HYPERSPECTRAL CHANGE DETECTION BY SPARSE UNMIXING WITH DICTIONARY PRUNING

Hyperspectral change detection is used in many applications ranging from environmental monitoring to city planning and military surveillance. Change detection by unmixing has the potential of not only providing the locations of the changes, but also the nature of the change, and sub-pixel level information. Change detection by sparse unmixing using spectral libraries, with respect to regular spectral unmixing, has the added benefits of circumventing the process of endmember extraction and providing specific mission-based information. However, sparse unmixing is generally a severely ill-conditioned and time-consuming problem. Recently, an approach that aims to identify a subset of signatures of the spectral library that contribute most to the observed data has been proposed in order to improve the conditioning of the problem, hence decreasing computation time and enhancing performance. In this work, sparse unmixing with dictionary pruning is explored for change detection in multi-temporal hyperspectral images. Experimental results validate the performance of the proposed approach.