files were not needed. What was necessary though, was to extract the spectrum of each endmember into a Spectra Plot file (.spl) with the help of the Spectra Extraction Dialog box (this step is further described in chapter 6.3.2). This step is not integrated in the Hyperspectral Analysis Package. The last step is the SAM classification. The SAM function is integrated in the Hyperspectral Analysis Package along with the function Spectral Unmixing. These functions can be found in Geomatica Focus under Geomatica Hyperspectral. Here all of the functions explained so far can be found, as well as many others. The SAM classification step (the method itself was explained in chapter 2.3) is described in chapter

29 Data Preparation 4 Data Preparation 4.1 CHRIS/PROBA The CHRIS/PROBA data set had already been geometrically 1 and atmospherically 2 corrected 3, and therefore only needed a few pre-processing steps done in both ENVI and Geomatica. The geometrical correction Root Mean Square Errors (RMSE) for the specific region of interest do generally not exceed one pixel (Kneubühler, M., 2006). The RMSE for ground control points selected within a high topography test site of a different CHRIS nadir are reported 0.42 for X and 0.64 for Y using the same approach (Kneubühler, M., 2005). 4.2 ENVI Metadata As mentioned above in chapter about the Geomatica software, metadata of the mission and the sensor are needed to support the image file. This counts for ENVI as well, only here the metadata already exist in a header file (.hdr). The header file is always delivered together with the image file (.bsq 4 ) directly from the CHRIS/PROBA sensor. When the image file is used in ENVI, the header file must be stored in the same directory, and both files must have the same name e.g.: Vordemwald.bsq (image file) Vordemwald.hdr (header file (metadata)) When this is the case, the header file is read by ENVI automatically (the header file information is listed in the appendix p.59) MNF-Transformation In ENVI the Minimum Noise Fraction-Transformation (MNF) is integrated in the Spectral Hourglass Wizard. The MNF is used to determine the inherent dimensionality of image data, to segregate and equalize the noise in the data, and to reduce the computational requirements for subsequent processing (ENVI Help). 1 The Geocorrection is based on a 3D physical model developed by Toutin. 2 The atmospheric correction was preformed using ATCOR-3. 3 The approach of these corrections is described in M. Kneubühler, et al., Band Sequential file format. 17

30 Data Preparation The MNF uses two cascaded Principal Components (PC) transformations. The first transformation is based on an estimated noise covariance matrix, and decorrelates and rescales the noise in the data. In this way the data has no band-to-band correlations, and the noise has a unit variance. The second transform action is a standard PC transformation of the data that appear affected (striped or white) by the noise. The resulting bands of the MNF transformed data are ranked with the largest amount of variance in the first bands and decreasing data variance with increasing band number. The CHRIS/PROBA dataset had about 24 transformed bands out of 37 without noise, according to the ENVI MNF-Transformation in the Spectral Hourglass Wizard. This result can be read out of an eigenvalue plot, where the higher values indicate higher data variance (and may also indicate data dimensionality). When the values get lower and approach 1 only noise is left in the transformed band. This is the case, because the noise has been scaled into unity in each output MNF band. The dimension of the data can indicate the actual number of endmembers contained in the dataset since each linearly independent component adds another dimension to a spectral dataset through mixing. The data dimensionality can be determined from the eigenvalues by finding where the slope of the eigenvalue curve breaks and the values fall off to 1, as already mentioned above (Fig.: 4.1). Fig. 4.1: Eigenvalue plot of CHRIS/PROBA data set from ENVI for all 37 bands. The eigenvalues and MNF images determine the inherent dimensionality of the data. The MNF data space can be devided into two parts. One part contains the large eigenvalues and the coherent eigenimages, and the other part contains the low eigenvalues and the noisedominated images. To separate the noise from the data a threshold must be defined using only the coherent parts of the MNF data space. The SHW can estimate this threshold automatically (Fig. 4.2), or it can be entered manually if the user is not satisfied with the automatic result. 18

31 Data Preparation Fig. 4.2: Spatial Coherence Value of CHRIS/PROBA dataset. Threshold: 0.02, Number of bands over threshold: 24. The SHW is designed to give a reasonable default calculation of the data dimensionality, but this processing step is difficult and scene dependent. It is therefore much better to overestimate the data dimensionality and add a few extra bands, than to underestimate it and lose important information. The threshold level for the CHRIS/PROBA dataset was estimated to 0.02, with 24 bands above the threshold. Since the data set only has 37 bands, and the main objective of this thesis is to compare two software packages, and not necessarily to do a perfect classification, none of the 37 bands were left out. 4.3 Geomatica Metadata The CHRIS/PROBA data set used in this thesis consists of two files; the image file and the header file. The header file contains the metadata. Both files have the ENVI formats BSQ and HDR. To be able to use these files in Geomatica they have to be converted into PCI format (PIX and XML). The ENVI image file (BSQ) was converted into PCI format (PIX) with the help of an integrated function in ENVI. The metadata, however, had to be formatted as an Extensible Mark-up Language (XML) document in a text file. The XML document file must be in the same directory and have the same base name as the image file (PIX) e.g.: Vordemwald.pix (image file) Vordemwald.xml (metadata) Importing metadata is important when working with hyperspectral data because of the additional information about the dataset and the sensor used to acquire the data. These data must be attached to the image data to make the processing and analyzing more efficient, or even possible. Metadata is read into the PIX file using the Geomatica program METAIN. This program can be opened through the Algorithm Librarian in Geomatica Focus. The program 19

32 Data Preparation METAOUT will read the information in the metadata segment and export it as an XML document (Geomatica Focus User Guide-Geomatica 10). The image metadata scheme describes two basic types of image metadata items; global metadata and band-specific metadata. The global metadata applies to the image as whole. Many global metadata items are optional; others, on the other hand, are required. All these items are listed in table 4.3. Required Global Metadata Text data set descriptions Number of image bands Number and sequence of radiometric transformations stored as band-specific metadata Optional Global Metadata Name of the sensor model Name of the sensor type Location of image acquisition Time of image acquisition Heading of the platform Fore-aft sensor tilt relative to gravity vector Total sensor field-of-view Tab. 4.1: Required and optional Global Metadata (Geomatica Focus User Guide-Geomatica 10). Some band-specific metadata are also optional. The following table (Tab.4.2.) shows a list of both required and optional band-specific metadata. Required Band-Specific Metadata Optional Band-Specific Metadata Band ID number Band quality The numbers of the file channel that Radiometric transformation stores the band sequence Response profile Tab. 4.2: Required and optional Band-Specific Metadata (Geomatica Focus User Guide-Geomatica 10). The required format of an image metadata XML document is specified by an XML scheme stored in the PCI ImageMetadata.xsd file in $PCIHOME/etc (an abbreviation of this file is listed in the appendix p.60). With the help of this document and the information in the ENVI header file, an XML document for the Geomatica image file (PIX) was written and imported (see appendix p.62) MNF-Transformation There are several PACE programs in Geomatica that can be used for noise removal, thereof two of these programs are of interest; generate a maximum noise fraction noise removal (MNFNR) and generate a maximum noise fraction linear transformation (MNFLT). The MNFLT program is used to compute the parameters of band-wise linear data transformations and their inverses, and to apply them. This is part of the process for random noise removal. The MNFLT is designed in a manner that the generated parameters of the linear transformations are decoupled from their application to an image. The parameters generated by MNFLT are for principal components (PC) and maximal noise fraction (MNF) 20

33 Data Preparation linear transformations, respectively, and their inverses. The transformation parameters are written to a MATLAB format file. The motivation behind using a PC or MNF transformation is that the image noise that is spread across many bands of the original image will often be concentrated into a few bands of the transformed image. A noise band in the transformed image could be very aggressively smoothed, or it could even be replaced by a band with pixels all set to the mean pixel value of the noise band. The risk of information loss in the inverse-transformed result is very small. The PC transformation parameters are computed from the band-vector covariance matrix, whereas the MNF transformation parameters additionally require a noise-value covariance matrix. MNFLT accepts either an explicit noise image that models the type of noise that is to be removed from the original image, or it can derive an approximation for certain kinds of noise (e.g. image striping). This noise image approximation consists of a between-neighbour differences band for each input band (Geomatica Help). The PC transformation result shows the bands ordered in terms of decreasing image quality, though only when the noise in the original bands is uncorrelated and has the same variance over all these bands. The MNF transformation result has bands that are ordered in terms of increasing signal-to-noise ratio with respect to the modelled noise. For some special cases of the application of the MNF transformation to noise removal, the single-step implementation MNFNR is very practical. The MNFNR is applied if one member of a set of image bands is considered to have more noise than other bands. The band which contains the most noise is transformed so that its noise content is close to that of the other bands. When using the MNFNR method it is not necessary to know the noise variance in any of the bands to be able to define the MNF transformation. None of these methods were applied to the CHRIS/PROBA data set because of the results already achieved in ENVI (chapter 4.1.2). To be certain not to lose important information all 37 bands were kept. 21

