The relative importance of model type and input features for water supply forecasting in snow-dominated basins of the southwestern US

• Published in: Journal of hydrology. Regional studies Vol. 60; p. 102548
• Main Authors: Pernat, Madeline R., Kasprzyk, Joseph, Zagona, Edith, Walker, Sydney D., Livneh, Ben
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.08.2025; Elsevier
• Subjects: Climate change; Feature selection; Machine learning; Nested cross-validation; Seasonal forecasting; Southwestern United States; Water supply forecasting; Feature selection; ERT; SVR; FFS; MLR; Southwestern United States; PDO; SWE; WIND; NSE; WY; AMO; ENSO; Seasonal forecasting; Nested cross-validation; WSF; SOI; Water supply forecasting; PCA; PA; Climate change; AMJJ; CV; RF; SNOTEL; RRMSE; TMP; Machine learning; SEFS; SPFH; NAO; PCR
• ISSN: 2214-5818; 2214-5818
• DOI: 10.1016/j.ejrh.2025.102548

This study focuses on five watersheds in the southwestern United States, where April–July (AMJJ) water supply forecasts (WSFs) inform water management. Climate change has altered long-relied-upon relationships between April 1st snow water equivalent (SWE) and AMJJ water supply, threatening the skill of traditional forecasting approaches. This work evaluates how the interaction between model type (e.g., multiple linear regression, random forest) and feature selection influences AMJJ WSF skill. Five machine learning model types are applied in each basin. A new wrapper-based feature selection method identifies the Best Feature Set—selected from a broad pool of station-based, meteorological, and climatological features—for each basin–model type combination. Results show that the most important features vary by both basin and model type, and that model types perform similarly when each is trained on its respective Best Feature Set. April 1st SWE is most important in highly snow-dominated basins, while April 1st precipitation accumulation becomes more important in less snow-dominated systems. Station-based features from multiple lag times are consistently selected, suggesting that earlier observations provide additional predictive value. Among meteorological and climatological features, specific humidity and the Atlantic Multidecadal Oscillation are frequently selected across basins and model types, indicating broad predictive utility. Overall, results suggest that feature selection has a greater influence on forecast skill than model type choice. [Display omitted] •Feature importance varies by both basin and model type for AMJJ WSFs.•Feature selection influences forecast skill more than model type choice.•SWE is important in highly snow-dominated basins; PA in less snow-dominated basins.•Meteorological and climate features improve skill across basins and model types.