Enhancing Sustainable Aquaculture: Applications of Artificial Intelligence in Fish Farming

The prediction of environmental conditions plays a crucial role in modern aquaculture, particularly in assessing and mitigating the risk of disease outbreaks. This thesis explores the development of deep learning models for forecasting key environmental variables in fish farms, leveraging advanced time-series prediction techniques. The study is structured into multiple phases, progressively refining both the dataset and the network architectures to enhance prediction accuracy. Initially, the project focused on understanding the dataset, its features, and their interactions. Early experiments tested different architectures and preprocessing methods, establishing a foundation for subsequent improvements. In the second phase, the introduction of Long Short-Term Memory (LSTM) networks allowed for better time-series modeling, enabling the prediction of multi-day sequences. In the third phase, the complexity of the prediction task was significantly increased by attempting to predict entire spatial matrices of environmental variables. The final phase of development returned to a more localized and precise prediction approach. Utilizing an improved dataset from CNR, the model leveraged an autoregression mechanism to predict future values iteratively. The findings of this research demonstrate that while deep learning methods—particularly LSTM-based architectures—are effective for environmental time-series forecasting, the complexity of the dataset and prediction target must be carefully considered. Future work will focus on integrating real farm mortality data and developing a classification model capable of providing early warnings for potential disease outbreaks. This advancement would represent a crucial step toward more predictive and preventive aquaculture management, helping farmers mitigate risks and improve overall fish health.