Estimation of Land Surface Downward Shortwave Radiation Using Spectral-Based Convolutional Neural Network Methods: A Case Study From the Visible Infrared Imaging Radiometer Suite Images

Surface downward shortwave radiation (DSR) is a key parameter in Earth’s surface radiation budget. Many satellite products have been developed, but their accuracies need further improvements. This study proposed an innovative deep learning method that combines radiative-transfer (RT) modeling with convolutional neural network (CNN) learning for estimating instantaneous DSR from Visible Infrared Imaging Radiometer Suite (VIIRS) observations. Unlike traditional CNN methods that rely on spatial contextual information and are not optimal for medium to coarse resolution satellite data, the proposed algorithm takes advantage of both spectral information as well as vertical information. The algorithm first estimates the atmospheric effective optical depth from top of the atmosphere (TOA) and surface reflectance by using the lookup table (LUT) created by RT simulations. We then constructed a spectral-wised virtual matrix to train the CNN using surface DSR measurements at 34 baseline surface radiation network sites globally during 2013. The developed CNN was also compared with four traditional machine learning algorithms. The validation results showed that the root-mean-square error (RMSE) and the bias were 91.42 and −0.94 W/m2, respectively. This article is the first spectral-wised CNN application to estimate surface biophysical parameters from satellite remote sensing data quantitatively. The comparison with previous LUT and optimization-based algorithms shows that the proposed algorithm outperforms by around 10–20 W/m2. We also explored how transfer learning can further improve the DSR estimation. Our results indicate that the universal model with local data transfer learning outperforms either the CNN with local data or the universal CNN by around 10–20 W/m2.