Predicting annual peak daily streamflow in natural basins using quantile regression forests

Flood risk is characterized by flood inundation areas influenced by hydroclimatic extremes such as peak streamflow events. Predicting peak streamflow discharge in ungauged basins upstream of dams or reservoirs is critical for forecasting inflows, aiding operational management, and mitigating downstream flood risk. We developed a Quantile Regression Forest (QRF) model to predict annual peak daily streamflow in ungauged basins, incorporating uncertainty quantification and variable influence analysis. The model integrates continental-scale data from PRISM, GAGES-II, NWIS Streamflow, and NLCD for the CONUS. Through hyperparameter tuning and recursive feature elimination (RFE), we optimized the QRF model to achieve an adjusted R² of 0.768 with low SMAPE scores (20.512% overall, median 9.444). Results reveal peak precipitation as the dominant driver of flood magnitude ($>$50% importance) in streamflow prediction, alongside significant contributions from other explanatory variables. The model effectively captures hydrological relationships and achieves realistic calibration to observed conditions. This approach provides actionable insights for water resources management and flood risk assessment.