An Artificial Neuron Network With Parameterization Scheme for Estimating Net Surface Shortwave Radiation From Satellite Data Under Clear Sky--Application to Simulated GF-5 Data Set

Net surface shortwave radiation (NSSR) is a key parameter that drives the surface material exchange and energy balance. Herein, we propose an improved artificial neuron network (ANN) with parameterized (ANN-P) method to first calculate the albedo at the top of atmosphere (TOA) by considering the surface non-Lambertian effect. Subsequently, the NSSR is estimated based on the relationship between TOA broadband albedo and the Earth’s surface-absorbed shortwave radiation using a parameterized method under clear sky. The modeling process is implemented with Chinese Gaofen-5 (GF-5) visible/near-infrared channels data simulated via MODTRAN. For comparison, a previously reported lookup table (LUT) with parameterized (LUT-P) method and an ANN method are also employed. The performances of all these methods are evaluated. In terms of model simulation part, the root-mean-square errors (RMSEs) are 15.01 (17.07), 10.04 (13.67), and 20.39 (29.99) W/m2 for land, water, and snow/ice surfaces, respectively, for the ANN-P (versus LUT-P) method. Their mean bias errors (MBEs) are within 0.9 W/m2. With respect to the direct ANN method, it shows the highest accuracy yet relatively large deviation for water surface. Additionally, the sensitivity analysis of water vapor content (WVC) confirms that the ANN-P method is more stable than the LUT-P and ANN methods and is, thereby, recommended for clear-sky NSSR estimation. Finally, the ground validations indicate that the mean RMSEs (MBEs) for the LUT-P, ANN-P, and ANN methods are 49.33 (−3.01), 47.55 (1.75), and 104.24 (−75.72) W/m2, respectively.