(1) This article uses MODIS products MOD09A1, MOD11A1, MOD13A2, MOD15A2H, and MCD43D58 (Table 2). 1 kilometer of LST per day is provided by MOD11A1, and 1 kilometer of 16 day EVI and NDVI is provided by MOD13A2. MOD15A2H provides 8d Leaf Area Index (LAI) with a spatial resolution of 500 m, while MCD43D58 provides daily albedo data with a spatial resolution of 30 arcseconds (∼ 1000 m). Some indices related to soil moisture, including NDWI, NSDSI, and Surface Water Index (LSWI), were produced by MOD09A1.

    (2) Terrain factors are closely related to SM, including altitude, slope, and aspect, obtained from the Land Process Distributed Activity Archive Center (LP DAAC)

    (3) The 1000 meter resolution soil data used in this study, including the proportions of sand, silt, and clay, was obtained from the China Soil Characteristics Dataset (CSCD) obtained from the National Qinghai Tibet Plateau Data Center

    (4) The in-situ SM measurement values were collected from the data provided by the Maqu monitoring network and Babao monitoring network.

    (5) The daily precipitation and temperature data come from 131 meteorological stations of the China Meteorological Data Service Center.

    (1) Based on selected variable indicators (mainly including terrain data, soil data, and some MODIS products) and machine learning methods, a SMAP SM downscaling framework was constructed using multiple machine learning methods, and machine learning methods widely used to construct SM and its related variable regression models were selected;

    (2) Firstly, all data needs to be preprocessed. The daily LST data may be affected by clouds, so we use its quality control (QC) band to perform quality control on MOD11A1 products and select high-quality cloud free pixels. All selected variables, including LST, albedo LAI、NDWI、LSWI、NSDSI、NDVI、EVI、DEM、 Slope, aspect, sand, silt, and clay are all aggregated in GeoTIFF format to a resolution of 1 km. Use nearest neighbor interpolation to further resample these variables to achieve spatial resolution of SMAP SM data (36 km);

    (3) Secondly, obtain valid samples and split them;

    (4) Thirdly, determine the regression model based on the training and testing sets. Considering that the sample size is crucial for the accuracy of the regression model, we only selected periods with more than 100 samples to construct the model;

    (5) Finally, rotate the hyperparameters and select the optimal model.