Estimation of the surface emission spectra from remotely sensed data based on the physics informed deep neural network

Surface emission spectra reflects the ability of a target to emit radiation of different wavelengths, and plays an important role in multiple fields as an inherent property of the target object. Currently, common methods require input of multiple data, whose acquisition will directly affect the accuracy and the timeliness of the results obtained. This article proposed a new physics-informed deep neural network for estimation of surface emission spectra, which not only utilizes the ability of deep learning models to handle nonlinear relationships and has been successfully applied to the inversion of various environmental parameters, but also combines physical knowledge of the maximum and minimum emissivity differences in multi band emission spectra to guide deep neural network training. The proposed model achieved accurate results in the simulated dataset. Compare the results of this article with ASTER surface emissivity products, which have similar spatial distributions and numerical ranges. However, the input of the new model only requires remotely sensed data, which has stronger robustness and timeliness.