34 Endmember Selection 5 Endmember Selection 5.1 Background Introduction An endmember is defined as a pure, and for the object of interest typical, spectrum. Because of sensor noise and the spectral variability of every object, the endmembers only exist as a convention or an idealisation. It is therefore essential to define what the object is. If the object class is e.g. grass, one must take into account that there might be parts of other objects mixed in with the grass, as e.g. soil or shadow, or both. This means that the definition of an endmember always depends on the observation scale. For this thesis ground truth data, which were collected in the field parallel to the CHRIS data takes, were used to define the endmembers. The ground truth is listed in the appendix on page 58 and the user defined endmembers are listed in table Convex Geometry The conventional digital image classification tries to assign pixels to broad classes, the hyperspectral image matching attempts to make more detailed identifications, e.g. to the specific mineralogy of soils and rocks. The spectral matching requires techniques that make it possible to separate pure pixels from impure pixels. This problem is well matched to the capabilities of convex geometry, which examines multidimensional envisioned data in n- dimensions. Each point (pixel) within this data space can be examined as linear combinations of an unknown number of pure components (Campbell, J. B., 1996). In figure 5.1a, the three points A, B and C represent three spectral observations at the extreme limits of the cloud of data points, which represent the pixels. These three points are representing endmembers that are defined as the pure pixels. The pixels in the middle of the point cloud are a mixture of all of the endmembers. With other words: the endmembers are always placed at the corners of the n-dimensional, convex triangle around the point cloud. Pixels that are located outside the triangle put together by the three endmembers A, B and C, are not representing any of these three endmembers. Since these pixels also consist of some material that is of interest for a classification, the convex triangle must be set larger. In figure 5.1(b), A, B and C are the observed approximations of the idealized spectra A', B' and C'. These are observations that can not be made on a specific image. 22

35 Endmember Selection Fig. 5.1: Spectral mixing analysis. (a) Simplex. (b) Endmembers (Tomkins, S., et al., 1993). When applying convex geometry to hyperspectral data, it is first necessary to define the data dimensionality Pixel Purity Index The Pixel Purity Index (PPI) is an algorithm used to find the most spectrally pure or extreme pixels in a multispectral or hyperspectral image data set. These spectrally pure pixels correspond with the endmembers needed for a SAM classification. The PPI is computed by repeatedly projecting n-dimensional scatter plots onto a random unit vector (the PPI process exploits convex geometry concepts which are explained in chapter 5.1.2). The extreme pixels in each projection are recorded and the total number of times each pixel is marked as extreme is noted. A threshold value is used to define how many pixels are marked as extreme at the end of the projected vector. Normally the threshold value should be about two or three times the noise level in the data. Higher thresholds cause the PPI to find more extreme pixels, but they are less likely to be pure endmembers (ENVI Help). In figure 5.1 of the previous chapter, this can be explained easier. According to the figure, a point is defined as pure when it is located at a corner in the convex triangle which is located around the point cloud. This means that the correlation between a point located in a corner and the other points is minimal, and exactly that is the purpose of the PPI. The PPI can give useful results in underdetermined data, but it can also miss certain endmembers since their corners of the data cloud are lost in projection of the data to the underdetermined spectral space (ENVI Help). 23

36 Endmember Selection 5.2 Applied Endmember Selection Ground Truth As already mentioned above, the endmembers applied were defined with the help of ground truth data which were collected in the field in parallel to the CHRIS/PROBA data set takes. The data were collected in the area of St.Urban in the RSL-Test site Vordemwald on the 17 th of August The ground truth contains 30 different types of land use classes, among which not all were qualified for identification in the CHRIS/PROBA data set. The reason why some classes could not be used was first of all that some of the samples were placed outside the image frame of the dataset. Another reason was the land use field sizes. The CHRIS/PROBA data set has a spatial resolution of 18 meters; this means that some of the sampled fields were simply too small to be taken into consideration because they would contain too many mixed pixels. The first elimination of ground truth classes was done without using data supported methods. The fields were selected by considering, only by optical means, whether they were located inside the image frame and whether they were large enough. The ground truth fields that were suitable to be used with the CHRIS/PROBA data set are listed in table 5.1 below. Land Use Class Ascertainment Abbreviation Maize M Stubble Field (harvested, stramineous) S2 Stubble Field (harvested, bare) S3 Stubble Field (seeded, partly green) S4 Field (harvested) A1 Field (seeded) A2 Field (ploughed) A3 Pasture Land We Grassland with trefoil Wi Alfalfa L Grassland without trefoil GS Sugar Beet Z Tab. 5.1: Modified ground truth with land use classes qualified for CHRIS/PROBA data set ENVI The new, modified ground truth data were used to define the endmembers for the classification in ENVI Spectral Hourglass Wizard (SHW) The SHW workflow was already described in chapter The workflow contains functions for deriving endmembers directly from the input data, but it is also possible to enter the endmembers from another source such as a spectral library, an ASCII file, a statistics file, or 24

37 Endmember Selection Regions of Interest (ROI). It is crucial that these files have the same data units, and the same spectra scaling, as the image data. The files are imported into the SHW through the Import menu bar item in the Endmember Collection dialog. For the classification of the CHRIS/PROBA image file the endmembers were imported from a ROI file (see chapter ). The next step in the SHW work flow is the Spectral Analyst. The Analyst is an aid to help identify the endmembers, but since they were already identified (ground truth), this step was not needed Regions of Interest (ROI) The Regions of Interest-Tool can be found in the ENVI Overlay toolbar under Regions of Interest. The ROI s were set with the help of the modified ground truth (table 5.1). As already mentioned above, the ground truth land use fields were selected only by considering whether their size was large enough to be identified, and sampled in the image file. By setting these samples with the ROI-tool, one can also study the spectrum of each land use class using the Z-Profile application in the Tools menu bar item, and in this way find out if the different spectra are separable. After setting all the image samples, further considerations of classes needed to be made. This was necessary because many classes were not separable. The three stubble field classes (S2, S3 and S3) and one field class (A1) had to be merged into one big class (see table 5.2). Since all of these classes have very sparse vegetation due to harvesting and ploughing, their spectra were not separable from each other. The new large class was named: Bare Fields. For three other classes: the Alfalfa (L), and the two field classes A2 and A3, the samples set in the image file turned out to be of too low quality to be used as endmembers. The samples contained too many mixed pixels and they therefore had to be excluded. The quality of the samples set for all of the other classes (table 5.1) was good and they could be used as endmembers for the classification. The CHRIS/PROBA image file also contains other thematic classes which can be identified and sampled by optical means. To avoid an image classification in which large parts would show no data, these items were sampled as well. The new classes were: Clouds, Water, Urban area, Forest, and Shadow. Three of these new classes, Forest, Clouds and Water, were defined with the help of 2- dimentional scatter plots and saved as Regions of Interest. The ROI classes set with scatter plots are saved as Point layers while the other classes are saved as polygon layers. The point layers turned out to be a problem later on. This problem will be further described in chapter The new endmembers are listed in the table below (table 5.2) along with a spectral plot (figure 5.2). Applied Endmembers Maize Bare Fields Pasture Land Grassland with trefoil Grassland without trefoil Sugar Beet 25

38 Endmember Selection Applied Endmembers Clouds Water Urban area Forest Shadow Tab. 5.2: Applied endmembers for SAM classification of CHRIS/PROBA dataset from August 17 th Fig. 5.2: Spectral profiles of applied endmembers for SAM classification of CHRIS/PROBA dataset acquired August 17 th Pixel Purity Index (PPI) As already mentioned above, the PPI is an algorithm used to find the most spectrally pure, or extreme pixels in a multispectral or hyperspectral image dataset. In ENVI this algorithm is integrated in the Spectral Hourglass Wizard (see chapter 3.2.1). The PPI was executed with the help of the SHW on the CHRIS/PROBA data set. This was done to try to identify endmembers other than the ones already identified in the ground truth. As a result of the PPI, a plot and an image are configured. To calculate the PPI the user needs to specify the number of iterations. The SHW recommends 5000 iterations for a useful result. The Pixel Purity Index plot shows the number of iterations and the cumulative number of pixels that have been found to be extreme. The curve in this plot usually starts steeply, as new pixels are found in each iteration, and it flattens out when all the extreme pixels are found (figure 5.3). 26

39 Endmember Selection Fig. 5.2: Pixel Purity Index Plot of CHRIS/PROBA dataset. The flattening of the PPI plot curve for the CHRIS/PROBA dataset was already the case after 2600 iterations (figure 5.3). When the iterations are completed, a PPI image is created in which the value of each pixel corresponds to the number of times that pixel was recorded as extreme. The bright pixels in the PPI image are normally the endmembers (see p.70 in the appendix). This was also the case for the CHRIS/PROBA PPI image. The bright pixels corresponded to the endmembers already identified in the ground truth (table 5.2), but did not identify new potential endmembers. The PPI process was therefore not necessary for the classification of the CHRIS/PROBA data set, and was left out of the classification work flow Geomatica For Geomatica a definition of endmembers was no longer needed, since this was already done in ENVI (see chapter ). The endmembers defined in ENVI were imported into Geomatica to make sure that the classifications performed in both software packages were based on exactly the same information. The endmembers, or Regions of Interest, were saved as shape files in ENVI and saved as bitmaps in Geomatica Bitmaps A bitmap is a segment (in Geomatica Focus) which is attached to an image file along with other segments such as vectors and rasters. The bitmaps are storage spaces for masks. In this case these masks represent the samples set in ENVI with the ROI-tool. The Regions of Interest were saved as shape files with the help of the ROI-tool in ENVI, and these shape files must be imported into bitmaps in Geomatica ROI-Import The ENVI shape files were imported into bitmaps with an algorithm called POLY2BIT which can be found in the Algorithm Librarian in Geomatica Focus. The POLY2BIT algorithm converts the polygons contained in the shape files into bitmaps. A geo-coded image segment can be chosen so that the image properties are added to the imported masks. This step is very 27

