Uncertainty-informed topology-aware transfer learning for robust stream salinity prediction

Study Region: Upper Red River Basin, Texas-Oklahoma, U.S. Study Focus: This paper presents a topology-aware transfer learning (TL) approach using machine learning models that leverage stream network connectivity and publicly available stream salinity data with varying temporal frequencies within a basin to generate robust daily stream salinity predictions. To ensure robust predictions from models with stable parameters and performance across potential uncertainties, the lower upper bound estimation (LUBE) method was used to quantify the uncertainty in the generated stream salinity data. A rich original record (∼2 years) of continuous sub-daily in-situ salinity measurements was used to verify estimated prediction intervals. New hydrological insights for the region: Stream salinity data availability varies across time and space within various tributaries of the Upper Red River Basin, USA, complicating water supply management. To address this challenge, this study generated robust stream salinity data for all monitoring sites within the basin. The topology-aware TL framework helped improve prediction accuracy by 6% lower RMSE for best fitted model and an average of 0.25 higher NSE for all 600 models fitted for a site. This improved accuracy of TL models is more pronounced at downstream sites compared to local models, while enhancing robustness and decreasing the uncertainty of predictions for all sites. Additionally, TL enabled fitting an accurate model with NSE higher than 0.8 for a site with only 26 pairs of observation, which was infeasible via local models. Comparison with the recent sub-daily in-situ data indicated the framework’s spatiotemporal generalizability, and confirmed the reliability of both the PIs and the generated synthetic salinity data.