Using Machine Learning Algorithms to Build Relationships between Spectral Surface Reflectances for Aerosol Optical Depth Retrieval over Land from Satellites

For the past two decades, quantitative retrievals of aerosol optical depth (AOD) have been made from both polar-orbiting and geostationary satellites, and the results have been widely used in numerous studies. Despite the tremendous progress made in improving the accuracy of AOD retrievals, there are still major challenges, especially over land. A notable one for the so-called dark-target algorithms is building the surface reflectance (SR) relationships to derive SR at visible channels (e.g. 0.47 and 0.64 µm) from SR at short-wave infrared (SWIR) channel, partly because these relationships are strongly subjected to entangled factors, such as scattering angle, surface type, vegetation state, etc. Previous algorithms mainly used linear regressions to build the SR relationships between visible and SWIR channels with consideration of scattering angle and Normalized Difference Vegetation Index at SWIR (NDVI). However, such linear regressions usually lead to relationships characterized by large standard deviation, which in turn serves as one of the major sources of uncertainties in AOD retrievals. In this study, we apply an ensemble machine-learning (ML) algorithm for establishing the relationships between the SR at 0.47, 0. 64, and 2.2 µm with multiple related inputs (scattering angle, NDVI, elevation, surface type, and month) for the Himawari-8 geostationary satellite. The algorithm demonstrates excellence performance across Eastern Asia, with a great reduction in random noise. The coefficients of determination (R²) for SR relationships are largely improved from 0.56-0.8 for linear regressions to 0.94-0.96 for the ML algorithm. Such improvements in SR relationships lead to considerable reductions in the uncertainties in the AOD retrievals. ML algorithm can generate an AOD products with better accuracy, especially for low albedo regions. The method used in this study directly benefits the algorithm development for Himawari-8. It can also be adopted for other geostationary or polar-orbiting satellites. Our study provides insights into how artificial intelligence might improve the retrieval of AOD from multi-spectral satellite observations.