DATASETS FOR MACHINE LEARNING BASED CONTINENTAL SCALE MODELING FOR PREDICTING RUNOFF-COEFFICIENTS

Accurate prediction of the event-based runoff coefficient (ERC) is fundamental to understanding watershed response, yet achieving reliable predictions across diverse climatic and geophysical regions remains a significant challenge. This study presents a robust, data-driven framework for predicting ERC at the continental scale across the Contiguous United States (CONUS). Utilizing a large-sample hydrology approach, we harmonized a comprehensive dataset of 479,872 rainfall-runoff events from USGS-gauged watersheds. The model integrates sub-daily atmospheric forcing from the NLDAS-2 dataset with static watershed attributes, including topography, land cover, and soil characteristics. To address spatial heterogeneity, we implemented a regime-based modeling approach that utilizes K-means clustering to categorize watersheds into distinct hydrological regimes before training. Predictive performance was evaluated using advanced machine learning algorithms, specifically eXtreme Gradient Boosting (XGBoost) and Support Vector Regression (SVR). Our results demonstrate that the regime-based ML framework significantly outperforms traditional global modeling approaches, particularly in regions with high seasonal variability. Sensitivity analysis reveals that antecedent moisture conditions and land-use patterns are the primary drivers of ERC variability at the continental scale. To support community research and reproducibility, the processed dataset and model scripts are made publicly available via HydroShare. This work provides a scalable tool for predicting watershed response in ungauged basins and offers new insights into the process controls governing runoff generation across the CONUS.