Unsupervised hierarchical spectral analysis for change detection in hyperspectral images

This paper describes a novel unsupervised approach to change detection in multi-temporal hyperspectral remote sensing images based on hierarchical spectral analysis and dimensionality reduction. The uniform feature design (UFD) strategy is implemented on original hyperspectral data for decreasing the data dimensionality and building different levels of data sets from coarse to fine spectral resolutions. Significant changes that can be easily extracted from low resolution data are then eliminated in the next high resolution level, in order to both avoid the computation burden and the complexity due to the increased number of channels, as well as to improve the detection accuracy. In each level, independent component analysis (ICA) is used on the hyperdimensional difference image to further separate specific change targets into independent components, which can help us to better identify the target change information. Bi-temporal Hyperion hyperspectral images are used in our experiment for the vegetation change detection in coastal wetland areas. The results confirm the effectiveness of the proposed technique. By using the hierarchical spectral analysis, more subtle changes can be detected to fully exploit the information contained in hyperspectral data.