The Effect of the Spectral Band Selection Method on Hyperspectral Images Segmentation Accuracy

Abstract

Abstract- In recent years, the image processing intelligent based systems has been the subject of interest for many researchers. In this manner, interpretation of urban aerial hyperspectral textured images is lionizing due to special features of this images. Image segmentation is the first step to reach this aim and extract features of these images. The correct selection of spectral bands is very important, because of multiplicity of spectral bands in this images and variety of texture in each of spectral bands. Since not all spectral bands include useful information, taking into account all of spectral bands decreases the speed of processing and accuracy of segmentation. In this paper, several dimension reduction or in other words spectral band reduction approach studied. In addition, the effect of each dimension reduction algorithms on accuracy of segmentation represented. Keywords-Hyperspectral Image, Dimension Reduction, Extended Mathematical Morphology, Region Growing I. INTROUDUCTION Image segmentation is one of the most important steps in image analysis and interpretation. In this way, hyperspectral image segmentation becomes interesting subject of many researches because of special properties of these images. The precision of segmentation in these images is depending on many parameters that one of most important parameters is proper spectral band selection. Suitable spectral band selection is important because of existing information in each band (e.g. texture) is different with other bands also none of spectral bands is contained of useful information. Therefore, spectral bands that contain more information must be select. There are many dimension reduction algorithms such as independent component analysis (ICA), principal component analysis (PCA) and discrete wavelet transform (DWT) [1]. All of existing approaches work depending on best band selection. Then, for processing, we need different approach when we are going to use several bands and using one band was not sufficient for segmentation. We use extended mathematical morphology for feature extraction. Image segmentation is one of the most critical tasks in automatic image analysis because the segmentation results will affect all the subsequent processes of image analysis, such as object representation and description, feature measurement and even the following higher-level tasks such as object classification and scene interpretation [2]. As mentioned before, discrimination of textures in different spectral bands is example of these features. On the other hand, we have different texture in various spectral bands. In general, this is considerable that none of segmentation algorithms is applicable at different application of image processing. Hence, propose different algorithm for any special application. With increasing number of algorithms for image segmentation, evaluation of performance of algorithms in studies is necessary. By take into account different hampers for image segmentation, proper algorithm selection is very important. However, texture characterization is particularly complex when the image data is composed of several spectral bands at different wavelengths, as in the case of remotely sensed hyperspectral images, in which hundreds of spectral bands are often available. Such images have two domains that can be analyzed: the spectral domain and the spatial domain. In this paper, we use joint direction and spectral features for hyperspectral image segmentation. By taking into account the complementary nature of spatial and spectral information in simultaneous fashion, it may be possible to alleviate the problems related to each of them taken separately and improve segmentation and classification results in urban analysis scenarios [3]. Under study image segmentation process, consists of three following stages: 1) Extracting of spectral and direction features and constructing feature image, using extended mathematical morphology. 2) Applying a suitable threshold on feature image, and create a binary image. 3) Finally using region growing algorithm and consider to spectral similarity between adjacent pixels, image segmentation is accomplished. The rest of this paper organized as follows. Section II describes extended mathematical morphology definition. In section III, we study several spectral band reduction (dimension reduction) algorithms. Experiment results discussed in Section IV and conclusions are given in Section V.