Ph.D. student under the direction of F. Pascal

Context :

        A hyperspectral image is made of hundreds of images corresponding to the same spatial area but for different wavelengths. This kind of imaging is particularly informative especially since the spectral resolution of these images is important. Thus, it enables to evaluate the kind of material or object present on an image more precisely than with only one image in one spectral band provided the materials spectral characteristics are known. Since hyperspectral images contain a wide range of information, this requires adapted statistical tools: this thesis concerns the development of these tools, with random matrix theory, and focus on unmixing and detection issues, often encountered in hyperspectral imaging. Unmixing consists in extracting the different materials or endmembers that are present in one area of the hyperspectral image when the spatial resolution is not sufficiently precise. As for detection issues, which aim at finding a particular object on a particular background, anomalies are to be distinguished from target: an anomaly is characterized by a statistical break in the background, and differs from target detection by the lack of knowledge of the source (the target or the anomaly) spectral characteristics.