A Deep Learning-Based Framework for Estimating the Earth’s TOA Shortwave Radiation Flux From DSCOVR EPIC

Space-based Earth observation is the most direct mean for observing the Earth’s top-of-atmosphere (TOA) radiation flux. Deep space-based Earth observation, such as the DSCOVR/Earth polychromatic imaging camera (EPIC) located around the Sun-Earth Lagrange point 1, has emerged as a promising perspective for Earth radiation budget studies. However, there are still no intended angular distribution models (ADMs) to convert sensor-measured radiance to TOA flux for DSCOVR/EPIC. This study aims to develop a deep-learning framework based on a combination of convolutional neural networks (CNNs) and a multilayer perception neural network (MLPNN) to derive Earth’s TOA shortwave flux and determine its application potential for the global shortwave flux. A total of 18 CNN shortwave flux retrieval models were developed for nine scene types and cloudy and clear-sky conditions, respectively. The models were trained with matchups of synchronous DSCOVR/EPIC observations and clouds and the Earth’s radiant energy system (CERES) SSF products, and compared with the EPIC-L2 albedo-derived TOA shortwave flux. MLPNN models were developed to relate the model-estimated flux to the global TOA shortwave flux from CERES SYN products. Results showed that CNN models produced excellent performance when compared with CERES SSF data, and obviously outperformed the EPIC-L2 albedo product; the EPIC-derived TOA shortwave flux could account for 93% variations of the global shortwave flux at the daily scale. Overall, this study demonstrated the effectiveness of the proposed deep learning-based framework for estimating TOA shortwave flux from deep space-based observations, and also revealed the potential of the DSCVOR/EPIC artificial neural network (ANN) framework for other future Earth’s radiation budget studies.