Hyperspectral satellite image classification using small training data from its samples

Abstract

The paper is aimed to develop an appropriate multi-stage algorithm to perform thematic classification of hyperspectral satellite images based on small training sets of data selected manually from localized image parts. This algorithm is needed in the case of the lack of sensing condition data and quite frequent in practice. The chosen technology includes the stages of pixel-wise and spatial pre-processing, image classification based on spatial and spectral factors, as well as spatial post-processing of the classification results. Experimental studies have shown a significant decrease in the classification quality under the conditions considered in the paper. As a result of the experiments, the highest results were achieved by the algorithm based on combining pixel-wise classification results and segmentation results obtained using k-means++ and connected components labeling, supplemented by nonlinear pre- and post-processing methods. The findings are supported by the results of comparative studies of different methods at each stage of the classification.