Advanced Techniques for Automatic Change Detection in Multitemporal Hyperspectral Images

The increasing availability of the new generation remote sensing satellite hyperspectral images provides an important data source for Earth Observation (EO). Hyperspectral images are characterized by a very detailed spectral sampling (i.e., very high spectral resolution) over a wide spectral wavelength range. This important property makes it possible the monitoring of the land-cover dynamic and environmental evolution at a fine spectral scale. This also allows one to potentially detect subtle spectral variations associated with the land-cover transitions that are usually not detectable in the traditional multispectral images due to their poor spectral signature representation (i.e., generally sufficient for representing only the major changes). To fully utilize the available multitemporal hyperspectral images and their rich information content, it is necessary to develop advanced techniques for robust change detection (CD) in multitemporal hyperspectral images, thus to automatically discover and identify the interesting and valuable change information. This is the main goal of this thesis. In the literature, most of the CD approaches were designed for multispectral images. The effectiveness of these approaches, to the complex CD problems is reduced, when dealing with the hyperspectral images. Accordingly, the research activities carried out during this PhD study and presented in this thesis are devoted to the development of effective methods for multiple change detection in multitemporal hyperspectral images. These methods consider the intrinsic properties of the hyperspectral data and overcome the drawbacks of the existing CD techniques. In particular, the following specific novel contributions are introduced in this thesis: 1) A theoretical and empirical analysis of the multiple-change detection problem in multitemporal hyperspectral images. Definition and discussion of concepts as the changes and of the change endmembers, the hierarchical change structure and the multitemporal spectral mixture is given. 2) A novel semi-automatic sequential technique for iteratively discovering, visualizing, and detecting multiple changes. Reliable change variables are adaptively generated for the representation of each specific considered change. Thus multiple changes are discovered and discriminated according to an iterative re-projection of the spectral change vectors into new compressed change representation domains. Moreover, a simple yet effective tool is developed allowing user to have an interaction within the CD procedure. 3) A novel partially-unsupervised hierarchical clustering technique for the separation and identification of multiple changes. By considering spectral variations at different processing levels, multiple change information is adaptively modelled and clustered according to spectral homogeneity. A manual initialization is used to drive the whole hierarchical clustering procedure; 4) A novel automatic multitemporal spectral unmixing approach to detect multiple changes in hyperspectral images. A multitemporal spectral mixture model is proposed to analyse the spectral variations at sub-pixel level, thus investigating in details the spectral composition of change and no-change endmembers within a pixel. A patch-scheme is used in the endmembers extraction and unmixing, which better considers endmember variability. Comprehensive qualitative and quantitative experimental results obtained on both simulated and real multitemporal hyperspectral images confirm the effectiveness of the proposed techniques.