40 Endmember Selection important, because it is crucial that the imported masks and the image file have the same goereferencing to be able to obtain correct results. The problem with importing ROI s derived from scatter plots (see chapter ) occurred during this work step. The three classes Water, Forest and Clouds were set as point layers in ENVI, using 2-dimensional scatter plots, while the other classes were all set as polygons. It turned out that Geomatica can not read these point layers, so the shape files of the three classes could not be imported into bitmaps. The classes saved as polygons could be imported without any problems. As an attempt to solve the problem with the point layers, they were converted into raster layers in the ESRI software Arc Map. These raster files were imported into bitmaps in Geomatica with an algorithm named RAS2BIT which converts rasters into bitmaps. The problem seemed to be solved at fist sight, but this turned out not to be the case later on. The program Arc Map uses a different calculation method to compute rasters and polygons than ENVI and Geomatica does. The converted raster layers were slightly dislocated, so that the samples aet in ENVI were no longer identical to the samples stored in the bitmaps in Geomatica. This mistake showed its effect when the classification results from both programs were compared. The Geomatica classification image result was very different from the one in ENVI. To correct this mistake, the samples for the three classes, Forest, Clouds and Water, had to be reset new. The classes were reset as polygons in ENVI and were then imported into bitmaps in Geomatica with the algorithm POLY2BIT. It that way the endmembers used in the two programs are identical Pixel Purity Index (PPI) In Geomatica the program ENDMEMB can automatically calculate user-defined input regions in the image with the help of an algorithm called iterative error analysis (IEA). The ENDMEMB estimates the number of non-redundant endmembers by a) numbering the eigenvalues for the image-region spectra in order of decreasing value, and b) determination after which eigenvalue the adjacent differences become close to uniform. The IEA algorithm computes a new endmember vector in each process-iteration, starting with one endmember in the first iteration. The algorithm executes N iterations, where N is the number of endmembers. The endmembers computed by ENDMEMB may correspond to materials in the user-defined image region, but there is no guarantee that this will be the case. Since the endmembers were already defined in ENVI, this step was not necessary. 5.3 Endmember Valuation The endmembers were defined using existing ground truth data of the image area. The classes defined in this ground truth were set as Regions of Interest in the image file using the ENVI ROI-tool. As already described above, some of these classes had to be eliminated at once, for reasons described in chapter Other classes were eliminated later because they could not be separated from each other. The endmembers of sufficient quality for the Spectral Angle Mapper classification are listed in table 5.2, but will be listed once more in table 5.3, along with a valuation flag to define the quality of each single endmember. The valuation is only based on the observation of the 28

41 Endmember Selection classification image results and the endmember spectra. The endmembers were defined as very good, good or acceptable. Applied Endmembers Endmember Valuation Maize Very good Bare Fields Good Pasture Land Good Grassland with trefoil Acceptable Grassland without trefoil Good Sugar Beet Acceptable Clouds Very good Water Very good Urban area Good Forrest Very good Shadow Very good Tab. 5.3: Applied endmembers with valuation flags. The valuation of the endmembers is depending on various factors. The most important one is the visual quality of the endmember spectra. Another important factor is the recognisability of each of the land use classes in the image, or with other words; how large are the land use fields in the image file. This is important, because large image objects are more likely to contain homogenous pixels than small image objects, especially if the spatial resolution of the image is low. Endmembers of very good quality are listed in figure 5.3 below. The classes: Forest, Shadow, Maize, Clouds and Water, are all among the main image objects in the CHRIS/PROBA data set. The samples set in ENVI with the ROI-tool were therefore easier to define for these classes than for the others. Not only because the size of the classes made them easier to be found, but also because the large classes contain more pixels, among which one is bound to find more pure pixels. The pure pixels were mainly located in the center of the class object; on the edges most pixels were mixed. Fig. 5.3: Spectral Profiles of Endmembers valuated as very good. The class of endmembers with good quality only contains two endmembers. The Bare Fields and the Urban Area are separable as can be seen from their spectral profiles in figure 5.4 below. However, because of the relatively low spatial resolution of the image dataset, the 29

42 Endmember Selection classes partly overlap in the classification image (see chapter 7.1 and 7.2). The reason for this overlapping is that the sample sets for both classes probably contain mixed pixels. The size of the land use fields (endmembers) in this class is much smaller than the classes mentioned above, the samples were therefore harder to identify. Fig. 5.4: Spectral Profiles of Endmembers valuated as good The third class contains the endmembers with an acceptable quality. These endmembers, Grassland without trefoil (-), Grassland with trefoil (+), Sugar Beet, and Pasture Land, are all classes which contain green vegetation, so their spectra are therefore very similar and hard to differentiate. Most of the classes even partly contain some of the same plants, e.g. grass. The differences are minimal, especially between the Grassland with trefoil and the Sugar Beet. The land use fields (endmembers) in the image file which belong to this class (valuation acceptable) were very small, so the samples were difficult to define; this fact combined with the low spatial resolution makes some overlapping of the classes in the classification image inevitable. The spectral profiles of the endmembers are listed in figure 5.5 below. Fig. 5.5: Spectral Profiles of Endmembers valuated as acceptable. 30

43 Spectral Angle Mapping (SAM) 6 Spectral Angle Mapper (SAM) Classification The Spectral Angle Mapper (SAM) algorithm was developed by J.W. Boardman at the University of Colorado in Boulder (Kruse, F. A., et al., 1993). The SAM programs in both ENVI and Geomatica are based on this algorithm. This chapter concentrates on the implementation of the SAM classification programs and on the necessary classification work steps in the two software packages. 6.1 ENVI ROI-Classes The Regions of Interest were set as samples in the CHRIS/PROBA image file, using the ENVI ROI-tool (see chapter 5.2.2). The ROIs and their valuations are listed in table 5.3 above. The regions serve as endmembers for the Spectral Angle Mapper classification Spectral Hourglass Wizard (SHW) The SAM classification function is integrated in the Spectral Hourglass Wizard along with two other classification methods; the Linear Spectral Unmixing and the Mixture-Tuned Matched Filtering SAM Classification The SAM classification is the last step in the SHW (see chapter 5.2.2). SAM computes a classification image based on a SAM Maximum Angle Threshold. If this threshold is set low it usually result in fewer matching pixels. In this case the threshold was set to 0.5 radian. The SAM output consists of a classification image and a set of rule images which correspond to the spectral angle calculated between each pixel and each endmember. One rule image is delivered for each endmember. The rule images are helpful if the SAM classification image does not show spatially coherent classes. They can then be examined individually by loading each rule image into a display window. The SHW automatically reverses the colour table and estimates a linear stretch of the image, applying the gray scale display range to the valid detections in the left part of a SAM histogram. The best matches are the small angles, so the colour table is reversed to give a more conventional detection image. To change the stretch maximum of the displayed rule image, the desired value can be entered in the Default Stretch Max text box. The ENVI Cursor Location-tool can be used to look at the SAM spectral angles for each pixel and identify a new SAM threshold if necessary (ENVI Help). The final panel of the ENVI SHW work flow is the Spectral Mapping Wizard Summary Report. This report contains a summary of each processing step made using the SHW and it also contains a list of the output files created. 31

44 Spectral Angle Mapping (SAM) 6.2 Geomatica ROI-Classes from ENVI The ROI-Classes configured in ENVI were imported into Geomatica to make sure the classifications performed in both programs were based on exactly the same sampled regions. Each ROI-Class was converted into a shapefile with an integrated function (export ROI s to shapefile) in the ENVI ROI-tool. These shapefiles were then converted into bitmaps by the program POLY2BIT in Geomatica. POLY2BIT converts polygons into bitmaps, which are storage spaces for masks Spectra Extraction To be able to use the endmembers (bitmaps) for a SAM classification, one must first extract the spectra of each endmember into a spectra file. This can be done by configuring the endmember data with the Spectra Extraciton Configuration dialog box, which is found in Geomatica Focus under Analysis. Here the input image file (vordemwald.pix) and the channels to work with must be specified. One can select an existing channel or a new one can be created. In this case a new channel was created. For this channel a Spectra Extraction dialog box appears, in which all the spectra from the bitmaps can be imported. Starting a Spectra Extraction Configuration with no open data in the Focus view, the dialog box will automatically create a new Map area for this work. If a map and an area are already open and the input file has the same georeferencing as the area, a new hyperspectral metalayer is added to the map and area. When the georeferencing is incompatible, which was not the case, a new area is added to the existing map Spectra Plot When the hyperspectral metalayer is configured, a Spectra Plot dialog box can be accessed, where the spectra from the Spectra Extraction dialog box mentioned above can be imported. These spectra can then be saved as a spectra file (.spl) which is needed for the SAM classification. The endmember spectra are listed in figure 6.1 below. 32

45 Spectral Angle Mapping (SAM) Fig. 6.1: Endmember spectra from SPL-file Regions of Interest (ROI) During the Spectra Extraction configuration, a hyperspectral metalayer is created in the Files tree in Focus. As already mentioned above, a sub menu can be accessed by right-clicking this metalayer. In this sub menu commands like Spectra Extraction and the Spectra Plotting are found. After running these two commands a sub layer called Regions of Interest, which is attached to the hyperspectral metalayer, appears. The Regions of Interest layer (ROI) contains the masks which were imported from the bitmaps. One can link the analysis between the image and the ROI layer and work with scatter plots or spectra plots SAM Classification The Spectral Angle Mapper program can be found in Geomatica Focus in the Algorithm Librarian, under Geomatica Hyperspectral. The SAM is listed in the Hyperspectral Analysis Package along with the program Spectral Unmixing. SAM is designed to classify image data on the basis of a set of reference spectra that define the classes (the ROIs saved in the spectra file). The SAM program computes the spectral angle between each band-vector in a specified region of the input image, in this case the whole image, and each of the spectra read from the spectra file. This angle is the measurement of spectral similarity between the band vector and the reference spectrum; the smaller the angle the greater the similarity. All 37 bands of the CHRIS/PROBA dataset were used for the classification. The spectra are treated as band vectors for the purpose of the angle computation. A classification channel indicates the input reference spectrum, with which it has the smallest angle, for each image pixel. The pixels are assigned to the NULL (0) class if the minimum spectral angle is greater than the threshold value. This threshold is optional and specifies the spectral angle in degrees (between 0 and 180). If this field is left open all pixels are assigned to the class corresponding to the smallest spectral angle, no matter how large that angle is. This method was not appropriate for the CHRIS/PROBA dataset, because a lot of pixels were assigned to classes they did not belong to. By using the try and error method, the threshold was set to 30 degrees. 33

