Examining the Potential for Remotely Sensed Snow Information to Predict Seasonal Water Supply across Dimensions of Aspect, Elevation, and Slope
Journal Article
Overview
abstract
Abstract; Seasonal water supply is largely dependent on interannual snow accumulation in the western United States, where snowmelt comprises the majority of spring runoff. Within the context of seasonal water supply prediction, we investigate whether remotely sensed snow observations may be able to supplement or replace limited in situ monitoring across the West. Across 85 snow-dominated basins, we apply a data-driven regression framework with robust feature selection to explore the skill of remotely sensed snow timing variables, e.g., the day of snow disappearance (DSD), in the prediction of seasonal streamflow for WY1985–2021. We construct terrain-informed subdomains using basin aspect, elevation, and slope information to identify where snowpack and streamflow are most strongly linked, comparing lumped and semidistributed forecasts. Models trained on remotely sensed predictors exhibit “very good” median predictive skill (< ±5% bias) for biweekly forecast dates from 1 March to 1 June. When in situ observations of snow are available, the median skill of “remotely sensed” models is comparable with “in situ” models, producing lower bias in approximately 40% of monitored basins. Across western U.S. snow regimes, we find snow timing information is a consistent (i.e., commonly selected) predictor of seasonal streamflow when aggregated over a basin’s highest elevations and is a highly weighted predictor when aggregated over a basin’s north-facing slopes. By quantifying the importance of these subdomains and their skill in forecasting seasonal water supply across the western United States, this analysis offers insights into how remotely sensed snow-cover data can most reliably predict seasonal streamflow in data-scarce regions.; ; Significance Statement; For 85 snow-dominated basins in the western United States, and for a period of record from WY1985–2021, we investigate the ability of remotely sensed snow information to predict seasonal water supply. To identify the most important predictors of snowmelt-driven streamflow, we apply an ensemble multiple regression framework with robust feature selection designed to increase model parsimony. We construct terrain-informed subdomains grouped by aspect, elevation, and slope and train predictive models using remotely sensed snow timing information, e.g., the day of snow disappearance, aggregated over these spatially discrete subdomains. This framework exhibits “very good” predictive skill (<±5% bias) and provides a greater understanding of how basin topography interacts with various climatic factors to inform streamflow forecasting.;
