Development of a NOAA One-kilometer Resolution SMOPS Climate Data Record

Soil moisture is one of the essential state variables of most numerical weather, climate and hydrological prediction models. The accuracy and reliability of this state variable substantially impact the performance of those model forecasts. The NESDIS Soil Moisture Operational Products System (SMOPS) blends global satellite soil moisture data products at 25km resolution using all currently available microwave satellite sensors (AMSR, SMAP, SMOS, ASCAT) for National Centers for Environmental Prediction (NCEP) users since 2013. However, the lack of a consistent long-term SMOPS product has led to pronounced fluctuations in data quality across distinct versions and remarkable uncertainties for climatological studies and prolonged data assimilation operations. To address these issues, NESDIS has reprocessed SMOPS with all available satellite soil moisture observations to generate a Climate Data Record (SMOPScdr). Responding to the data needs of high-resolution Numerical Weather Prediction (NWP) and water forecast models, algorithms based on machine learning approaches have been tested for downscaling the 25 km SMOPScdr blended soil moisture products to 1 km scales. The 1 km SMOPScdr data product takes advantage of four advances: 1) Instead of the thermal infrared land surface temperature, we used the microwave 36 GHz and 89 GHz brightness temperature to represent the surface temperature, allowing us to generate an all-sky 1 km SMOPScdr; 2) The developed 1 km SMOPScdr was bias-corrected to the climatology of National Water Model (NWM) 0-10 cm soil moisture climatology, and thus can be directedly assimilated into the NWM without the need of a bias correction preprocessing; 3) The trained machine learning model can preserve the reference data information beyond the training period, ensuring that it can be used to estimate real-time and historical long-term 1 km SMOPScdr without a new training procedure; 4) All input datasets used in training and implementing the selected machine learning model are purely based on satellite observations. Compared to the coarse resolution data product, the developed 1 km SMOPScdr shows better accuracy with much more spatial details.