Reconstructing MODIS aerosol optical depth and exploring dynamic and influential factors of AOD via random forest at the global scale

Aerosols pose a significant impact on the atmospheric environment and public health. Aerosol optical depth (AOD) is one of the essential parameters for quantifying aerosols. Currently, existing AOD products suffer from severe pixel value missing issue due to clouds and other interferences globally, which severely affects the ac-curacy of fine particulate matter estimation and related epidemics researches. At the global scale, this study first adopted population as one of the influencing factors of AOD, and fully considered the effects of humidity, surface pressure, cloud cover, and evapotranspiration on AOD. Meanwhile, the spatiotemporal resolution of meteoro-logical data was promoted to improve the AOD estimation precision. A random forest model was fitted to estimate the relationship between AOD and influencing factors in non-missing areas, and the missing AOD values were filled in based on this model. The spatiotemporal dynamic characteristics of AOD were revealed, and the importance of each influencing factor on AOD was elucidated. The results demonstrated that the scheme can effectively fill in missing AOD values, and the accuracy of the filled-in values was significantly improved with an R2 of 0.6683 using AERONET as the validation sites. After filling in the missing values, the global AOD spatial coverage rate approached 100%. AOD exhibited significant spatiotemporal heterogeneity, with higher values in East Asia, South Asia, Southwest Asia, West and Central Africa, and northern South America, possibly due to natural factors and human activities. The importance of each influencing factor on AOD was ranked in descending order as follows, surface temperature (0.15), population density (0.14), surface pressure (0.11), U-shaped wind (0.10), humidity (0.10), total cloud cover (0.09), V-shaped wind (0.09), boundary layer height (0.08), evapotranspiration (0.07), and total precipitation (0.03). This study can provide insights and references for reconstructing other remotely sensed datasets.