Multiple geometry atmospheric correction for image spectroscopy using deep learning

The goal of this research is to develop a general deep learning solution for atmospheric correction and target detection using multiple hyperspectral scenes. It is assumed that the scenes differ only in range and viewing angles, that they are acquired in rapid sequence using an airborne sensor orbiting a target, and that the target and the atmosphere remain invariant within the time scale of the collection. Several hundred thousand hyperspectral simulations were performed using the MODTRAN model and were used to train the deep learning solution, as well as to validate the proposed method. The input to the deep learning solution is a matrix of the simulated radiances at the sensor as function of wavelength and elevation angles. The output is atmospheric upwelling, downwelling, and transmission. This solution is repeated for all or a subset of pixels in the scene. We focus on emissive properties of targets, and simulations are performed in the longwave infrared between 7.5 and 12 μm. Results show that the proposed method is computationally efficient and it can characterize the atmosphere and retrieve the target spectral emissivity within one order of magnitude errors or less when compared with the original MODTRAN simulations.