A framework for estimating all-weather fine resolution soil moisture from the integration of physics-based and machine learning-based algorithms

Due to the effects of radio frequency interference and the limitations of algorithms under specific conditions, most of the currently available microwave-based soil moisture (SM) products are spatially discontinuous and have coarse spatial resolution, whereas optical observations also reveal various data gaps due to cloud contamination. Hence, the prediction of SM over invalid pixels and disaggregation from coarse to high scales are two main processes for obtaining SM at fine spatiotemporal resolution (e.g., daily/1-km). In the present study, two methods with respect to disaggregation-first or prediction-first were investigated from the synergetic use of the widely recognized European Space Agency-Climate Change Initiative (ESA-CCI) SM product and Moderate Resolution Imaging Spectroradiometer (MODIS) images over the Tibetan Plateau (TP) region. Specifically, the Disaggregation based on Physical And Theoretical scale Change (DisPATCh) algorithm and the generalized regression neural network (GRNN) were implemented in the disaggregation and prediction, respectively. In DisPATCh, spatially complete land surface temperature (LST), normalized difference vegetation index (NDVI) and digital elevation model (DEM) were provided as essential inputs to downscale the microwave-based ESA-CCI to a spatial resolution of 1 km, whereas MODIS-derived LST, NDVI, land surface albedo and DEM were considered in the GRNN prediction. Following the two methods, the daily/1-km SM dataset over a period of three years was finally estimated. Assessments with ground in-situ SM measurements over the TP region reveal an acceptable accuracy with unbiased root mean square errors of ∼ 0.06 m3/m3, indicating the potential to obtain operational daily/1-km spatially continuous SM products in future developments.