Improved soil moisture mapping using an integrated cyclic modeling and bias correction approach

Abstract

Soil moisture (SM) is crucial for climate change, crop growth estimation, and environmental hazard monitoring. Existing SM products often have low spatial resolution, limiting their use in local-scale studies. While various machine learning (ML) methods have been applied to downscale SM, few studies have explored multiple cyclic modeling or improved downscaling accuracy by recycling qualified stations. In this study, we performed quality control and bias correction on data from the International Soil Moisture Network (ISMN) stations. We obtained qualified sites by cyclic modeling using an extreme gradient boosting (XGBoost) regression model. The predicted bias from cyclic modeling was combined with dynamic environmental variables to correct errors at unqualified sites. Finally, a surface (0–5 cm) soil moisture product with a temporal and spatial resolution of 500 m/day was produced: (1) the XGBoost model described the relationship between SM and environmental variables well, achieving a correlation coefficient (R) of 0.98 and a root mean square error (RMSE) of 0.007 m³/m³ (2) The generated 500 m SM data was comparable to the Soil Moisture Active Passive Level 4 (SMAP-L4) SM data, with 83.2 % of the 1996 points having R > 0.6. The downscaling accuracy is improved by robust cyclic modeling and bias correction techniques, with R, RMSE, and mean absolute error (MAE) improved by 6.5 %, 9.3 %, and 9.6 %, respectively, over single-shot modeling. The estimated results of the surface layer (0–5 cm) soil moisture at 500 m/day can supplement the regional soil moisture database and provide ideas for downscaling soil moisture research.