Machine Learning-Based Prediction of Streamflow Volume and Peak Discharge Using Snow, Temperature, and Soil Moisture Data in Logan River, Utah

Accurate prediction of annual streamflow volume and peak discharge is essential for effective water resource management, particularly in snowmelt-dominated watersheds where hydrologic responses are highly sensitive to climate variability. This study applies three regression-based modeling approaches—Multivariate Linear Regression (MLR), k-Nearest Neighbors (KNN), and Random Forest (RF)—to forecast annual streamflow volume and peak discharge using a combination of snow, temperature, and soil moisture data from the Logan River watershed in northern Utah. The input features included maximum snow water equivalent (SWE), average maximum temperature over the three weeks following peak SWE (as a proxy for melt energy), and antecedent soil moisture measured at 2\", 8\", and 20\" depths. The dataset spanned water years 1980 to 2023, with soil moisture data available for 2006 onward. Feature engineering was performed to align predictors with hydrologically relevant timing, such as melt onset and pre-saturation conditions. MLR models using only SWE and temperature achieved R² values of 0.749 for annual streamflow volume and 0.772 for peak discharge, while adding soil moisture improved performance to 0.847 and 0.899, respectively. KNN models using SWE and temperature achieved R² values of 0.804 for volume and 0.776 for peakflow, making it the second-best method when soil data were unavailable. The RF model showed strong performance for peakflow (test R² = 0.746) but was less accurate for volume (R² = 0.646). Across all models, SWE consistently emerged as the most influential predictor, while temperature and soil moisture improved model accuracy by capturing melt energy and runoff potential. These results demonstrate that incorporating multiple hydrometeorological variables improves prediction of streamflow behavior, and that machine learning methods such as KNN and RF offer valuable alternatives to traditional regression, particularly when data availability varies. This framework provides a robust approach to streamflow forecasting under changing climatic and hydrologic conditions.