46 Spectral Angle Mapping (SAM) The algorithm assumes that the number of measurement values in the reference spectrum equals the number of bands in the image, and that the band centre wavelength values match the wavelength coordinates of the reference spectrum measurement values. The angle between each reference spectrum and each band vector is computed, and these angles are assigned to output channels that automatically appear in the Focus view. Then, each pixel is assigned to the class to which it has the smallest angle (defined by the reference spectrum). The class assigned to each pixel is saved in the output classification channel which appears automatically in the Focus view. 34

47 Results 7 Results In this chapter the classification results and the accuracy assessments from both the ENVI and the Geomatica software packages are displayed and discussed. 7.1 ENVI The classification result image in figure 7.1 shows some overlapping (mixed pixels) between the different land use classes. These classes are mostly the ones with spectra similar to each other (see chapter 5.2), as e.g. between the Grassland with trefoil and Maize, and between Bare Fields, Urban areas and Grassland without trefoil. The overlapping is particularly obvious on the borders of the land use fields. Another reason for these overlapping effects is the choice of the spectral angle value for the SAM classification. This angle was set to 0.5 radian in the SAM classification in ENVI (see chapter 6.1.3). The best pixel matches in the SAM classification are always the smaller angles. To be able to classify all the image pixels, a higher angle was necessary. Higher angle values may result in a more spatially coherent image, however, the overall pixel matches will not be as good as for the smaller angle. Hereby some of the classes (see below) recognized more pixels during the classification and in this way contain pixel spectra with higher similarities to those of other classes. These facts led to overlapping effects between well separable classes (e.g. Clouds and Grassland without trefoil) as well as the less separable classes mentioned above. 35

48 Results Fig. 7.1: ENVI SAM classification result. 36

49 Results 7.2 Geomatica Comparing the classification results performed in ENVI (Fig.: 7.1) and Geomatica (Fig.: 7.2), there are no differences visible at first sight. The overlapping effects in the Geomatica classification image are the same as the ones explained in chapter 7.1 above. The reasons for these effects are the similarity of class spectra (see below) and the size of the spectral angle chosen for the SAM classification. In Geomatica this angle was set to 30 degrees (see chapter 6.2.3). 37

50 Results Fig. 7.2: Geomatica SAM classification result. 38

51 Results 7.3 Accuracy Assessment To be able to statistically differentiate between the two classification results, accuracy assessments have to be preformed. Such an assessment task can be found in Geomatica Focus under the Post Classification Analysis dialog box. The accuracy assessments determine the correctness of classified images, which are based on pixel groupings. The measure of accuracy is the correlation between a standard that is assumed to be correct and an image classification of unknown quality. In this thesis, the ground truth data consist of 39 verification samples which were set as vector layers in the image file. These samples were located as described in chapter and chapter The verification samples were used as a standard for the accuracy assessments of the classifications made in both ENVI and Geomatica. The training classes which were added to the classification by optical means (Clouds, Forest, Shadow, Water, and Urban Area) were excluded from the classified images prior to the accuracy assessments. This was done using the Geomatica Class Labelling, and Class Editing dialog boxes. These classes were deleted because they all contained samples which were set over such large areas, that it was not possible to locate verification samples without overlapping the classification samples. The classification samples were only set by optical means, not with the help of ground truth data like the other samples, and could therefore not be assumed to be correct with absolute certainty. The modified classification images, now containing only the classes Maize, Bare Fields, Pasture Land, Grassland with trefoil, Grassland without trefoil, and Sugar Beet, were saved as channels in the image file. In the Accuracy Assessment dialog box these channels were selected along with the reference image, here the image file, and the verification samples, which were stored as vector segments in the image file. For each verification sample a reference attribute from the classification image was selected. The Geomatica Accuracy Assessment produces an Accuracy Report which contains a Random Sample Listing, a Confusion Matrix and Accuracy Statistics. The Random Sample Listing shows an information table for all the randomly generated test pixels or in this case the verification samples (see table 7.1 and 7.2). This report reveals first differences between the classification image and the reference data without showing the statistics. It is only possible to see if the classified values are effectively the same as the verification values, which was not always the case for both the ENVI and the Geomatica classification images. Sample Number Georeferenced Position Easting Northing Database Position Pixel Line Classified Value Name Reference Value Name Bare Fields 7 Bare Fields Grassland + 5 Grassland Sugar Beet 2 Sugar Beet Pasture Land 6 Pasture Land Grassland - 3 Grassland - 39

52 Results Sample Number Georeferenced Position Easting Northing Database Position Pixel Line Classified Value Name Reference Value Name Sugar beet 3 Grassland Bare Fields 7 Bare Fields Bare Fields 7 Bare Fields Pasture Land 7 Bare Fields Bare Fields 7 Bare Fields Bare Fields 7 Bare Fields Bare Fields 7 Bare Fields Pasture Land 6 Pasture Land Pasture Land 6 Pasture Land Pasture Land 6 Pasture Land Pasture Land 7 Bare Fields Bare Fields 7 Bare Fields Bare Fields 7 Bare Fields Pasture Land 7 Bare Fields Pasture Land 5 Grassland Pasture Land 5 Grassland Bare Fields 7 Bare Fields Grassland + 8 Maize Grassland + 8 Maize Maize 8 Maize Grassland + 5 Grassland Grassland + 5 Grassland Bare Fields 7 Bare Fields Bare Fields 7 Bare Fields Pasture Land 6 Pasture Land Bare Fields 7 Bare Fields Pasture Land 5 Grassland Grassland + 5 Grassland Pasture Land 6 Pasture Land Grassland - 3 Grassland Grassland + 3 Grassland Grassland + 3 Grassland Maize 8 Maize Maize 8 Maize Tab. 7.1: ENVI Random Sample Listing. Sample Number Georeferenced Position Easting Northing Database Position Pixel Line Classified Value Name Reference Value Name Bare Fields 7 Bare Fields Grassland + 5 Grassland Sugar Beet 2 Sugar Beet Pasture Land 6 Pasture Land Grassland - 3 Grassland Sugar beet 3 Grassland Bare Fields 7 Bare Fields Bare Fields 7 Bare Fields Pasture Land 7 Bare Fields Bare Fields 7 Bare Fields 40

53 Results Sample Number Georeferenced Position Easting Northing Database Position Pixel Line Classified Value Name Reference Value Name Bare Fields 7 Bare Fields Bare Fields 7 Bare Fields Pasture Land 6 Pasture Land Pasture Land 6 Pasture Land Pasture Land 6 Pasture Land Pasture Land 7 Bare Fields Bare Fields 7 Bare Fields Bare Fields 7 Bare Fields Pasture Land 7 Bare Fields Pasture Land 5 Grassland Pasture Land 5 Grassland Bare Fields 7 Bare Fields Grassland + 8 Maize Grassland + 8 Maize Maize 8 Maize Grassland + 5 Grassland Grassland + 5 Grassland Bare Fields 7 Bare Fields Bare Fields 7 Bare Fields Pasture Land 6 Pasture Land Bare Fields 7 Bare Fields Pasture Land 5 Grassland Grassland + 5 Grassland Pasture Land 6 Pasture Land Grassland - 3 Grassland Grassland + 3 Grassland Grassland + 3 Grassland Maize 8 Maize Maize 8 Maize Tab. 7.2: Geomatica Random Sample Listing. There were several differences between the reference data and the classification images, especially between the land use classes Bare Fields and Pasture Land, Grassland with trefoil and respectively Pasture Land, Maize and Grassland without trefoil. The reasons for these overlapping effects are due to the spectrally indifferent behaviour of the land use classes which were explained above in chapter 5. These differences between the reference data and the classification images were identical in the classifications made in ENVI and Geomatica (see tables 7.1 and 7.2). The Error (Confusion) Matrix displays the results of the Accuracy Assessment process. The reference data is listed in the columns of the matrix and represent the number of correctly classified samples. The ENVI classification Error Matrix and the matrix from the Geomatica classification are identical, as can be observed in Tab.:7.3 and Tab.:7.4 below. In an Error Matrix, two different types of errors can occur: Commission Errors and Omission Errors. The Commission Errors are all the pixels which are assigned to wrong classes, and the Omission Errors are all of the picture elements which are incorrectly excluded by their own classes. 41

54 Results Classified Data Reference Data Sugar Grassland - Grassland Pasture Bare Maize Totals Beet + Land Fields Sugar Beet Grassland Grassland Pasture Land Bare Fields Maize Unknown Totals Tab. 7.3: ENVI Error (Confusion) Matrix. Classified Data Reference Data Sugar Grassland - Grassland Pasture Bare Maize Totals Beet + Land Fields Sugar Beet Grassland Grassland Pasture Land Bare Fields Maize Unknown Totals Tab. 7.4: Geomatica Error (Confusion) Matrix. One sample in the tables above is listed as Maize(red) in the reference data and unknown in the classified data. To find out which unknown class this sample belongs to in the classified image, one can add the samples made for the classes which were added to the classification by optical means (Clouds, Forest, Shadow, Water, and Urban Area) and run another Accuracy Report. The sample belonged to the class: Forest in the classified data. The last part of the Accuracy Report is the Accuracy Statistics Report. This report lists different statistical measures of accuracy (as discussed in chapter 2.4) for each class (see tables 7.5 and 7.6). Class Producer s User s Cappa Accuracy Accuracy Statistic Maize 60.00% %

