1. MCD19A2 data

From MODIS official website (https://modis.gsfc.nasa.gov//) to download.

2. MERRA-2 tavg1_2d_aer_Nx data

From NASA's Goddard earth science data and information service center (https://disc.gsfc.nasa.gov/) to download. (This data needs to be downloaded worldwide)

3. ERA5-Land data

From the European centre for medium-range weather forecasts (https://cds.climate.copernicus.eu/) to download. (This data needs to be downloaded for the range 0° E - 150° E, 81° N - 11° S)

1. Remote sensing data preprocessing:MERRA-2(0.5°×0.625°) and ERA5-Land(0.1°×0.1) were sampled to a resolution of 1 km by bilinear interpolation, respectively. ERA5-Land also needed to fill in the missing values by nearest neighbor interpolation before this, and then extracted the required data by spatio-temporal matching according to the nearest pixel. The aerosol optical thickness (AOD) with wavelength index (AE) <1 was selected from the AERONET ground station observation data in the study area as the target truth value of dust AOD.

2. Inversion method:The two models were trained respectively using the LightGBM model: (1) The complete model, that is, the model with no missing predictors, was used as the input training model (MCD19A2 data had a large number of missing data, for the non-missing part); (2) Incomplete model, that is, MCD19A2 has missing data, and only the predictors extracted from MERRA-2 and ERA5-Land data are used as input training models. The output of the two models needs to be averaged with the sand AOD data of MERRA-2 after downscaling before it can be used as the final inversion result.

Reference：Ke G, Meng Q, Finley T, et al. Lightgbm: A highly efficient gradient boosting decision tree[J]. Advances in neural information processing systems, 2017, 30.