Time series forecasting of chlorophyll-a concentrations in the Chesapeake Bay

Declining water quality poses serious environmental and public health risks, with chlorophyll-a serving as a key biological indicator of harmful algal blooms. This study evaluates the use of a Long Short-Term Memory (LSTM) neural network to forecast chlorophyll-a concentrations in the Chesapeake Bay, a critical estuarine ecosystem supporting over 17 million people. Weekly satellite-derived chlorophyll-a measurements from 1997 to 2020 were collected for three geographic regions of the bay. The LSTM model was benchmarked against traditional statistical models, including ARIMA and TBATS, and trained to capture complex seasonal patterns in the time series data. Across all three regions, the LSTM consistently outperformed the other models, achieving lower root mean squared error (RMSE) values of 0.121, 0.155, and 0.199 mg/m³. These results demonstrate the LSTM model's superior ability to learn spatiotemporal dynamics and accurately predict future chlorophyll-a levels. This work highlights the potential of deep learning approaches to improve water quality forecasting, inform timely policy decisions, and support sustainable management of aquatic ecosystems.