55 Results Class Producer s User s Cappa Accuracy Accuracy Statistic Bare Fields 80.00% % 1.00 Pasture % 50.00% 0.41 Land Grassland 57.14% 57.14% 0.48 with trefoil Grassland 40.00% % 1.00 without trefoil Sugar Beet % 50.00% 0.49 Tab. 7.5: ENVI Accuracy Statistics Report, Overall Accuracy: 71.80%. Class Producer s Accuracy User s Accuracy Cappa Statistic Maize 60.00% % 1.00 Bare Fields 80.00% % 1.00 Pasture % 50.00% 0.41 Land Grassland 57.14% 57.14% 0.48 with trefoil Grassland 40.00% % 1.00 without trefoil Sugar Beet % 50.00% 0.49 Tab. 7.6: Geomatica Accuracy Statistics Report, Overall Accuracy: 71.80%. The statistical measures for both classifications show identical values (see tables 7.5 and 7.6). These values are calculations based on the information found in the error matrixes (tables 7.3 and 7.4). The statistical measure Overall Accuracy is the percentage of correctly classified pixels in the classification images (see chapter 2.4.1). The classifications made in ENVI and Geomatica both have an Overall Accuracy of 71.80%. These results are satisfying, considering the difficulties of endmember selection due to the low spatial resolution and heterogeneity of the dataset mentioned in chapter The Producer s Accuracy shows what percentage of a particular reference class was correctly classified (see chapter 2.4.2). The accuracy for all of the reference classes (verification classes) in the two classifications is identical (see table 7.5 and 7.6). The accuracy values range between 40% and 100% in both classifications. These values show what percentage of the reference class was correctly classified. In the case of Sugar Beet and the Pasture Land, both classes have an accuracy of 100%. Bare Fields (80%) and Maize (60%) also have a relatively high Producer s Accuracy, while the Grassland with (57.14%) and without (40%) trefoil show low accuracy values. The accuracy values of the four last reference classes (Bare Fields, Maize, Grassland with trefoil 43

56 Results and Grassland without trefoil) are low because they contain pixels which belong to other classes (errors of commission). As already mentioned above, these pixels are not considered in the process of the Producer s Accuracy measure. The User s Accuracy is calculated by dividing the number of correct pixels for a class by the number of classified pixels for that class (see chapter 2.4.3). The Users s Accuracy is also identical for all of the classified classes in the ENVI and Geomatica classifications. In this case the accuracy values are ranging between 50% and 100%. The values show what percentage of a certain class was correctly classified. The Maize, Bare Fields and the Grassland without trefoil have an accuracy value of 100%. The Pasture Land (50%), Grassland with trefoil (57.14&) and the Sugar Beet (50%) all show slightly lower accuracy values. The pixels which are not identified by these classes are not taken into consideration in the process of the User s Accuracy measure. The Kappa Coefficient is a statistical measure of the agreement, beyond chance, between two maps (see chapter 2.4.4). The classes Maize, Bare Fields and Grassland without trefoil all show a kappa value of 1. The Pasture Land (0.41), Grassland with trefoil (0.48) and the Sugar Beet (0.49) all show lower kappa values, which indicate that these classes are not as accurate in the classification output as the other classes. When the Accuracy Assessment results are compared with the observations of spectra and classification images made in chapter 5.3, there are many similarities. The classes of low quality in the assessment are the same classes which were valuated as acceptable, or in some cases as good in the endmember valuation chapter (5.3). The classes with a very good valuation correspond to the assessment classes of good quality. 7.4 Conclusions The classifications performed in both software packages show identical results, even though the classification workflow was different on some points. Based on the obtained results one can assume that the implementation of the SAM algorithm in ENVI and Geomatica is equal (see chapter 7.3). Without considering how or if the classification work steps are integrated in programs of the two software packages, both ENVI and Geomatica are equally qualified to perform a SAM classification. However, taking the integration of work steps into consideration it must be said that ENVI lays a step ahead with the Spectral Hourglass Wizard approach (see figure 3.1). This is the case because the work steps for all hyperspectal classification methods are implemented into the software package, which also contains dialog boxes where every step of the classification is explained and tips are given. The Hyperspectral Analysis Package which is contained in Geomatica is also a very useful tool for a hyperspectral classification, but the fact that the work steps are only listed in the Geomatica help-function and not integrated into a workflow, like in ENVI, the proceeding is not as fluent as it is in ENVI (see table 3.2). 44

57 Concluding Observations 8 Concluding Observations The SAM classification in ENVI and Geomatica showed identical results despite differences in the classification workflow tools. As already mentioned in the chapter 7.4, the ENVI Spectral Hourglass Wizard workflow was implemented into the software package, while in Geomatica the Hyperspectal Analysis Package workflow was only listed in a help-function. Both classification tools contain all the work steps needed to perform a SAM classification, however, the ENVI Spectral Hourglass Wizard lays a step ahead because of the integrated workflow, which makes the classification procedure much easier and clearly arranged. The main differences in the work steps were the missing step Pixel Purity Index in the HAP in Geomatica, which was integrated in the ENVI SHW workflow. Since this step was not necessary for the classification of the CHRIS/PROBA data set, its absence from the HAP workflow list was not a problem. If the PPI process would have been necessary, the function can be found in the Algorithm Library in Geomatica Focus. Another difference was the necessity of importing metadata into the image file in Geomatica. This process is done automatically in ENVI, and is therefore not found in the SHW. The Spectra Extraction dialog box in Geomatica was another difference between the two software workflows. In this box the training area samples had to be imported and a file with the spectra, and Regions of Interest were configured. In ENVI the training area samples could be defined as ROI s from the beginning on, which made this step unnecessary. The last difference between the two workflows was the SAM angle which had to be chosen for the classification. In ENVI this angle was defined in radian, and in Geomatica as degree. The angles were chosen using the try an error method, where the angle with the best classification result was chosen for each software package. These angles slightly differed from each other in the two software packages, though the result image was the same. In ENVI the 0.5 radian ( 28.7 ) was the SAM angle that showed the best result. In Geomatica this angle was defined as 30 degrees. As already mentioned above the endmember (training areas) selection was done mainly in ENVI using the Regions of Interest tool. These endmembers were then imported into Geomatica. The possibility to do this work step in Geomatica was, however, also there. When the endmembers are defined, one has to decide if the spectra are of good enough quality for the classification. To find out if this is the case the user has to run a testclassification, observe the result image and decide if the spectra represent their mixed components well. To be able to do this, some pre-knowledge of the composition of mixed materials in the picture elements are useful. Because of the low spatial resolution of the image file, the image pixels were relatively heterogeneous. This fact made the endmember selection very difficult, because the image spectra were in this way hard to separate. Since most of the image classes consist of vegetation the spectra were all very similar. Image classes that were represented only by very small areas in the image file therefore had to be excluded from the classification. The classification results from both ENVI and Geomatica were compared visually and later also by using an Accuracy Assessment. Visually there were no differences to be seen between the two classification results. The images showed matching overlapping effects at the exact same locations and the areas without overlapping effects were also identical in both images. To be certain that these optical observations were correct, an Accuracy Assessment in Geomatica Focus was executed to statistically differentiate between the two classification results. These accuracy results confirmed the observations made by optical means, that the two classification results are identical. The Accuracy Assessment also confirmed the 45

58 Concluding Observations observations made on the quality of the class spectra which were made before running the classifications (see chapter 5.3). 46

59 Outlook 9 Outlook The Geomatica Hyperspectral Analysis Package is a very helpful tool for the classification of hyperspectral image data. However, it must be mentioned that this package is still in need of improvements and user-friendliness to be able to measure up with the ENVI Spectral Hourglass Wizard. Since the HAP is relatively new, these improvements are probably going to appear in the new Geomatica versions in the years to come. Until then, the SHW is the easier choice. To be able to obtain better classification results than the ones achieved in this thesis, one has different options. The best option would be to use data from a sensor with a higher spatial resolution and more band-information than the one obtained with CHRIS/PROBA (e.g. AVIRIS). A higher spatial resolution would make the endmember selection process a lot easier because the land use classes would contain less mixed pixels, which would deliver training areas with spectra of better quality and the classification results would show less overlapping effects. 47

60 Glossary Glossary The terms of definition contained in the tables below are listed in order of appearance. Geomatica Geomatica Focus OrthoEngine Modeler EASI Geomatica Hyperspectral Analysis Package (HAP) SAM (Spectral Angle Mapper) SPUNMIX (Spectral Unmixing) MNF-Transformation Principal Components (PC) MNFNR MNFLT Software package developed by the Canadian corporation PCI Geomatics. It includes a series of integrated work environments, e.g. Focus, OrthoEngine, Modeler and EASI Geomatica s main data visualization environment Photogrammetric tool designed to handle small and large production workloads to efficiently produce geospatial products Provides an interactive methodology for the development of both simple and complex data processing flows. It provides access to a number of standard operations such as data import and export, as well as most PACE/EASI processing algorithms Geomatica s full-featured programming language. The command-line based interface and scripting provide users with tools costomization and programming An aid designed for processing and analyzing images acquired with imaging spectrometers Computes the angle between the endmember spectra and each pixel vector in n- dimensional space Finds the abundances of materials in the image pixels by assuming that the pixel reflectance in each band is equal to a linear combination of the reflectances of the endmember material present within the pixel Maximum Noise Fraction-Transformation is an algorithm used to remove noise from image data. It uses two Principal Components (PC) transformations. This algorithm is called Minimum Noise Fraction- Transformation in ENVI. Consists of two programs for an MNF- Transformation (MNFNR, MNFLT) The band that contains the most noise is transformed so that its noise content is close to that of the other bands Computes the parameters of band-wise linear data transformations and their inverses, and to apply these 48

