Classification of hyperspectral data using best-bases feature extraction algorithms

Mapping landcover type from airborne/spaceborne sensors is an important classification problem in remote sensing. Due to advances in sensor technology, it is now possible to acquire hyperspectral data simultaneously in more than 100 bands, each of which measures the integrated response of a target over a narrow window of the electromagnetic spectrum. The bands are ordered by their wavelengths and spectrally adjacent bands are generally statistically correlated. Using such high dimensional data for classification of landcover potentially provides greatly improved results. However, it is necessary to select bands that provide the best possible discrimination among the classes of interest. Treating the wavelength ordered bands of a hyperspectral sensor as a "signal", we adapt the local discriminant bases (LDB) algorithm, developed for classification of signals, to the task of classifying hyperspectral data. In a previous work we proposed a top-down algorithm based on LDB for finding a set of discriminating bases for pairwise classification of hyperspectral data. In this paper, we propose a bottom-up algorithm for finding the best-bases that combines highly correlated adjacent bands for maximum discrimination between class pairs. The proposed algorithm extends LDB in three ways. Firstly, it can find bases localized anywhere in time (wavelength or band number in case of hyperspectral data) and can be of any length, secondly, the criteria for selection of bands favors grouping highly correlated adjacent bands that, when projected on their Fisher dimension, yield maximum discrimination, and finally, the search for the best discriminatory bases is done for each of the ((C)(2)) two class problems into which a C-class problem can be decomposed. Experiments on a 183 band AVIRIS data set for a 12 class problem show significant improvements in both classification accuracies and the number of features required for all 66 pairs of classes relative to LDB. The overall accuracy for the C class problem is also increased significantly. Domain knowledge regarding the importance of certain bands for discriminating specific groups of classes can also be extracted from the results.