Direct aerosol optical depth retrievals using MODIS reflectance data and machine learning over East Asia

Abstract

Anthropogenic aerosols have rapidly increased since the industrial revolution and are harmful to human health. Moderate Resolution Imaging Spectroradiometer (MODIS) data are critical for retrieving aerosol properties worldwide. However, current MODIS aerosol optical depth (AOD) products require extensive computations and a precalculated lookup table. This study proposes assumption-free high-resolution AOD retrieval models based on the light gradient boosting machine method using MODIS data and ground-based observations over East Asia. The models were developed with three spatial resolutions: 250 m, 500 m, and 1 km. The results showed that 77.8% of the 250 m AOD values were within the MODIS expected error (EE) range, while 76.5%, 76.3%, and 70.08% of the 500 m, 1 km, and Multi-Angle Implementation of Atmospheric Correction (MAIAC) AOD values were within the EE range, respectively. Furthermore, an analysis of the time series and detailed spatial distribution of the proposed model-derived AOD based on data from the Korea–United States Air Quality campaign demonstrated the excellent quality of the 250 m AOD via further validation using a spatially independent dataset. The Shapley Additive exPlanations analysis identified the sensor zenith angle and top-of-atmosphere reflectance of the blue band as the key contributors to the models. In addition, while MAIAC has limited spatial coverage, the spatial frequency of the proposed direct AOD retrieval was nearly 1.5-times higher than that of the MAIAC AOD. Our findings confirmed that machine learning-based high-resolution AOD estimates can be obtained using only satellite data.