Retrieval of aerosol optical depth in Beijing based on deep learning

Based on the moderate-resolution imaging spectroradiometer（MODIS） remote sensing data and the aerosol robotic network（AERONET） global aerosol observation network data， a retrieval algorithm for aerosol optical depth （AOD） based on deep learning was proposed. Firstly， the advantage of massive samples of MODIS data was fully utilized， and a training dataset was constructed by combining with its original AOD products to complete the model pre‑training. Secondly， through the spatiotemporal accurate matching of high‑precision AOD data from AERONET and corresponding MODIS observation features， a high reliability input and output dataset was built， and the pre‑trained model was fine tuned and optimized with a small learning rate. This design realized the complementary advantages of the global feature learning capability from MODIS large‑sample data and the error correction capability from AERONET small sample high precision data. Finally， the retrieval results were compared with AERONET data. The proportion of samples falling within the expected error （EE=±（0.05+0.15τ）） was 83.3%，and the correlation coefficient was 0.91. Both core indicators were superior to the corresponding indicators of MODIS products （65.03% and 0.83）.