Abstract
Soil moisture (SM) is a critical factor influencing plant growth, agricultural yield, ecosystem functions, and water resource management. Existing SM products, such as SMAP, ERA5, and ESA-CCI, provide daily SM data. However, their coarse spatial resolution limits their application, necessitating downscaling for improved effectiveness. This study integrates multisource remote sensing, reanalysis, and ground observation data through machine learning models to generate two-layer SM data, aiming to improve the spatial resolution and accuracy of SM data. First, the ESTARFM spatiotemporal fusion algorithm was applied to combine high-resolution MODIS data with long-term GLASS data, generating daily SM driving variables (NDVI and LST) at a 1-km resolution. Subsequently, the LightGBM downscaling algorithm was used to reduce the spatial resolution of GLEAM SSM and RZSM data from 0.25 degrees to daily 1-km resolution. Finally, a bias correction method based on convolutional neural networks with transfer learning was employed for point-to-area fusion calibration to improve data accuracy. Experimental results show that the Pearson correlation coefficient (PCC) of SSM and RZSM data downscaled by LightGBM are 0.699 and 0.754, with root mean square error (RMSE) values of 0.053 and 0.055 m(3)/m(3), respectively. After calibration with ground-based observations, the PCC of the SM data ranges from 0.792 to 0.860, and the RMSE values range from 0.028 to 0.031 m(3)/m(3), showing significant improvement in accuracy. Further spatiotemporal and comparative analyses confirm that the generated two-layer SM data excels in capturing spatial and temporal dynamics. The study successfully generated high-precision, long-term time series SM data for the Korean Peninsula, providing reliable support for agricultural drought monitoring and drought forecasting, and offering valuable references for SM research in similar regions.