Evaluation Of A Multi-Model Prediction Ensemble For Soil Erosion In A Typical Watershed On The Loess Plateau

How to effectively quantify and simulate soil erosion is one of the important research elements in erosion control,Despite the existence of many different soil erosion models,there are still few studies that use integrated techniques to simulate soil erosion.To investigate whether integration techniques can effectively reduce the uncertainty of soil erosion simulated by different models,this study uses Bayesian model averaging(BMA),Bates-Granger(BG)integration,and Granger-Ramanathan(GR)integration techniques in the Xinshui River basin on the Loess Plateau to investigate the widely used Morgan-Morgan-Finney(MMF),Revised Universal Soil Loss Equation(RUSLE),and China Soil Loss Equation(CSLE).The integrated simulations were rate-determined by watershed export sediment observations,and the models were simulated and evaluated based on Nash coefficient(NSE),correlation coefficient(R),root mean square error(RMSE),and other evaluation metrics.The main findings of the study are as follows:1.the soil erosion simulated by the three single models has some differences,among which MMF is closer to the real soil erosion condition of the watershed,and its evaluation indexes NSE,R and RMSE are better than RUSLE and CSLE models in most cases.2.The integrated simulation can increase the prediction accuracy of single-model simulation to some extent,and the NSE,R,and RMSE increase by 47.2%,23.6%,and 46.4%before and after integration,respectively.3.By comparing the confidence intervals and dispersion coefficients of the three integration methods,we find that Bayesian average and BG integration have higher confidence and lower dispersion coefficients compared to GR integration,and are more reliable in terms of uncertainty.Due to the wide range of sources involved in model uncertainty,the application of soil erosion model integration only provides partial quantification of model uncertainty,and it will be particularly important to use integration methods to reduce model prediction uncertainty to a greater extent in the future.