Improved 1-km-Resolution Hourly Estimates of Aerosol Optical Depth Using Conditional Generative Adversarial Networks

Abstract

Aerosol Optical Depth (AOD) is a crucial parameter for various environmental and climate studies. Merging multi-sensor AOD products is an effective way to produce AOD products with more spatiotemporal integrity and accuracy. This study proposed a conditional generative adversarial network architecture (AeroCGAN) to improve the estimation of AOD. It first adopted MODIS Multiple Angle Implication of Atmospheric Correction (MAIAC) AOD data to training the initial model, and then transferred the trained model to Himawari data and obtained the estimation of 1-km-resolution, hourly Himawari AOD products. Specifically, the generator adopted an encoder–decoder network for preliminary resolution enhancement. In addition, a three-dimensional convolutional neural network (3D-CNN) was used for environment features extraction and connected to a residual network for improving accuracy. Meanwhile, the sampled data and environment data were designed as conditions of the generator. The spatial distribution feature comparison and quantitative evaluation over an area of the North China Plain during the year 2017 have shown that this approach can better model the distribution of spatial features of AOD data and improve the accuracy of estimation with the help of local environment patterns.