A generalized neural network for simultaneous retrieval of atmospheric profiles and surface temperature from hyperspectral thermal infrared data

This paper makes an attempt to establish a generalized neural network for simultaneously retrieving atmospheric profiles and surface temperature from hyperspectral thermal infrared data. To generate the simulated data covering the whole actual situations, the distributions of surface material, temperature and atmospheric profiles are elaborated carefully. The simulated at-sensor radiances are divided into two sub-ranges, one in atmospheric window and another in water absorption band. The simulated data are transformed in the eigen-domain in both sub-ranges and used as the network inputs. The atmospheric profiles, surface temperature and emissivity are used as the outputs after the eigen-domain transformation. The validation of the trained network indicates that a RMSE of surface temperature around 1.6K, a RMSE of temperature profiles around 2K in troposphere and a RMSE of total water content around 0.3g/cm² can be obtained. The results from the net can be used as initial guess of the physical retrieval model.