Regional stream temperature modeling in pristine Atlantic salmon rivers: A hybrid deterministic–Machine Learning approach

Study region: Pristine Atlantic salmon rivers located across northeastern Canada and the U.S. Study focus: To simulate water temperature in ungauged rivers, we explore the regionalization of thermal parameters within the CEQUEAU model—a deterministic, semi-distributed hydrological and water temperature model. Additionally, a global sensitivity analysis is conducted to identify the most sensitive thermal parameters within the study region. We employed the support vector regression algorithm (SVR), to map the dependence of these parameters with climatic and watershed characteristics. New hydrological insights for the region: Parameters controlling radiative and sensible heat fluxes are the most critical for CEQUEAU water temperature modeling within the study region. Key explanatory variables include low cloud coverage, high wind speed quantiles, upstream land cover areal coverage, distance to the coast, watershed orientation, and topographical features describing surface curvature and elevation. The machine learning-based regionalization approach provides a robust approach for deriving water temperature model parameters from watershed attributes, provided flow measurements are available. Using leave-one-out cross-validation, support vector regression (SVR) significantly outperformed the traditionally used multiple linear regression (MLR), achieving a mean regional RMSE of 1.89 °C.