Spatiotemporal Soil Erosion Dataset for the Yarlung Tsangpo River Basin (1990–2100)
This dataset includes historical erosion data (1990–2019), future soil erosion projections under SSP126, SSP245, and SSP585 scenarios (2021–2100), and predicted R and C factors for each period.
Future R factors We incorporated 25 Global Climate Models (GCMs) from CMIP6 for calculating the future R factors, selected via the NASA Earth Exchange Global Daily Downscaled Projections (NEX-GDDP-CMIP6) project (Table S3). The selection was based on the completeness of their time series and their alignment with the selected scenarios. Rainfall projections were corrected using quantile delta mapping (QDM) (Cannon et al., 2015) to address systematic biases in intensity distributions while preserving the projected trends in mean rainfall and extremes—critical for soil erosion analysis (Eekhout and de Vente, 2019). Bias correction was conducted using a 25-year baseline (1990–2014), with adjustments made monthly to correct for seasonal biases. The corrected bias functions were then applied to adjust the years (2020–2100) of daily rainfall data using the "ibicus" package, an open-source Python tool for bias adjustment and climate model evaluation. A minimum daily rainfall threshold of 0.1 mm was used to define rainy days, following established studies (Bulovic, 2024; Eekhout and de Vente, 2019; Switanek et al., 2017). Additionally, the study employed QDM to correct biases in historical GCM simulations, ensuring the applicability of the QDM method for rainfall bias correction in the YTRB. A baseline period of 1990–2010 was selected to establish the bias correction function, which was subsequently applied to adjust GCM simulations for 2011–2014. To evaluate the effectiveness of this calibration, we compared the annual mean precipitation from bias-corrected GCMs during 2011–2014 with observed precipitation data at the pixel level (Figs. S2, S3), using R² as the evaluation metric. The results showed a significant increase in R² after bias correction, confirming the effectiveness of the QDM approach.
Future C factors To ensure the accuracy of the C factor predictions, we selected five CMIP6 climate models (table S4) with high spatial resolution compared to other CMIP6 climate models. Of the five selected climate models, CanESM5, IPSL-CM6-LR, and MIROC-ES2L have high equilibrium climate sensitivity (ECS) values. The ECS is the expected long-term warming after doubling of atmospheric CO2 concentrations. It is one of the most important indicators for understanding the impact of future warming (Rao et al., 2023). Therefore, we selected these five climate models with ECS values >3.0 to capture the full range of potential climate-induced changes affecting soil erosion. After selecting the climate models, we constructed an XGBoost model using historical C factor data and bioclimatic variables from the WorldClim data portal. WorldClim provides global gridded datasets with a 1 km² spatial resolution, including 19 bioclimatic variables derived from monthly temperature and precipitation data, reflecting annual trends, seasonality, and extreme environmental conditions (Hijmans et al., 2005). However, strong collinearity among the 19 bioclimatic variables and an excessive number of input features may increase model complexity and reduce XGBoost's predictive accuracy. To optimize performance, we employed Recursive Feature Elimination (RFE), an iterative method for selecting the most relevant features while preserving prediction accuracy. (Kornyo et al., 2023; Xiong et al., 2024). In each iteration, the current subset of features was used to train an XGBoost model, and feature importance was evaluated to remove the least significant variable, gradually refining the feature set. Using 80% of the data for training and 20% for testing, we employed 5-fold cross-validation to determine the feature subset that maximized the average R², ensuring optimal model performance. Additionally, a Genetic Algorithm (GA) was applied in each iteration to optimize the hyperparameters of the XGBoost model, which is crucial for enhancing both the efficiency and robustness of the model (Zhong and Liu, 2024; Zou et al., 2024). Finally, based on the variable selection results from RFE, the bioclimatic variables of future climate models were input into the trained XGBoost model to obtain the average C factor for the five selected climate models across four future periods (2020–2040, 2040–2060, 2060–2080, and 2080–2100).
RUSLE model In this study, the mean annual soil loss was initially estimated using the RUSLE model, which enables us to estimate the spatial pattern of soil erosion (Renard et al.,1991). In areas where data are scarce, we consider RUSLE to be an effective runoff dependent soil erosion model because it requires only limited data for the study area (Haile et al.,2012).
