<p>Atmospheric aerosols are closely related to climate phenomena such as Earth&#8217;s energy budget and the formation of clouds. Anthropogenic aerosols have rapidly increased since the industrial revolution, which is damaging to human health, leading to cardiovascular, respiratory, and allergic diseases. Consequently, comprehensive knowledge of aerosol distribution is critical, particularly at detailed spatial scales. AOD measures the vertically integrated extinction of solar radiation by atmospheric aerosol particles. Ground-based sun photometers and satellite remote sensing are mainly used to retrieve AOD as a trade-off relationship. Ground-based measurements have been considered ground truth data, and satellite remote sensing has been used to derive the spatial variability of AOD over vast areas in near real-time. The Moderate Resolution Imaging Spectroradiometer (MODIS) sensor onboard the Terra and Aqua satellite is one of the main operation instruments to retrieve AOD, which has conducted atmospheric observations for almost two decades. MODIS has two well-known aerosol retrieval algorithms, Dark Target (DT) and Deep Blue (DB). Recently, the Multi-Angle Implementation of Atmospheric Correction (MAIAC) algorithm was developed to retrieve high-resolution AOD at a 1 km scale with better performance than DT and DB products. However, DT, DB, and MAIAC algorithms used radiative transfer models (RTM) and lookup tables (LUTs). LUTs was precalculated for a specific aerosol model using meteorological data, atmospheric gases, and constant geometry viewings, which required a high computation. The current LUT-based AOD model has reported uncertainties by aerosol model assumptions. Thus, there is room for complementing the existing AOD retrieval. Recently, machine learning (ML) has been applied with great performance for AOD retrieval. The ML-based AOD retrievals can be processed much faster and simpler without sensitive assumptions of the existing MODIS AOD algorithms. This study developed ML-based AOD retrievals that produce different resolutions of AODs (250m, 500m, and 1km) using MODIS data. The developed AODs at 250m, 500 m, and 1 km showed comparable performance, and 250 m AOD especially caught the spatial dynamics over urban areas well. When compared to MAIAC, 77.8 % of the 250 m AOD values are within the MODIS expected error (EE) envelope of &#177; (0.05 + 15%), followed by 500 km (76.5 % within EE), 1 km (76.3 % within EE), and MAIAC (70.08% within EE). Even ML-based AOD showed similar performance to MAIAC with three times more samples in the region where MAIAC AOD was unavailable. Our findings suggest the feasibility of ML-based estimation of high-resolution AOD using only satellite data.</p>
