A Deep-Learning Framework for the Generalized Emulation of Atmospheric Radiative Transfer at Microwave Frequencies

Abstract

This paper presents an application of machine learning to the problem of the computational expense of atmospheric radiative transfer calculations from physics-based models. While in recent studies individual emulators have been created that emulate certain radiative transfer schemes for different sensors, we show that a general framework can be applied to emulate the radiative transfer method itself. A fully connected deep neural network has been trained to predict top-of-atmosphere microwave radiances in all-sky conditions over various land, ocean and sea ice surfaces at all frequencies from 1 to 300 GHz, and for all incidence angles from 0 to 65 degrees. The reference models against which the emulator has been trained include an Eddington two-stream radiative transfer model for one-dimensional and pseudo-three-dimensional microwave brightness temperature simulations. This neural network architecture has also been separately trained to reproduce the explicit calculations of the Community Radiative Transfer Model (CRTM) for a single microwave imager. In all cases, the model architecture is successfully able predict microwave brightness temperatures by being trained directly on the optical properties of the atmosphere and surface and therefore directly emulating the radiative transfer method. Compared to the explicit schemes used to produce the training data, the GPU-enabled emulator is able to produce brightness temperature simulations more than three orders of magnitude faster and with a root-mean-squared-error less than 0.0348 K, 0.0531 K, and 0.0246 K against the independent test data for the two Eddington models and CRTM, respectively. Additionally, preliminary training using a full 3D Monte Carlo scheme for high-resolution simulations shows promising results for the application of this method to "exact" radiative transfer methods, and these results are discussed in Section IV.