Authors: Brian Maughan and Jim Nelson – Brigham Young University
Title: Hourly Stream Temperature Prediction with LSTMs: From CAMELS Basins Toward Southern Alaska
Presentation Type: Poster Presentation
Abstract: Stream temperature is a critical indicator of aquatic ecosystem health and a key driver of cold-water habitat suitability. Current models operate at a daily time step, missing the sub-daily thermal dynamics that govern biological stress events and diurnal warming cycles. This study develops a long short-term memory (LSTM) deep learning model, implemented within the NeuralHydrology framework, to predict stream temperature at hourly resolution across minimally impaired reference basins (CAMELS) in the conterminous United States (CONUS), with the broader goal of extending hourly stream temperature prediction to data-sparse subarctic regions such as Southern Alaska. Training data were drawn from the CAMELS dataset, filtered to USGS stream gages with co-located 15-minute stream temperature and discharge records. Hourly ERA5-Land meteorological forcings, alongside observed streamflow and static CAMELS basin attributes, were used as model inputs. Model performance is evaluated on held-out CONUS basins to assess generalization across diverse hydroclimatic regimes, establishing a foundation for future transfer to Southern Alaskan watersheds where co-located streamflow and temperature observations remain limited.