61 Glossary Geomatica Algorithm Librarian Hyperspectral Add-on Analysis Hyperspectral Analysis PACE Preprocessing Sensor-related Calibration Geometric Correction Noise Removal ENDMEMB Iterative Error Analysis (IEA) Atmospheric Correction Local Analysis Spectral Handling Metadata I/O Spectra Extraction Configuration Dialog box Spectra Plotting Contains listings of all the existing algorithms in Geomatica Library listing the algorithms designed to process and analyze images acquired with imaging spectrometers Application grouping in Geomatica Focus Task box for processing and analyzing image file with the Hyperspectral Add-ons Application program that consist of different functions like Preprocessing, Atmospheric Correction, Local Analysis, Spectral Handling and Metadata I/O Application program with tasks to preprocess the image file (sensor-related calibration, geometric correction and noise removal) Subtracts dark-reference values from band image values in order to obtain band image values that are more closely proportional to at-sensor radiance Removes geometric distortions made by e.g. aircraft roll from the image date (if accurate sensor position and attitude data are acquired together with the image data) Consists of different programs for image noise removal (MNFLT, MNFNR etc.) Program that helps to locate endmembers based on the iterative error analysis (IEA) algorithm Algorithm that computes a new endmember vector in each process-iteration Reduces or removes atmospheric effects from the image file Contains programs that can be used for local analysis of imaging spectrometer data (ENDMEMB, SAM, SPUNMIX) Contains programs that provide a variety of spectra handling capabilities. Program with two different functions: METAIN imports metadata into an image file (PIX), METAOUT reads the information in the metadata segment in the image file and exports it as an XML document Extracts spectra either directly from image file or imports the spectra from external files Function which displays and saves spectra from the Spectra Extraction Configuration Dialog into a SPL file 49

62 Glossary Geomatica Hyperspectral Metalayer Geomatica Hyperspectral Pixel Purity Index (PPI) POLY2BIT RAS2BIT Post Classification Analysis Accuracy Assessment Random Sample Listing Confusion Matrix Accuracy Statistics Class Labelling Class Editing ATCOR-3 MODTRAN-4 Overall Accuracy Producer s Accuracy User s Accuracy Kappa Coefficient During spectra extraction, a metalayer is created in the Files-tree in Geomatica Focus Library of algorithms for hyperspectral data Algorithm used to find the most spectrally pure or extreme pixels in a multispectral or hyperspectral image data set Program that converts polygons contained in the shape files into bitmaps Program that converts rasters into bitmaps Contains a group of algorithms used for analyzing classification results Program in the Post Classification Analysis used to determine the correctness of a classified image, which is based on pixel groupings Product of the accuracy assessment. Contains an information table for all the randomly generated test pixels Displays the results of the accuracy assessment process. The reference data is listed in the columns of the matrix and represent the number of correctly classified samples Report that lists different statistical measures of overall accuracy for each class Class editing and aggregation Combining of several classes Program for atmospheric corrections. Calculates a ground reflectance image using elevation data. Is based on MODTRAN-4 MODerate spectral resolution atmospheric TRANSsmittance algorithm and computer model, developed by AFRL/VSBT in collaboration with Spectral Sciences, Inc. The percentage of correctly classified pixels in a classification image Shows what percentage of a particular reference class was correctly classified Shows what percentage of a certain class was correctly classified Statistical measure of the agreement, beyond chance, between two maps ENVI ENVI Environment for Visualizing Images, software package developed by the American company ITT Visual Information Solutions 50

63 Glossary ENVI Spectral Hourglass Wizard (SHW) Spectral Analyst Spectral Angle Mapper (SAM) MNF-Transformation Mapping Method panel Mixture Tuned Matched Filtering (MTMF) Linear Spectral Unmixing Principal Components (PC) transformations Pixel Purity Index (PPI) Spectral Library Regions of Interest (ROI) Import Menu Application integrated in ENVI to guide the user step-by-step through the ENVI hourglass processing flow to find and map image spectral endmembers from hyperspectral data An aid to identify endmembers and to review mapping results Matches image spectra to reference spectra in n-dimensions. It computes the angle between the endmember spectra and each pixel vector in n-dimensional space Minimum Noise Fraction-Transformation is an algorithm used to remove noise from image data. It uses two Principal Components (PC) transformations. This algorithm is called Maximum Noise Fraction-Transformation in Geomatica. Panel with listing of classification methods Finds the abundances of endmembers using a partial unmixing approach. It uses a Matched Filter to maximize the response of the known endmember and surpresses the response of the composite unknown background Finds the abundances of materials in the image pixels by assuming that the pixel reflectance in each band is equal to a linear combination of the reflectances of the endmember material present within the pixel Consists of two programs. One decorrelates and rescales the noise in the Data, the other transforms data that appear affected (striped or white) by the noise Algorithm used to find the most spectrally pure or extreme pixels in a multispectral or hyperspectral image data set Contains spectra that are acquired in a laboratory, they represent hundreds of different minerals and vegetation types Tool used to set samples in an image file Menu where ROI s are imported from spectral libraries, individual spectral plots, text files, statistic files or from an already existing ROI file 51

64 Glossary ENVI Endmember Collection Overlay toolbar Tools Z-Profile ENVI Cursor Location-tool SAM Maximum Angle Threshold Spectral Mapping Wizard Summary Report Dialog Window in the SHW. Allows standardized selection of endmember spectra. These can come from spectral libraries, individual spectral plots, text files, ROI s, or statistics files Contains different functions to handle the image file (e.g. ROI, classification, annotation, contour lines) Contains different functions to handle and edit the image file (e.g. Profiles, Build Mask, Scatter plots) A tool used study the spectra of ROI s A tool used to look at the SAM spectral angles for each pixel and identify a new SAM threshold if necessary SAM classifications are based on this threshold. Small vales usually result in fewer matching pixels. Higher values may result in a more spatially coherent image, however, the overall pixel matches will not be as good as for the lower threshold The final panel of the ENVI Hyperspectral Process Flow Wizard. Contains a summary report of the processing steps preformed using the Wizard and lists of the output files Others Arc Map MATLAB Geographical Information System program developed by ESRI Commercial software developed by MathWorcs Inc. 52

65 Bibliography Bibliography Albertz, J., 2001: Einführung in die Fernerkundung, Grundlagen der Interpretation von Luft- und Satellitenbildern, 2. überarbeitete und erweiterte Auflage, Wissenschaftliche Buchgesellschaft, Darmstadt. Barnsley, M. J., Settle, J. J., Cutter, M. A., Lobb, D. R., Teston, F., 2004: The PROBA/CHRIS Mission: A Low-Cost Smallsat for Hyperspectral Multiangle Observations of the Earth Surface and Atmosphere, IEEE Transactions on Geoscience and Remote Sensing, Vol. 42, No. 7. Boardman, J. W., 1989: Inversion of Imaging Spectrometry Data Using Singular Value Decomposition, CSES/CIRES, Geoscience and Remote Sensing Symposium, 12th Canadian Symposium on Remote Sensing, 1989 International Volume 4, pp Berk, A., Bernstein, L. S., Robertson, D. C., 1989: MODTRAN: A moderate resolution model for LOWTRAN7, Report GL-TR , Air Force Geophysics Laboratory, Bedford, MA. Boardman, J. W., 1990: Inversion of High Spectral Resolution Data, Proceedings of SPIE Vol. 1298, Imaging Spectroscopy of the Terrestrial Environment, Greg Vane, Editor, pp , Boardman, J. W.: Automating Spectral Unmixing of AVIRIS Data Using Convex Geometry Concepts, Summaries of the Fourth Annual JPL Airborne Geoscience Workshop, Robert O. Green, JPL Publication, Vol. 1, pp , Pasadena. Campbell, J. B., 1996: Introduction to Remote Sensing, Second Edition, Taylor & Francis, London (GB). Congalton, R. G., Green, K., 1998: Assessing the Accuracy of Remotely Sensed Data: Principles and Practices (Mapping Science Series), Lewis Publishers, USA. 53

66 Bibliography ENVI Help, 2006: ENVI Help, ENVI 4.3 Help, ITT Industries Inc., Boulder, CO, USA. ENVI Software, 2006: ENVI Version 4.3, ITT Industries Inc., Boulder, CO, USA. Geomatica 10 Brochure, 2006: Geomatica 10 Brochure, Leading the Image-Centric Revolution, PCI Geomatics Enterprises Inc., Ontario, Canada. Geomatica Focus User Guide, 2005: Geomatica Focus User Guide version 10, PCI Geomatics Enterprises Inc., Ontario, Canada. Geomatica Help, 2005: Geomatica Version 10 Help, PCI Geomatics Enterprises Inc., Ontario, Canada. Geomatica Software, 2005: Geomatica Version 10.0, Software Solutions, PCI Geomatics Enterprises Inc., Ontario, Canada. Goetz, A. F. H., 1995: Imaging Spectroscopy for Remote Sensing: Vision to Reality in 15 Years, Imaging Spectroscopy, Michael R. Descour, Jonathan M. Mooney, David L. Perry, Luanna Illing, Proc. SPIE 2480, pp. 2-13, Bellingham. Goetz, A. F. H., Vane, G., Solomon, J. E., Rock, B. N., 1985: Imaging Spectrometry for Earth Remote Sensing, Science, New Series, Vol. 228, No. 4704, pp Green, R. O., Eastwood, M. L., Sarture, C. M., Chrien, T. G., Aronsson, M., et al., 1998: Imaging Spectroscopy and the Airborne Visible/Infrared Imaging Spectrometer (AVIRIS), Remote Sensing of Environment, Vol. 65, pp Kneubühler, M., Koetz, B., Huber, S., Schopfer, J., Itten, K. I., 2006: Monitoring Vegetation Growth using Multitemporal CHRIS/PROBA Data, IEEE Geoscience and Remote Sensing Symposium (IGARSS 2006), pp , Denver, USA. 54

