Modeling Stream Water Temperatures for A Montane Catchment Using Observations, Statistical Models and Machine Learning

The water temperature of streams in montane catchments is a key harbinger of ecosystem health and water resource quality for nearby and downstream communities, a dependency that is ever increasing and sensitive to change, in the western United States and worldwide. In recent decades, representative snowfall-dominated, montane catchments such as the Sagehen Experimental Forest (hereafter, “Sagehen”), located in the eastern Sierra Nevada mountains of California, have been studied to better understand how disturbances ranging from climate-induced events, i.e., drought, wildfire, and extreme precipitation events; to human-caused events, i.e., forestry experimentation, affect stream flow and stream water temperature (SWT). Sagehen, like many catchments in the mountain West, experiences cold, wet winters and warm, dry summers, with both the quantity and timing of snow and rain being vitally important for sustaining spring and summer streamflow and buffering SWT for ecosystem resiliency. Alarmingly, climate projections for Sagehen indicate an earlier snowmelt season and more rain-on-snow events, both of which are likely to result in unknown consequences. Additionally, rising global temperatures may exacerbate the risk of hard-to-predict disturbances (i.e., wildfires and insect infestations) and resulting impacts on hydrologic systems. Stream hydrologic response, including SWT, to such events remains poorly understood due to limitations such as lack of field data and/or lack of years-long records.

To address this knowledge gap, we leverage a 12-year dataset of streamflow and SWT observations collected across Sagehen to first calibrate and then compare statistical and machine learning models for SWT. Currently, performance metrics using TempEst-NEXT, a CONUS-scale, statistical SWT forecasting model show a RMSE of 4.09°C, R² of 0.88, NSE of 0.43 and percent bias (PBIAS) of 48% for mean daily SWT. Performance metrics for the machine-learning neural network model using daily SWT, air temperature and snow input show a strong validation period RMSE of 0.71°C, R² of 0.98, NSE of 0.98, and PBIAS of 0.11%. Using this unique dataset, which encompasses both dry and wet periods, droughts and extreme precipitation events, as well as forest treatments, we consider the following objectives: 1) examine how climatic factors have influenced SWT response during the period of record and how response may change in the future using climate scenarios, and 2) identify what, if any, physical patterns can be discerned from observations, modeling results, and model comparisons. Preliminary analysis of daily SWT in Sagehen shows that summer 2020 had the highest daily mean SWT for the 12-year record, followed by summer 2021, and 2022. In terms of SWT variability, preliminary analysis has identified a possible relation between SWT variability and slope-face, where SWT is most buffered on the main stem, followed by the north-facing, then south-facing tributaries. Pending model analysis and cross-comparison is expected to illuminate differences in model prediction  of SWT for the Sagehen basin for both the near-term and the future. Broadly, this research is expected to provide new insights on the evolution of hydrology in a montane catchment as it responds to climate variability and disturbance events.