Learning in high dimensions for an automatic and robust target detection in hyperspectral imagery

The hyperspectral remote sensing system has four basic parts: (1) the illumination source (e.g. the Sun in passive remote sensing); (2) the atmospheric path; (3) the targeted scene; and (4) the airborne hyperspectral imaging sensor. The hyperspectral sensor is capable of simultaneously acquiring the same spatial scene in a contiguous and multiple narrow spectral wavelength color bands. When all these spectral bands are stacked together, this results a hyperspectral data cube with two spatial and one spectral dimension. Each pixel of the hyperspectral image follows a specific distribution which is usually completely unknown to the user. This distribution, even if known, can depend on two unknown parameters such as the mean and the covariance matrix. Due to the presence of the atmospheric path, the reflected energy (initially captured by the solar illumination and then modified by the atmosphere) from the material surface will be different from the one who reaches the airborne hyperspectral imaging sensor since it passes back through the atmosphere. In this regard, some effects that produce variability to the material spectra (e.g. atmospheric conditions, sensor noise, material composition, scene geometry) have to be taken into consideration. It is well known that the signal model for hyperspectral test pixels is fundamentally different from the additive model used in radar and communications applications. We can regard each test pixel as being made up of αt + (1 - α)b,  with 0 ≤ α ≤ 1, "t" the spectrum of the target of interest, and "b" the spectrum of the background (located at the same spatial location of t). A prior information about the target spectral signature may be available to the user and which can be exctrated from some online spectral libraries. Different classical target detectors (e.g. matched filter, matched subspace detector, adaptive subspace detector) have been proposed in the literature. The performance of these detectors mainly depend on the estimation quality of the mean and the covariance matrix. During this doctoral thesis, the shrinkage and sparsity techniques may be exploited to develop new estimators of the covariance matrix in high dimensions in order to achieve a robust target detection in hyperspectral imagery. Due to the lack of real hyperspectral images available online, the doctoral candidate will be engaged in generating new hyperspectral images that can help working in some more realistic scenarios usually encountered in the real world of military target detection. For example: (a) camouflage targets; (b) targets mixed with the background with very low target fill-fraction α; (c) targets embedded in a background containing a lot of small heterogeneous and/or high contrast regions; and (d) targets having spectra very close to the surroundings. The latter task is of high importance since the paradigm in military applications for hyperspectral imagery seems to center on finding the target but ignoring all the rest. Sometimes, that rest is important especially if the target is well matched to the surroundings. 

Research Unit: Laboratoire Informatique et Société Numérique (LIST3N)

Website LIST3N : http://recherche.utt.fr/list3n

Applicants must have an engineering or master's degree in the fields of signal and image processing. Skills are particularely expected in applied mathematics, statistical signal processing, and matlab programming. Knowledge of hyperspectral or optical imaging will be a plus.