Improving spatial resolution of satellite soil water index (SWI) maps under clear-sky conditions using a machine learning approach

•A new Soil Water Index (SWI) downscaling approach is introduced.•Land surface temperature had the greatest effect on the spatial variation of SWI.•Biophysical properties had greater impact than topographic and geographical.•The machine learning based approach showed strong potential in improving SWI. One of the limitations of daily Soil Water Index (SWI) products obtained from satellite imagery is the low spatial resolution, limiting their precise applications. The purpose of this study was to present a machine learning based approach to improve the spatial resolution of the SWI obtained from the Advanced Scatterometer (ASCAT). Surface biophysical, topographic, and geographical properties (environmental parameters) maps of three field sites from the United States of America (USA), France, and Iran were prepared with a spatial resolution of 30, 1000 and 10,000 m and their effects on SWI were investigated. A SWI estimation model was constructed based on a Random Forest (RF) regression using effective environmental parameters and used to map SWI at 1,000 and 30 m spatial resolutions. The final SWI map with an improved spatial resolution was prepared after applying a correction due to a residual error. Finally, the efficiency of the proposed model was evaluated based on measured soil moisture (SM) data recorded at ground stations. The results showed that land surface temperature had the greatest effect on the spatial distribution of SWI. The impact of surface biophysical properties on the SWI was greater than topographical and geographical properties. The mean SWI error in USA, France, and Iran at spatial resolution of 10,000 (improved 1000 m) for warm season were 23.6 % (15.8 %), 14.2 % (9.8 %) and 10.7 % (7.4 %), respectively. These values for cold season were 27.9 % (17.2 %), 15.3 % (13.2 %) and 15.5 % (8.8 %), respectively. Mean of R2 and RMSE between measured SM values and SWI 10,000 m (1000 m and 30 m) were 0.13 (0.43 and 0.73) and 17.6 (12.1 and 7.2 %), respectively. These values for cold season were 0.10 (0.39, 0.67), and 20.7 (14.3, 7.2 %), respectively. The proposed machine learning based approach showed strong potential in improving the spatial resolution of SWI and giving the opportunity for various precise applications.