Simulation of visible light at night from infrared measurements using deep learning technique

Abstract

Improving visible imagery of clouds, especially at night, is of great importance for weather forecasting and warning. This study considered full-disc data obtained by the FengYun-4B satellite with the Advanced Geostationary Radiation Imager (AGRI) as an example, and adopted a deep neural network model to simulate the full-disc visible image using observed dynamic infrared data and surface background data to compensate for missing visible image data at night. The generated daytime visible image had the same characteristics as a corresponding GeoColor image, with an R2 value of 0.99. The simulated night-time visible light, especially in the visualization of low cloud, clearly displayed and highlighted the characteristics of low cloud and fog, and could play an important role in application to near real-time weather forecasting at night. The simulation method could also be applied to the FY-3D/MERSI and Himawari-9/AHI satellite systems and to other infrared bands, similar to the FY-4B/AGRI system.