Saltwater intrusion (SWI), exacerbated by climate change, poses a significant threat to estuarine ecosystems, impacting drinking water supplies, agriculture, aquaculture, biodiversity, and habitat stability. Improving our ability to understand and predict the SWI is crucial to support decision-making for estuarine and coastal resource management. We developed machine learning (ML) models that estimate daily SWI length (Ls) in the Chesapeake Bay main stem and its eight major tributaries, supporting both historical reconstructions and future projections. The ML models were trained on two decades (2001-2020) of simulations from a validated, high-resolution three-dimensional hydrodynamic model. The ML model effectively reproduced variability of SWI, providing a simplified and computationally efficient alternative to traditional numerical modeling. Model skills were consistently high, with correlation coefficients ranging from 0.88 to 0.95 and root-mean-square errors (RMSE) between 1.53 and 5.03 km. In addition, the ML model demonstrated predictive capability for 7-day and 14-day forecasts using the preceding 90 days of discharge data. Scenario experiments simulating increased or reduced river discharge confirmed the model's utility for management applications. A key advantage of the ML approach is its ability to reconstruct historical or future SWI under limited forcing data conditions, providing valuable insights into long-term hydrological changes and the effects of climate variability on SWI in the Chesapeake Bay.
Shen et al. (Mon,) studied this question.