Machine learning-based inversion of aerosol optical depth inversion from FY-4A data

This study aims to develop a machine learning algorithm based on the characteristics of AGRI data to generate an aerosol dataset with a high spatiotemporal resolution. Using aerosol data from 67 aerosol robotic network (AERONET) sites in China and its surrounding areas in 2021, this study selected data of factors such as apparent reflectance, observation angles, elevation, and MODIS surface reflectance acquired from FY-4A advanced geostationary radiation imager (AGRI)-a new generation geostationary meteorological satellite of China. Then, this study performed the inversion of aerosol optical depth (AOD) using four machine learning methods-random forest (RF), gradient boosting regression tree (GBRT), extreme gradient boosting (XGBoost), and back propagation neural network (BPNN). Using the optimal model determined based on evaluation metrics, this study achieved the AOD inversion with a spatial resolution of 4 km × 4 km based on FY-4A data. Then, this study compared the inversion results with MODIS aerosol products of roughly the same periods. The results indicate that the AOD inversion models based on the four machine learning algorithms yielded correlation coefficients (R) exceeding 0.90, mean absolute errors (MAEs) of less than 0.09, and root mean square errors (RMSE) below 0.14. This indicates that it is feasible to conduct AOD inversion using machine learning-based models. The GBRT-based model exhibited the highest inversion accuracy among the four methods, with a correlation coefficient of 0.82, MAE of 0.16, and RMSE of 0.25, as indicated by the verification results. Additionally, 47% of the inversion results fell within the expected error ranges, indicating that the results of AOD inversion from FY-4A data using the GBRT-based model were generally consistent with observed values. The comparison between the GBRT model-derived AOD inversion results and the results of MODIS aerosol products shows that the former exhibited high consistency with the latter in terms of spatial distribution, with 83.57% of grid deviations falling within the range from -1.0 to 0 and the former slightly higher than the latter.