Location: Northwest Watershed Research Center
Project Number: 2052-13610-015-008-S
Project Type: Non-Assistance Cooperative Agreement
Start Date: Sep 1, 2025
End Date: Aug 31, 2026
Objective:
Agricultural productivity in the West is dependent upon accurate streamflow forecasts before and during the growing season. The mountain watersheds where the majority of streamflow is generated are subject to a great deal of interannual water input variability, causing peak streamflow amounts to vary year to year between 10% to 250% of historical averages. Due to the dynamic complexity of mountain ecosystems, snowmelt timing, snowmelt rates, and the subsequent streamflow variability have major consequences for downstream water consumption. The goal of this research will be to determine the primary factors affecting streamflow by leveraging over four decades of available hydrometeorological data from the Reynolds Creek Experimental Watershed and cutting-edge machine learning algorithms. One to two peer-reviewed publications will result from this effort. These findings will then be implemented into a predictive framework coupled with an energy balance snowmelt model to improve future runoff forecasts.
Approach:
Through collaboration with the Cooperator and a postdoctoral researcher at the Cooperating Institution, we will use hourly weather inputs from an existing gridded 40-year hydroclimatic dataset to force the iSnobal energy balance snow model in the Reynolds Creek Experimental Watershed in Southwest Idaho. Then, the hourly model-estimated surface water input and hourly measured streamflow leaving the basin will be used to develop machine learning models with Python packages such as Scikit-Learn, TensorFlow, and PyTorch. The use of ARS SCINet infrastructure and computing power will be key to the success of the model development. These models will be developed for a calibration period in the dataset, then the predictive performance will be evaluated for a smaller subset of years referred to as the validation period. The models that are developed will represent the relationships between the hydrologic inputs and streamflow over each spring runoff season across the 40-year record. Many different climatic conditions are represented in the dataset, enabling the application of these models to similarly representative years in the future.