Spatial Gap-Filling of SMAP Soil Moisture Pixels Over Tibetan Plateau via Machine Learning Versus Geostatistics

Soil moisture (SM) is a key variable in ecology, environment, agriculture, and hydrology. The Soil Moisture Active Passive (SMAP) satellite provides global SM products with reliable accuracy since 2015. However, significant gaps of SMAP SM appeared over Tibetan Plateau. Considering the important role of the Tibetan Plateau in global climate and environment, it is essential to develop methods to infill the gaps to generate seamless SMAP SM data. To address this issue, we proposed two methods, machine learning and geostatistics technique. For the machine learning technique, we train a Random Forest algorithm which aims to match the output of available SMAP L3 SM using a series of input variables such as SMAP brightness temperature (TB_H and TB_V) in ascending orbits (6:00 PM local time), surface temperature, MODIS NDVI, land cover, DEM, and other auxiliary data. Then, the established RF estimators were applied to the SMAP brightness temperature from descending orbits (6:00 AM local time) to reconstruct complete SM data over the Tibetan Plateau. For the geostatistics technique, the Ordinary kriging was applied to the available SMAP L3 SM pixels to interpolate complete SM data. To cross-validate the performances of the algorithms, we assume certain areas with available SMAP SM values as missing, and then compared the gap-filling results with the actual ones. The cross-validations show that the gap-filling results from two algorithms were highly correlated to the official SMAP SM products with high coefficients of determination (R²_RF = 0.97 and R²_OK = 0.85) and low RMSE (RMSE_RF = 0.015 cm³/cm³ and RMSE_OK = 0.036 cm³/cm³). Furthermore, the gap-filling SM data present a better correlation with the Soil Moisture and Ocean Salinity SM data (R = 0.55–0.7) than the Global Land Data Assimilation System simulations (R = 0.18–0.62). The reconstructed SM from RF (R = 0.71) and OK (R = 0.55) algorithms are well related to the Maqu network measurements. Thus, the machine learning and geostatistics algorithms have the potential to reproduce the missing SMAP SM products over the Tibetan Plateau.