Global seamless Aerosol Optical Depth (AOD) dataset with high accuray is essential for understanding the aerosols impacts on human health, weather, and climate. Traditional AOD reconstruction models are designed to fill gaps in satellite-derived AOD caused by cloud cover, precipitation, or limited satellite overpasses. These models typically rely on a single satellite product as the target variable, with auxiliary inputs such as meteorological and reanalysis data. However, their applicability at the global scale is limited by sample imbalance (e.g. sparse coverage in polar regions and nighttime) and by the inability to effectively incorporate lower-quality satellite products. To overcome these limitations, this study proposes an AOD fusion framework that integrates measurement adjustment theory with a machine learning algorithm to combine multiple satellite AOD products with the MERRA−2 reanalysis dataset. Validation results show that the fused AOD estimates exhibit stronger agreement with AERONET Solar/Lunar observations (R2 = 0.76/0.66) compared to both traditional reconstruction models (R2 = 0.68/0.53) and MERRA−2 AOD (R2 = 0.66/0.60), for daytime/nighttime conditions. The fusion model effectively addresses sample imbalance issues, which often cause traditional models to overestimate AOD, especially in polar regions (e.g. Antarctica, by 0.14) and at night (by 0.06), due to the absence of satellite observations. This improvement stems from the model’s ability to leverage the complementary strengths of multiple datasets through measurement adjustment theory. By using MERRA−2 as a spatiotemporal continuous reference during training, the model not only fills data gaps but also benefits from the enhanced regional performance of individual satellite products. For example, incorporating NOAA and SNPP DT AOD, previously underutilized in reconstruction due to low standalone accuracy, led to further accuracy gains over dark surfaces such as oceans, improving R2 by 0.01 and reducing RMSE by 0.0027. Notably, the fusion model reduces the annual mean AOD by ~40% compared to reconstruction model, highlighting its potential to correct systematic overestimation in data-sparse regions and produce more reliable high-frequency global AOD datasets.
Ding et al. (Tue,) studied this question.