Leveraging Machine Learning to Enhance Water Quality Predictions in Small Agricultural Streams

Effective water quality management at the catchment scale requires robust tools to predict pollutant loads under diverse environmental conditions. This study introduces an innovative methodology employing machine learning models to estimate phosphorus (P) concentrations and loads in small catchments contributing to the Northeast Arm of Lake Champlain (NALC). The region is characterized by agricultural land use and dynamic hydrological conditions. Current challenges include the lack of monitoring data for small direct drainage streams and the high uncertainty in existing P load estimates.

To address these gaps, we propose a dual modeling framework using Random Forest (RF) and Long Short-Term Memory (LSTM) models. Both models will be trained and validated using project-specific monitoring data, alongside extensive datasets from the USGS and regional monitoring programs. RF models, known for their interpretability and efficiency, will quantify predictor variable importance and generate insights into the key drivers of P loading. Complementarily, LSTM models, capable of capturing complex temporal dynamics, will provide high-resolution predictions of daily P loads and concentrations.

Our methodology highlights innovative monitoring strategies, including the deployment of stream gauging and water sampling stations at representative sites, capturing flow rates and concentrations of total phosphorus (TP), total dissolved phosphorus (TDP), and total suspended solids (TSS). These observations will be integrated into the machine learning framework, allowing a targeted validation of the model results. Preliminary analyses indicate disproportionately high P loading in small agricultural watersheds, underscoring the need for targeted interventions informed by reliable model predictions.

Expected outcomes of this study include the identification of source areas and processes driving P and sediment transport, as well as validated machine learning tools capable of estimating loads in ungauged basins. These models are designed to accommodate future scenarios of land use and climate change, providing resource managers with actionable insights to design effective mitigation strategies. By integrating high-resolution empirical data with state-of-the-art machine learning techniques, this work advances the understanding and management of nutrient dynamics at the catchment scale.