Sparse Unmixing-Based Change Detection for Multitemporal Hyperspectral Images

In recent years, the increased availability of spectral libraries has resulted in a growing interest in sparse unmixing, which aims to find an optimal subset of library signatures to represent the pixels of remotely sensed hyperspectral datasets as linear combinations of these signatures. Sparse unmixing sidesteps two important drawbacks of the regular spectral unmixing process, namely the difficulty of estimating the number of endmembers, and the process of extracting the endmembers itself, the result of which will vary according to the utilized extraction method. In this work, sparse unmixing is exploited for the first time in the context of multitemporal hyperspectral data analysis and change detection. Change detection by sparse unmixing based on spectral libraries has the important advantage of providing not only pixel-level but also subpixel-level change information for the hyperspectral data. The changes that occur in multitemporal datasets due to time or as a result of a significant event are revealed, at subpixellevel, as the abundances of underlying endmembers within a pixel, or as variations in the distribution of these endmembers throughout the scene. The proposed approach is validated by experimental studies on both carefully prepared synthetic datasets and real datasets, using different spectral libraries.