Developing machine learning models with multisource inputs for improved land surface soil moisture in China

• The land surface soil moisture (SSM) was generally under- or overestimated by Fengyun-3C (FY-3C) data. • Five machine learning (ML) models were developed and compared to estimate accurate regional SSM. • The random forest regression (RfR) model provided the most accurate (R = 0.789) SSM monitoring results. • The geographical location played an important role for monitoring the regional SSM over China. • The spatio-temporal variation trend of the estimated SSM well matched the precipitation patterns. Accurate and spatially continuous land surface soil moisture (SSM) data will greatly benefit analyses of heat transfer, energy exchange and agricultural dryness. To obtain spatiotemporally consistent SSM information, five machine learning (ML) models, i.e., polynomial regression (PR), ridge regression (RR), lasso regression (LR), elastic net regression (EnR) and random forest regression (RfR) models, were generated to map the regional SSM in the 0–10 cm soil layer across the study area. Multiple features, including the geographical location, elevation, vegetation coverage, soil texture, seasonal patterns and satellite-retrieved SSM product from Fengyun-3C (FY-3C), were selected as the input variables for the proposed ML models. In situ SSM measurements from the Chinese Automatic Soil Moisture Observation Stations (CASMOS) were used as the reference dataset. The error metrics, including the coefficient of correlation (R), mean relative error (MRE), unbiased RMSE (ubRMSE) and mean absolute error (MAE), between the measured SSM values and those estimated using the different models were calculated. Among those ML models, the RfR model showed the best performance in the training (R = 0.981, MRE = 7.3%, ubRMSE = 0.021 cm 3 /cm 3 , and MAE = 0.015 cm 3 /cm 3 ) and testing (R = 0.789, MRE = 22.2%, ubRMSE = 0.065 cm 3 /cm 3 , and MAE = 0.047 cm 3 /cm 3 ) processes and was applied to map the regional SSM values and measure the importance of each input feature. The results indicated that geographical location, i.e., latitude (35.84%) and longitude (16.96%), contributed the most to the SSM estimation model, followed by elevation (14.88%), vegetation coverage (9.75%), the FY-3C SSM product (8.30%), the soil texture (8.04%) and seasonal patterns (6.23%). In addition, the SSM estimations across mainland China matched the spatiotemporal patterns of historical precipitation well, which indicated the feasibility of achieving accurate and consistent land surface (0–10 cm) soil moisture monitoring results using the established RfR model with appropriately selected input features.