67 Bibliography Kneubühler, M., Koetz, B., Richter, R., Schaepfman, M., Itten, K. I., 2005: Geometric and radiometric pre-processing of CHRIS/PROBA data over mountainous terrain, Proc. 3 rd CHRIS/PROBA Workshop, Frascati (I), March, ESA Publications Division, Noordwijk (NL), SP-593, CD-ROM. Kruse, F. A., Lefkoff, A. B., Boardman, J. W., et al., 1993: The Spectral Image Processing System (SIPS) Interactive Visualization and Analysis of Imaging Spectrometer Data, Remote Sensing of Environment, Vol. 44, pp Landgrebe, D., 1997: On Information Extraction Principles for Hyperspectral Data, A White Paper, School of Electrical & Computer Engineering, Purdue University, West Lafayette. Liu, G., Allen, J., Lu, K., Kleppel, G., Parkey, J., 2006: Hyperspectral Signal Processing Applications in Environmental Monitoring Identification and Mapping of the Invasive Plant Species Purple Loosestrife, IEEE Conference, 8 th International Conference on Signal Processing, ICSP, Nov , Guilin, China. PCI Geomatics Hyperspectral Data Primer, : Hyperspectral Image Analysis, Geomatica 10, PCI Geomatics Enterprises Inc., Ontario, Canada. Richter, R., 1998: Correction of satellite images over mountainous terrain, Applied Optics 37, pp Singer, M., 1998: Anwendung des Spectral Unmixing auf einen Bildspektrometrie-Datensatz, Diplomarbeit am Geographischen Institut der Universität Zürich. Toutin, T., 2004: Rewiev article: Geometric processing of remote sensing images: models, algorithms and methods, Int. J. Remote Sensing 25 (10), pp

68 Bibliography Ustin, S. L., 2004: Remote Sensing for Natural Resource Management and Environmental Monitoring, Manual of Remote Sensing, Third Edition, Volume 4, Andrew N. Rencz, Editor-inchief, American Society for Photogrammetry and Remote Sensing, John Wiley & Sons, Inc., Hoboken, New Jersey, USA. Vane, G., Green, R. O., et al., 1993: The Airborne Visible/Infrared Imaging Spectrometer (AVIRIS), Remote Sensing of Environment, Vol. 44, pp Vane, G., Goetz, A. F. H., 1988: Terrestrial Imaging Spectroscopy, Remote Sensing of Environment, Vol. 24, pp What s new in Geomatica 10, 2005: What s new in Geomatica 10, An Overview of the new Capabilities of Geomatica Version 10, PCI Geomatics Enterprises Inc., Ontario, Canada. Internet: CHRIS-PROBA Data Exploitation Web Site: The CHRIS-on-PROBA Mission Web Pages: ESA s Official CHRIS-PROBA Web Page: PCI Geomatics Homepage: ITT Industries Inc. Homepage: 56

69 Appendix 1. CHRIS/PROBA nadir data set acquired on August 17 th

70 2. Complete table of original ground truth data Class Maize Barley (green) Barley (yellow) Wheat (green) Wheat (yellow) Oat Spelt Rye Stubble Field (being harvested) Stubble Field (harvested, stramineous) Stubble Field (harvested, bare) Stubble Field (seeded, partly green) Stubble of Maize and Barley Mixed Fallow Sugar Beet Colza Oilseed Potatoes (green) Potatoes (yellow) Alfalfa Grassland without trefoil Grassland with trefoil Pasture Land Natural Pasture Land with Yarrow Field (harvested) Field (seeded) Field (ploughed) Synthetic Silo balls Water Waste Water Treatment Plant Abbreviation M G1 G2 W1 W2 H D Ro S1 S2 S3 S4 MSG MSB Z R O K1 K2 L GS Wi We SG A1 A2 A3 SB AQ ARA 58

71 3. ENVI Header File ENVI samples = 616 lines = 845 bands = 37 header offset = 0 data type = 2 byte order = 1 interleave = BSQ file type = Other description = {ATCOR3 scene: 17/08/05 solar zen=35.9, solar azi=153.3, fcref= 10.0, fctem= 10.0, offtem= 0} map info = {Arbitrary, 1, 1, , , e+01, e+01, 1, units=meters} z plot titles = { wavelength [!7l!3m]!N, reflectance [%*10] } band names = {band 1 refl [%*10], band 2 refl [%*10], band 3 refl [%*10], band 4 refl [%*10], band 5 refl [%*10], band 6 refl [%*10], band 7 refl [%*10], band 8 refl [%*10], band 9 refl [%*10], band*10 refl [%*10], band*11 refl [%*10], band*12 refl [%*10], band*13 refl [%*10], band*14 refl [%*10], band*15 refl [%*10], band*16 refl [%*10], band 17 refl [%*10], band 18 refl [%*10], band*19 refl [%*10], band*20 refl [%*10], band*21 refl [%*10], band 22 refl [%*10], band*23 refl [%*10], band*24 refl [%*10], band 25 refl [%*10], band 26 refl [%*10], band*27 refl [%*10], band*28 refl [%*10], band*29 refl [%*10], band*30 refl [%*10], band*31 refl [%*10], band*32 refl [%*10], band*33 refl [%*10], band*34 refl [%*10], band*35 refl [%*10], band*36 refl [%*10], band 37 refl [%*10]} wavelength = { , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , } fwhm = { , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , } 59

72 4. Abbreviation of the required format of an image metadata XML document (specified by an XML schema stored in the PCI ImageMetadata.xsd file in $PCIHOME/etc) The two first and the last band (224) is listed below. < xml version="1.0" > <PCI-ImageMetadata xmlns=" xmlns:xsi=" xsi:schemalocation=" PCIImageMetadata.xsd"> <!-- *************** --> <!-- global metadata --> <DatasetDescription>Reformatted from Free AVIRIS Standard Data Product f970619t01p02_r02.tar.gz (Cuprite)</DatasetDescription> <NumberOfBands>224</NumberOfBands> <NumberOfRadTrans>1</NumberOfRadTrans> <SensorModelName>AVIRIS-1997</SensorModelName> <SensorType>whiskbroom</SensorType> <NominalLocation> <Latitude> </Latitude> <Longitude> </Longitude> <Height>21036</Height> </NominalLocation> <NominalTime> <Year>1997</Year> <Month>06</Month> <Day>19</Day> <Hour>20</Hour> </NominalTime> <NominalHeading>179.8</NominalHeading> <ForeAftTilt>0</ForeAftTilt> <FieldOfView> <Left>15</Left> <Right>15</Right> </FieldOfView> <!-- ********************** --> <!-- band-specific metadata --> <Band> <StorageChannel>1</StorageChannel> <ID>1</ID> <Quality>plot</Quality> <ResponseProfile> <Type>gaussian</Type> <Centre>369.85</Centre> <FWHM>9.61</FWHM> </ResponseProfile> <RadiometricTrans> <Index>1</Index> <Quantity>apprad</Quantity> <Gain>0.0020</Gain> 60

73 <Offset>0</Offset> </RadiometricTrans> </Band> <Band> <StorageChannel>2</StorageChannel> <ID>2</ID> <Quality>plot</Quality> <ResponseProfile> <Type>gaussian</Type> <Centre>379.69</Centre> <FWHM>9.58</FWHM> </ResponseProfile> <RadiometricTrans> <Index>1</Index> <Quantity>apprad</Quantity> <Gain>0.0020</Gain> <Offset>0</Offset> </RadiometricTrans> </Band>... <Band> <StorageChannel>224</StorageChannel> <ID>224</ID> <Quality>plot</Quality> <ResponseProfile> <Type>gaussian</Type> <Centre> </Centre> <FWHM>10.03</FWHM> </ResponseProfile> <RadiometricTrans> <Index>1</Index> <Quantity>apprad</Quantity> <Gain>0.0010</Gain> <Offset>0</Offset> </RadiometricTrans> </Band> </PCI-ImageMetadata> 61

74 5. CHRIS/PROBA XML File <?xml version="1.0"?> <PCI-ImageMetadata xmlns=" xmlns:xsi=" xsi:schemalocation=" PCIImageMetadata.xsd"> <!-- *************** --> <!-- global metadata --> <DatasetDescription> ATCOR3 scene: 17/08/05 solar zen=35.9, solar azi=153.3, fcref=10.0, fctem=10.0, offtem=0 </DatasetDescription> <NumberOfBands>37</NumberOfBands> <NumberOfRadTrans>0</NumberOfRadTrans> <SensorModelName>CHRIS/PROBA</SensorModelName> <SensorType>pushbroom</SensorType> <NominalTime> <Year>2005</Year> <Month>08</Month> <Day>17</Day> </NominalTime> <!-- ********************** --> <!-- band-specific metadata --> <Band> <StorageChannel>1</StorageChannel> <ID>1</ID> <Quality>plot</Quality> <ResponseProfile> <Type>gaussian</Type> <Centre>442.0</Centre> <FWHM>9.0</FWHM> </ResponseProfile> </Band> <Band> <StorageChannel>2</StorageChannel> <ID>2</ID> <Quality>plot</Quality> <ResponseProfile> <Type>gaussian</Type> <Centre>489.0</Centre> <FWHM>9.0</FWHM> </ResponseProfile> </Band> 62

