FORECASTING STREAM COMEBACKS: AI MODELS OF POST-RESTORATION WATER QUALITY IN THE NORTHEASTERN UNITED STATES

Stream restoration projects such as dam removals and culvert replacements are accelerating across the northeastern United States to improve aquatic habitat, enhance climate resilience, and restore water quality. Yet, practitioners still lack tools to reliably forecast the outcomes of these interventions—especially in predicting changes in sediment and particulate dynamics reflected by stream turbidity. This study develops and evaluates machine learning (ML) models to forecast post-restoration turbidity recovery trajectories across a regional cohort of restored stream sites. Using publicly available data from USGS sensors, the EPA Water Quality Portal, and North Atlantic Aquatic Connectivity Collaborative (NAACC) inventories, we train two model types: a Random Forest regressor and a Long Short-Term Memory (LSTM) neural network. These models predict daily turbidity over 6 to 24 months following restoration activities. Results show that both model types significantly outperform seasonal climatology baselines, with the Random Forest achieving lower RMSE and the LSTM better capturing temporal dynamics such as storm-driven turbidity spikes and post-disturbance recovery. Feature importance and uncertainty analyses highlight the roles of precipitation, streamflow, and restoration attributes (e.g., dam removal vs. culvert replacement) in shaping water-quality improvements. The findings demonstrate that AI-based forecasting can complement traditional monitoring and support data-driven decision-making in restoration planning, particularly under variable hydrologic conditions. This modeling framework provides a scalable, transferable approach for anticipating sediment stabilization and stream recovery following restoration across diverse watersheds in the Northeast.