Spatiotemporal Evolution and Drivers of Soil Erosion in the Yarlung Tsangpo River Basin: A RUSLE-Based 30-Year Analysis and Future Scenario Projections
Description
This dataset includes gridded soil erosion data for the Yarlung Tsangpo River Basin from 1990 to 2019, derived using the Revised Universal Soil Loss Equation (RUSLE) model. It also contains projected rainfall erosivity (R-factor) rasters for four future periods (2021–2040, 2041–2060, 2061–2080, and 2081–2100) under three CMIP6 emission scenarios (SSP126, SSP245, and SSP585), based on the bias-corrected outputs of 25 GCMs using the quantile delta mapping (QDM) method. In addition, the dataset includes C-factor rasters simulated using the XGBoost machine learning algorithm, trained with historical WordClim climate variables for each future scenario and time slice. Finally, projected average soil erosion rasters for each period and scenario across the entire basin are also provided.
Files
Steps to reproduce
This dataset includes future rainfall erosivity (R factor) and vegetation cover management (C factor) inputs for soil erosion modeling in the Yarlung Tsangpo River Basin (YTRB), based on CMIP6 climate projections. For R factor, we used daily precipitation data from 25 CMIP6 GCMs provided by the NEX-GDDP-CMIP6 dataset. Bias correction was performed using the Quantile Delta Mapping (QDM) method with a 1990–2014 baseline. Corrections were applied monthly using the ibicus Python package. A rainfall threshold of 0.1 mm was used to define rainy days. To validate the QDM performance, we compared bias-corrected GCM outputs with observed precipitation from 2011–2014 using R². For C factor, five CMIP6 models with high equilibrium climate sensitivity (ECS > 3.0) were selected. We trained an XGBoost model using historical C factor values and 19 bioclimatic variables from WorldClim. To reduce overfitting, we applied Recursive Feature Elimination (RFE) to select key input variables and used a Genetic Algorithm (GA) to optimize model hyperparameters. The model was trained with 5-fold cross-validation. The trained model was then used to predict future C factors under three SSP scenarios (SSP126, SSP245, SSP585) across four future periods: 2021–2040, 2041–2060, 2061–2080, and 2081–2100.
Institutions
- Yunnan University