ICA-based Algorithm for Hypespectral image Classification, FastICA versus JadeICA

Hyperspectral imagery is an efficient way to urban scene description and for detecting and identifying of building roof materials, because of its fine spectral and spatial resolution and its ability to acquire simultaneously great number of spectral bands. The potential of such data can lead to overcome the different basic problems in remote sensing, including spectral mixing. The spectral unmixing and sub-pixel classification techniques are also the relatively new research area of remotely sensed data analysis which permits the sub-pixel estimation of land cover. These techniques consist of both supervised and unsupervised algorithms. In this paper, we have used two ICA (Independent Component Analysis)-based methods for hyperspectral images classification. These methods allow us to separate the sources related to different materials proportion and distribution whichexist in the scene, without any necessity to a priori knowledge, which can be considered as an unsupervised unmixing approach. The estimated components can be interpreted asdifferent types of materials with relatively pure spectra. Because of the inherent ambiguities of ICA-based methods, the estimated sources corresponding to materials abundances could not be used directly as the class membership values. So, to have thematerial maps, another decision step is needed. The Fuzzy logic technique has been applied to obtain a classification map. This technique includes either two hard and soft classification results. Both of two methods have been applied on two different scenes of CASI (Compact Airborne Spectrographic Imager), image sets over an urban area of Toulouse city in the south of France. The hard classification results have been evaluated by comparing each other and with ground truth data and also the Soft classification results have been evaluated by comparing each other and with the soft classification results of two capable sub-pixel classification methods, SAM (Spectral Angle Mapper)and MF (Matched Filtering), as referencedata. Two ICA-based methods have beenevaluated in either classification precision and processing time, too. We observed thatthis algorithm for two methods is capable for detection of natural and man-madedifferent materials which exist within the scene and also individual small targets.The only rest limitation of this algorithm is its need to an interpretation step by a human operator for selection of the meaningful components resulting from ICA.