Atmospheric correction of airborne infrared hyperspectral images using neural networks

The retrieval of surface emissivity and temperature from infrared radiances measured by an airborne hyperspectral sensor closely depends on the ability to correct the acquired data from atmospheric effects. In this paper we present a new atmospheric correction scheme based on sounding techniques and neural networks. A key problem of neural network is to select relevant entries and outputs. Therefore, a preliminary sensitivity analysis that takes into account atmospheric conditions as well as the surface emissivity and temperature variations is carried out. It shows that only the first three or four PCA coefficients of atmospheric profiles have a significant influence on the radiance measured in the 4.26 mu m carbon dioxide and the 6.7 mu m water absorption bands. But these coefficients allow to rebuilt temperature and water profiles with enough accuracy for the addressed problem. This lead us to develop two groups of neural networks, the first one to estimate the main PCA coefficients of temperature profile, and the second one to retrieve the related water PCA coefficients. The atmospheric profiles thus obtained are then used to derive the "ground" radiances. Eventually we evaluate the accuracy of surface temperature and emissivity obtained with the derived atmospheric profiles.