Downscaling SMAP soil moisture using a wide & deep learning method over the Continental United States

•A new SM downscaling framework based on the Wide & Deep Learning method is proposed to improve the spatial resolution of SMAP SM.•The proposed downscaling framework is conducted for the entire CONUS.•The downscaled SM at 1-km spatial resolution is validated using in-situ measurements collected from 211 sites distributed across the CONUS. Soil moisture (SM) plays a critical role in drought monitoring, agricultural management, flood forecasting, and other practical applications. However, the relatively coarse spatial resolutions of SM products derived from passive microwave satellite retrievals (approximately 25–55 km) greatly hamper their local-scale applications. In this research, we proposed an SM downscaling framework based on the Wide & Deep Learning (WDL) method to improve the spatial resolution of the level-3 daily composite of Soil Moisture Active Passive (SMAP) radiometer SM product (L3_SM_P). In this method, horizontally and vertically polarized Brightness Temperature (TBh, and TBv, respectively), surface reflectance and Land Surface Temperature (LST), topographic attributes, soil properties, climate types, and landcover types collected in the Continental United States (CONUS) during the annual unfrozen season (April 1st to November 1st) from 2015 to 2017 were used as auxiliary datasets to downscale the spatial resolution of the SMAP SM (L3_SM_P) product from its original 36 km to 1 km. Precipitation and in-situ SM measurements obtained from 211 sites distributed across the International Soil Moisture Network (ISMN) over the CONUS were utilized to validate the downscaled SM. The results demonstrated that the correlation (R) between the downscaled and the in-situ SM ranged from 0.325 to 0.997; the average R value was 0.715. The unbiased Root Mean Square Error (ubRMSE) values ranged from 0.010 to 0.141 m3/m3, with an average ubRMSE of 0.041 m3/m3, which meets the accuracy of SMAP SM requirement of ubRMSE approximately 0.04 m 3/m3. The downscaled SM also showed good temporal consistency with the in-situ SM and exhibited a high response to the precipitation data. The downscaled SM not only maintained high spatial consistency with the original SMAP SM but also provides more detailed spatial SM variations.