Applying Deep Learning to Predict Adverse Environmental Conditions in Fish Aquaculture Pens

• Main Authors: Khan, Abdul Baseer, Korus, Jennie, Sclodnick, Tyler, Whidden, Christopher
• Format: Conference Proceeding
• Language: English
• Subjects: Aquaculture; Biological system modeling; Deep learning; Dissolved Oxy-gen; Environmental Monitoring; Fish; Forecasting; Machine Learning; Predictive models; Real-time systems; Schedules; Sustainable development; Sustainable Fish-eries; Thin film transistors; Transformers
• ISSN: 2996-1882
• DOI: 10.1109/OCEANS55160.2024.10754551

Aquaculture is critical for meeting the increasing global seafood demand as wild fish stocks decline. Effective management of aquaculture environments is vital for sustainability and operational efficiency. This study applies advanced deep learning models, specifically Temporal Fusion Transformers (TFT), to predict dissolved oxygen (DO) levels in fish pens. Accurate DO forecasting is essential for optimizing feeding schedules, enhancing fish health, and reducing environmental impact. Using extensive datasets from aquaculture sites in Canada and Scotland, our models were trained and validated to deliver reliable 24-hour DO predictions. Our phased approach, starting with single-pen data and scaling to multiple pens across various farms and regions, achieved mean absolute error (MAE) values below 0.7 for 24-hour forecasts. The findings demonstrate the potential of deep learning models in assisting real-time decision-making and mitigation strategies in aquaculture, fostering sustainable and efficient practices. Future work could incorporate additional environmental variables or farm inputs and extend model applications to diverse aquaculture systems.