A soil temperature dataset based on random forest in the Three River Source Region

Changes in soil temperature (ST) in the Three River Source Region (TRSR) significantly influence regional climate, ecology, and hydrological processes. However, existing models and reanalysis data exhibit considerable deviations in ST due to limitations in physical processes and parameterization schemes. To address this issue, we developed a new ST dataset using the Random Forest method (RFST), integrating observed ST data with relevant gridded datasets. RFST provides monthly ST data at nine layers with a spatial resolution of 0.01° × 0.01° from 1982 to 2015. Validation against two soil observation networks and six meteorological stations shows that the Nash-Sutcliffe Efficiency (NSE) of RFST exceeds 0.7 at all depths. Compared to ERA5 and CRA40, RFST corrects the cold bias, improves NSE, and reduces RMSE from 4 °C-8 °C to 1 °C-2 °C. RFST not only corrects the underestimation of ST and its warming rate but also aligns more closely with observed values for surface freezing and thawing indices as well as soil freeze-thaw periods, providing a more accurate representation of soil thermal conditions in the TRSR.