- IMPROVING CLASS SEPARABILITY - A COMPARATIVE STUDY OF TRANSFORMATION METHODS FOR THE HYPERSPECTRAL FEATURE SPACE

Principle and practical aspects of three transformation methods for the hyperspectral feature space (MNF, DAFE, DBFE) are described. In two application cases (urban case: selected man-made materials, water and shadow; rural case: tree species) these methods are used for the discrimination of spectrally similar classes. Supervised classifications are conducted in the original feature space as well as in each of the transformed feature spaces. The achieved separabilities of the selected classes are compared and discussed with respect to the different transformation methods. The analysis is made on 2003 HyMap data (3 m GSD) for an area in and close to the city of Osnabrück in Lower Saxony. INTRODUCTION Hyperspectral sensors measure the surface reflectance of electromagnetic irradiation using a huge number of narrow adjacent bands. Due to the small spectral distance between two adjacent bands, they are often highly correlated, so that for a specific classification problem usually not all bands are needed. In fact, having to many bands can lead to worse classification results caused by two reasons: First, if there were some “good” bands that contain information for the separation of the given classes and some others that don’t and to all bands the same weight was given in classification, then the other bands would degrade the classification accuracy. To solve this problem, feature selection methods (e.g. Bhattacharyya distance, transformed divergence (i)) were developed to enable the selection of n bands which, for each number n, would result in the best classification accuracy. Figure 1 shows the achieved classification accuracy for the urban study case depending on the number of bands which were selected with Bhattacharyya distance. In this figure the second reason for reducing the number of bands before classification can be seen. For a limited number of training samples adding more bands will first increase the classification accuracy, raise to a maximum and then decrease again. This decrease is due to an inaccurate estimation of class statistics and is known as the Hughes phenomenon. Figure 1: Classification accuracy for the classification of training (red line) and test samples (green line) © EARSeL and Warsaw University, Warsaw 2005. Proceedings of 4th EARSeL Workshop on Imaging Spectroscopy. New quality in environmental studies. Zagajewski B., Sobczak M., Wrzesień M., (eds) Moreover, in some cases the original feature space is suboptimal for separating spectrally similar classes. Consequently, transformation methods were developed that generate a smaller set of new bands by transforming the hyperspectral feature space with the aim of improving the separability of given classes. The methods have different principles, assets and drawbacks. Therefore, in this study three selected transformation methods – the Minimum Noise Fraction Transformation (MNF) (ii), the Discriminant Analysis Feature Extraction (DAFE) (i) and the Decision Boundary Feature Extraction (DBFE) (iii) – are compared and examined for their ability to aid discrimination between spectrally similar classes in two application cases. DATA AND TEST SITES The analysis was done for an urban and a rural test site in and close to the city of Osnabrück (Lower Saxony, Germany). The urban analysis was done on a hyperspectral dataset of DAIS 7915 (iv) with a ground resolution of 5 m, collected by the DLR in July 2002. The rural analysis was done on hyperspectral HyMap (v) data with a ground resolution of 3 m, collected by the DLR in July 2003. The DAIS data were processed by the DLR including rectification and radiometric calibration and were delivered as radiance data. Further pre-processing included atmospheric correction with FLAASH and the elimination of noisy bands. The HyMap data was delivered as radiometric calibrated radiance data. Rectification was done with PARGE (vi), followed also by an atmospheric correction with FLAASH and the elimination of noisy bands. For the determination of noisy bands their signal-to-noise ratio (SNR) was calculated utilizing the “homogeneous area method” (e.g. vii). This method implies that the image data contains an area which is homogeneous for the sensor. Thus, the variance in this area is due to the noise. For the SNR calculation the overall mean of each band was taken for the magnitude of the signal and the magnitude of noise was approximately calculated by the standard deviation of a water area assumed to be homogeneous. This was done on radiance data. Figure 2 shows the SNR quality of both datasets which show a significant difference in the SWIR. Figure 2: Comparison of the Signal-to-Noise-Ratios (SNR) of DAIS 7915 and HyMap For both analyses – the urban and rural one – the focus of attention lies on the differentiation of selected spectrally similar classes. Training fields for the analyses were build from larger homogeneous regions – if available – otherwise from big single trees or single roofs, respectively, and were validated by high resolution image data and by ground truthing. The classes’ spectra (mean spectra of training fields) are shown in figures 3 and 4. The urban classes are all characterized by a low albedo and by a lack of strong absorption bands. The spectra of the tree species are characterized by a very similar shape.