LAUSR

In this work, the Learning-Based Approach to Atmospheric Compensation (LBAC) of hyperspectral data proposed by Acito et al. is extended. LBAC makes use of machine learning methods to directly estimate the spectral reflectance from the at-sensor radiance accounting for the variability induced by one or more unknown atmospheric parameters and by-passing their estimation. LBAC training is obtained by exploiting a spectral reflectance library and accounting for the effects of both the atmosphere and the noise. However, depending on the spectral library adopted, some specific spectra may be reconstructed with lower accuracy. To overcome this drawback, two solutions are proposed referring to two application scenarios. The former deals with small and rare anomalous pixels with unknown reflectance and could be of interest in many applications such as man-made targets detection. It leverages the strengths of LBAC and those of the empirical line method (ELM). The second scenario refers to the case of materials with a priori known spectral reflectance and is defined for applications such as mining exploration and contaminant detection. It directly acts on the training phase of LBAC by introducing the spectra of interest in the generation of the training set. An extensive analysis is carried out on simulated data to test the effectiveness of the proposed solutions, to discuss their strengths and weakness, and to compare them with a classical physics-based approach. Results on a real hyperspectral image acquired by an airborne sensor provide a demonstration of the effectiveness of the proposed strategies in a real application environment.