75 63 <Band> <StorageChannel>3</StorageChannel> <ID>3</ID> <Quality>plot</Quality> <ResponseProfile> <Type>gaussian</Type> <Centre>530.0</Centre> <FWHM>9.0</FWHM> </ResponseProfile> </Band> <Band> <StorageChannel>4</StorageChannel> <ID>4</ID> <Quality>plot</Quality> <ResponseProfile> <Type>gaussian</Type> <Centre>551.0</Centre> <FWHM>10.0</FWHM> </ResponseProfile> </Band> <Band> <StorageChannel>5</StorageChannel> <ID>5</ID> <Quality>plot</Quality> <ResponseProfile> <Type>gaussian</Type> <Centre>570.0</Centre> <FWHM>8.0</FWHM> </ResponseProfile> </Band> <Band> <StorageChannel>6</StorageChannel> <ID>6</ID> <Quality>plot</Quality> <ResponseProfile> <Type>gaussian</Type> <Centre>631.0</Centre> <FWHM>9.0</FWHM> </ResponseProfile> </Band> <Band> <StorageChannel>7</StorageChannel> <ID>7</ID> <Quality>plot</Quality> <ResponseProfile> <Type>gaussian</Type> <Centre>661.0</Centre> <FWHM>11.0</FWHM> </ResponseProfile> </Band> <Band>

76 64 <StorageChannel>8</StorageChannel> <ID>8</ID> <Quality>plot</Quality> <ResponseProfile> <Type>gaussian</Type> <Centre>672.0</Centre> <FWHM>11.0</FWHM> </ResponseProfile> </Band> <Band> <StorageChannel>9</StorageChannel> <ID>9</ID> <Quality>plot</Quality> <ResponseProfile> <Type>gaussian</Type> <Centre>683.0</Centre> <FWHM>11.0</FWHM> </ResponseProfile> </Band> <Band> <StorageChannel>10</StorageChannel> <ID>10</ID> <Quality>plot</Quality> <ResponseProfile> <Type>gaussian</Type> <Centre>697.0</Centre> <FWHM>6.0</FWHM> </ResponseProfile> </Band> <Band> <StorageChannel>11</StorageChannel> <ID>11</ID> <Quality>plot</Quality> <ResponseProfile> <Type>gaussian</Type> <Centre>703.0</Centre> <FWHM>6.0</FWHM> </ResponseProfile> </Band> <Band> <StorageChannel>12</StorageChannel> <ID>12</ID> <Quality>plot</Quality> <ResponseProfile> <Type>gaussian</Type> <Centre>709.0</Centre> <FWHM>6.0</FWHM> </ResponseProfile> </Band> <Band> <StorageChannel>13</StorageChannel>

77 65 <ID>13</ID> <Quality>plot</Quality> <ResponseProfile> <Type>gaussian</Type> <Centre>716.0</Centre> <FWHM>6.0</FWHM> </ResponseProfile> </Band> <Band> <StorageChannel>14</StorageChannel> <ID>14</ID> <Quality>plot</Quality> <ResponseProfile> <Type>gaussian</Type> <Centre>722.0</Centre> <FWHM>6.0</FWHM> </ResponseProfile> </Band> <Band> <StorageChannel>15</StorageChannel> <ID>15</ID> <Quality>plot</Quality> <ResponseProfile> <Type>gaussian</Type> <Centre>728.0</Centre> <FWHM>7.0</FWHM> </ResponseProfile> </Band> <Band> <StorageChannel>16</StorageChannel> <ID>16</ID> <Quality>plot</Quality> <ResponseProfile> <Type>gaussian</Type> <Centre>735.0</Centre> <FWHM>7.0</FWHM> </ResponseProfile> </Band> <Band> <StorageChannel>17</StorageChannel> <ID>17</ID> <Quality>plot</Quality> <ResponseProfile> <Type>gaussian</Type> <Centre>742.0</Centre> <FWHM>7.0</FWHM> </ResponseProfile> </Band> <Band> <StorageChannel>18</StorageChannel> <ID>18</ID>

78 66 <Quality>plot</Quality> <ResponseProfile> <Type>gaussian</Type> <Centre>748.0</Centre> <FWHM>7.0</FWHM> </ResponseProfile> </Band> <Band> <StorageChannel>19</StorageChannel> <ID>19</ID> <Quality>plot</Quality> <ResponseProfile> <Type>gaussian</Type> <Centre>755.0</Centre> <FWHM>7.0</FWHM> </ResponseProfile> </Band> <Band> <StorageChannel>20</StorageChannel> <ID>20</ID> <Quality>plot</Quality> <ResponseProfile> <Type>gaussian</Type> <Centre>762.0</Centre> <FWHM>7.0</FWHM> </ResponseProfile> </Band> <Band> <StorageChannel>21</StorageChannel> <ID>21</ID> <Quality>plot</Quality> <ResponseProfile> <Type>gaussian</Type> <Centre>770.0</Centre> <FWHM>7.0</FWHM> </ResponseProfile> </Band> <Band> <StorageChannel>22</StorageChannel> <ID>22</ID> <Quality>plot</Quality> <ResponseProfile> <Type>gaussian</Type> <Centre>777.0</Centre> <FWHM>15.0</FWHM> </ResponseProfile> </Band> <Band> <StorageChannel>23</StorageChannel> <ID>23</ID> <Quality>plot</Quality>

79 67 <ResponseProfile> <Type>gaussian</Type> <Centre>792.0</Centre> <FWHM>8.0</FWHM> </ResponseProfile> </Band> <Band> <StorageChannel>24</StorageChannel> <ID>24</ID> <Quality>plot</Quality> <ResponseProfile> <Type>gaussian</Type> <Centre>800.0</Centre> <FWHM>8.0</FWHM> </ResponseProfile> </Band> <Band> <StorageChannel>25</StorageChannel> <ID>25</ID> <Quality>plot</Quality> <ResponseProfile> <Type>gaussian</Type> <Centre>872.0</Centre> <FWHM>18.0</FWHM> </ResponseProfile> </Band> <Band> <StorageChannel>26</StorageChannel> <ID>26</ID> <Quality>plot</Quality> <ResponseProfile> <Type>gaussian</Type> <Centre>886.0</Centre> <FWHM>10.0</FWHM> </ResponseProfile> </Band> <Band> <StorageChannel>27</StorageChannel> <ID>27</ID> <Quality>plot</Quality> <ResponseProfile> <Type>gaussian</Type> <Centre>895.0</Centre> <FWHM>10.0</FWHM> </ResponseProfile> </Band> <Band> <StorageChannel>28</StorageChannel> <ID>28</ID> <Quality>plot</Quality> <ResponseProfile>

80 68 <Type>gaussian</Type> <Centre>905.0</Centre> <FWHM>10.0</FWHM> </ResponseProfile> </Band> <Band> <StorageChannel>29</StorageChannel> <ID>29</ID> <Quality>plot</Quality> <ResponseProfile> <Type>gaussian</Type> <Centre>915.0</Centre> <FWHM>10.0</FWHM> </ResponseProfile> </Band> <Band> <StorageChannel>30</StorageChannel> <ID>30</ID> <Quality>plot</Quality> <ResponseProfile> <Type>gaussian</Type> <Centre>925.0</Centre> <FWHM>10.0</FWHM> </ResponseProfile> </Band> <Band> <StorageChannel>31</StorageChannel> <ID>31</ID> <Quality>plot</Quality> <ResponseProfile> <Type>gaussian</Type> <Centre>940.0</Centre> <FWHM>20.0</FWHM> </ResponseProfile> </Band> <Band> <StorageChannel>32</StorageChannel> <ID>32</ID> <Quality>plot</Quality> <ResponseProfile> <Type>gaussian</Type> <Centre>955.0</Centre> <FWHM>10.0</FWHM> </ResponseProfile> </Band> <Band> <StorageChannel>33</StorageChannel> <ID>33</ID> <Quality>plot</Quality> <ResponseProfile> <Type>gaussian</Type>

81 <Centre>965.0</Centre> <FWHM>11.0</FWHM> </ResponseProfile> </Band> <Band> <StorageChannel>34</StorageChannel> <ID>34</ID> <Quality>plot</Quality> <ResponseProfile> <Type>gaussian</Type> <Centre>976.0</Centre> <FWHM>11.0</FWHM> </ResponseProfile> </Band> <Band> <StorageChannel>35</StorageChannel> <ID>35</ID> <Quality>plot</Quality> <ResponseProfile> <Type>gaussian</Type> <Centre>987.0</Centre> <FWHM>11.0</FWHM> </ResponseProfile> </Band> <Band> <StorageChannel>36</StorageChannel> <ID>36</ID> <Quality>plot</Quality> <ResponseProfile> <Type>gaussian</Type> <Centre>997.0</Centre> <FWHM>11.0</FWHM> </ResponseProfile> </Band> <Band> <StorageChannel>37</StorageChannel> <ID>37</ID> <Quality>plot</Quality> <ResponseProfile> <Type>gaussian</Type> <Centre>1019.0</Centre> <FWHM>33.0</FWHM> </ResponseProfile> </Band> </PCI-ImageMetadata> 69

82 6. ENVI Pixel Purity Index Image for CHRIS/PROBA Dataset (see chapter ) 70