A novel hyperspectral segmentation algorithm—concept and evaluation

Nowadays, the hyperspectral imaging is the focus of intense research, because its applications can be very useful in the natural disaster monitoring and agricultural monitoring to enumerate only a few. The main problem of systems using hyperspectral imaging is the cost of labelling, because it requires the domain experts, who label the region or prepare the labelled learning set for machine learning methods. The article presents a novel Hyperspectral Segmentation Algorithm which is a part of a general framework used for image classification. The algorithm is based on an image decomposition into homogeneous regions using a novel similarity measure. Three different region representations are proposed using the matrix notation. An additional procedure merges similar regions into larger ones to reduce human expert engagement in region labelling. The algorithm has been evaluated on the number of benchmark datasets to investigate the influence of algorithm parameters on the final performance. Comparison with competing methods proved that the considered algorithm is an interesting proposition in hyperspectral image analysis tasks.