A time-invariant bias correction strategy for improving CLM5.0 evapotranspiration simulation by random forest method for mainland China

Abstract

Reliable and accurate estimation of large-scale evapotranspiration (ET) is fundamental for research in Earth System Science, however, large-scale ET estimation remains a challenge. Bias correction of the land surface model (LSM) simulated ET is a popular approach for providing large-scale, long-term estimations. Previous correction methods often ignore stochastic errors, making them unsuitable for complex conditions. Therefore, this study evaluated the performance of the Community Land Model 5.0 (CLM 5.0) in simulating ET based on the Global Land Evaporation Amsterdam Model (GLEAM) product and in situ observations across mainland China, analyzed the main factors influencing model performance, developed a time-variant bias correction strategy using the random forest (RF) method. Results show that precipitation is crucial in influencing the performance and uncertainty of CLM 5.0 simulations since it determines whether ET is limited by energy or water. The CLM 5.0 performed worse but with lower uncertainty in water-limited regions due to soil water stress, while it performed better but with higher uncertainty in energy-limited regions. Future research should focus on refining the depth limits of dry surface layer (DSL) parameterization to improve model performance in water-limited regions. Furthermore, the CLM 5.0 overestimated ET in the Northern China (NC) region but underestimated ET in the other seven regions. The overestimation is attributed to the model's overestimation of leaf area index and the underestimation of GLEAM data in farmland. After bias correction, the national average correlation coefficients (R) increased by 0.102, root mean square errors (RMSE) decreased by 0.178 mm/d, absolute mean bias (BIAS) values decreased by 0.006 mm/d, Kling-Gupta efficiency (KGE) values increased by 0.154, and uncertainty coefficients with 95 % confidence level (U_95%) decreased by 0.195 mm/d. The performance of bias corrected simulations is significantly improved. The trends of both overestimation and underestimation of ET simulations by CLM 5.0 at a regional scale have also been alleviated. The time-invariant bias correction strategy proposed in this study demonstrates more reliable performance compared to the previous studies that applied monthly scaling factors. This advancement is essential for estimating more reliable large-scale ET under complex conditions.