Surface Soil Moisture Retrieval of China Using Multi-Source Data and Ensemble Learning

Abstract

Large-scale surface soil moisture (SSM) distribution is very necessary for agricultural drought monitoring, water resource management, and climate change research. However, the current large-scale SSM products have relatively coarse spatial resolution, which limits their application. In this study, we estimate the 1 km daily SSM in China based on ensemble learning using a multi-source data set including in situ soil moisture measurements from 2980 meteorological stations, MODIS Surface Reflectance products, SMAP (Soil Moisture Active Passive) soil moisture products, ERA5-Land dataset, SRTM DEM and soil texture. Among them, in situ measurements are used as independent variables, and other data are used as dependent variables. In order to improve the spatio-temporal completeness of SSM, the missing value in SMAP soil moisture products were reconstructed using the Discrete Cosine Transformation-penalized Partial Least Square (DCT-PLS) method to provide spatially complete background field information for soil moisture retrieval. The results show that the reconstructed soil moisture value has high quality, and the DCT-PLS method can fully utilize the three-dimensional spatiotemporal information to fill the data gaps. Subsequently, the performance of four ensemble learning models of random forest (RF), extremely randomized trees (ERT), extreme gradient boosting (XGBoost), and light gradient boosting machine (LightGBM) for soil moisture retrieval was evaluated. The LightGBM outperformed the other three machine learning models, with a correlation coefficient (R2) of 0.88, a bias of 0.0004 m³/m³, and an unbiased root mean square error (ubRMSE) of 0.0366 m³/m³. The high correlation between the in situ soil moisture and the predicted values at each meteorological station further indicate that LightGBM can well capture the temporal variation of soil moisture. Finally, the model was used to map the 1 km daily SSM in China on the first day of each month from May to October 2018. This study can provide some reference and help for future long-term daily 1 km surface soil moisture mapping in China.

Keywords

• soil moisture
• multi-source data
• ensemble learning
• China