Dataset of runoff variation and its driving mechanisms in headwater catchments of the Yellow River water conservation area (2000-2020)

The contribution of driving factors is obtained by constructing a data-driven framework: using XGBoost (eXtreme Gradient Boosting) machine learning regression model with Optuna automatic hyperparameter optimization for runoff simulation, coupled with SHAP (Shapley Additive Explanations) interpretability analysis method based on cooperative game theory, to quantify the marginal contributions of climate and underlying surface factors to runoff output.

The uncertainty and error range of the model are evaluated through five fold stratified cross validation, and the 95% confidence interval is directly calculated based on the statistical distribution of the validation samples to characterize the robustness of feature contributions. The dataset contains two table files stored in a universal CSV (comma separated values) format, which can be read using conventional spreadsheet software such as Excel, WPS, or programming languages.

The first file (Figure 3. csv) records the spatiotemporal evolution data of runoff patterns in a small watershed, containing a total of 31 records, with row labels indicating specific hydrological station names (such as Huanghe River, Tangke, etc.); The column variables record the average flow rate, multi-year average values of low and high flow rates (measured in m ³/s), slope of the trend, and Boolean identification through significance testing (1 represents significant, 0 represents insignificant) for each watershed.

The second file (Figure 4a - left. csv) records the contribution statistics of 19 climate and underlying surface characteristic factors under average flow conditions, including 19 data records with row labels of characteristic factor names (such as precipitation, potential evapotranspiration, forest proportion, etc.); The column variables cover the feature contribution strength index (mean-abs_SHAP) based on the SHAP absolute value over the years, as well as the upper and lower bounds of the 5% quantile (abs_SHAP_5quantile) and 95% quantile (abs_SHAP_95quantile) that reflect model uncertainty.