AUGMENTING A HIERARCHICAL CLASSIFIER FOR HYPERSPECTRAL DATA BY EXPLOITING SPATIAL CORRELATION

Classification of hyperspectral data is a challenging problem because of large dimensionality of data involved and non-binary nature of input classes. Most current methods do not consider the continuity of geographical features. In this work, augmenting the feature set of a pixel by appending data from its spatial neighborhood is experimentally seen to improve performance. Random projection is also seen to achieve processing speedup without unacceptable loss of accuracy. Using the two techniques in conjunction is shown to improve both accuracy and time characteristics of a binary hierarchical classifier. The output of the above techniques is generated, measured and analyzed. 1. DATA Spectral data is in the form of a three dimensional array I(m,n, d) where m and n refer to the spatial position of a pixel on a map and d determines the number of frequency bands at which the image was taken. The input to the system is hyperspectral data of a geographical region and the output is the classified map of the region. We measure the accuracy of the classifier from the ground truth points. The data set is the AVIRIS satellite image of the area around the Kennedy Space Center [2], taken on March 23,1996. AVIRIS acquires data in 224 bands of 10 nm width with center wavelengths from 400-2500 nm. The KSC data, acquired from an altitude of approximately 20 km, have a spatial resolution of 18 m. After removing water absorption and low SNR bands, 176 bands remain for analysis. The data was classified into 13 classes. Training data were selected using land cover map