Simulation and analysis of evapotranspiration from desert grasslands based on a random forest regression model.

Evapotranspiration (ET) is a critical component of the water and energy cycles in desert grassland ecosystems. However, modeling ET in arid grasslands faces significant challenges due to data scarcity, high spatiotemporal heterogeneity, and complex interactions among climatic drivers. To address these challenges, this study developed a Random Forest Regression (RF-R) model integrated with high-resolution PML-V2 ET data and CRU meteorological datasets (2001-2020) to simulate ET in China's desert grasslands. The RF-R model achieved superior performance, with R² values of 0.953 (training) and 0.931 (testing), RMSE of 3.421 and 4.182 mm/month, and an average prediction bias of 11.815%. The comparative analysis between BPNN and SVR models confirms the robustness of RF-R estimates. Key climate factors were identified through multi-scale importance assessments: precipitation and wet-day frequency were the primary drivers, followed by cloud cover and diurnal temperature range. This study provides a reliable framework for ET simulation in data-scarce arid regions and supports targeted water management strategies for desert grassland restoration.