Accounting for snowpack and time-varying lags in statistical models of stream temperature

Water temperature plays a primary role in driving ecological processes in streams due to its direct impact on biogeochemical cycles and the physiological processes of stream fauna, such as growth, development, and the timing of life history events. Streams influenced by snowpack melt are generally cooler in the summer and demonstrate less sensitivity to climate variability in what is commonly referred to as “climate buffering”. Despite the substantial influence of snowpack on stream temperature and expected changes in snowpack accumulation and melt timing with climate change, methods for representing snowpack in statistical models for stream temperature have not been well explored. In this investigation, we quantified the extent of stream temperature buffering in free-flowing streams across a geographically diverse region in the Pacific Northwest USA. We demonstrated that statistical models of daily mean stream temperature can be improved by explicitly accounting for temporal variability in a small number of climate covariates believed to be mechanistically related to stream temperature. Our novel statistical approach included as predictors combinations and interactions between the following variables: (1) air temperature, (2) lagged air temperature (where the lag duration varied according to its relationship with flow on a given day at that site), (3) flow, (4) snowpack in the upstream catchment, and (5) day of year. We found that sites with substantial snow influence were associated with increased air temperature buffering during the warm season and longer air temperature lags (>30 days during spring high flows and ∼ 10 days during late summer low flows) compared to sites where precipitation predominantly fell as rain (<6 days year-round). By accounting for snowpack and temporal variation in lagged heat transfer processes, our models were able to accurately predict seasonal patterns and interannual variability in stream temperature in validation data from years not used in model fits using publicly available data sources (average RMPSE ∼ 0.80).