Deep Learning Algorithm Forecasts Shellfish Toxicity at Site Scales in Coastal Maine

The aquaculture industry is the fastest growing food sector in the world, resulting in both wild harvest and farmed shellfish industries becoming increasingly important to coastal communities and their local economies. However, every year harmful algal blooms (HABs) lead to often long-term shutdowns of harvesting sites. Due to the effects of climate change as well as the unprecedented growth of the aquaculture industry, the intensity, frequency, and resulting impacts of HABs are exacerbated. Contaminated shellfish, and their resultant cases of paralytic shellfish poisoning, is one major impact in Maine. Currently, there are no site-specific forecasts of paralytic shellfish poisoning outbreaks in Maine. Deployed models operate on larger time and geospatial scales. Machine learning and other powerful technological tools are being increasingly utilized across many industries such as government organizations, marketing teams, and in science as well to increase the resolution and predictive power of models. Neural networks, a machine learning tool, is an extremely powerful predictive tool for complex dynamic systems such as ecosystems. We used the Keras neural network application programming interface along with Maine Department of Marine Resources’ toxicity data to forecast paralytic shellfish poisoning outbreaks in the Gulf of Maine. The predictions were made for each harvesting site and used four classification levels, the highest indicating a closure, and intermediate levels indicating warning of potential closure at a particular location. We also conducted a variety of tests with various metrics and configurations to determine the predictive power of the neural network. The model was predicting with at least 95% accuracy with zero false predictions at the closure level; the variability in predictive power was largely in the lower two levels of toxicity. Compared with simpler methods, the neural networks was a superior predictive tool, suggesting the utility of this machine learning approach for high resolution ecological forecasts.