Interpretable AI for Fisheries Data

Abstract

Sustainable use of fish resources is essential, and decision makers such as the Norwegian Directorate of Fisheries (NDF) must take proactive measures to prevent Illegal, Unreported, and Unregulated (IUU) fishing activities. With access to large volumes of open datasets, machine learning (ML) models can play a key role in automating the detection of hidden patterns indicative of such activities. One valuable dataset is the collection of catch reports, where fishermen record the details of their fishing operations. Previous research has explored the use of ML models to predict expected catch quantities. By comparing these predictions with the actual reported values, potential violations of regulations can be identified. However, to ensure trust in the model’s outputs and to gain deeper insight into the data, this paper applies interpretable Artificial Intelligence (AI) methods and visualization techniques to analyze prediction errors. We investigate feature importance and examine how the most influential features affect the model's output patterns. The results are promising, demonstrating that it is possible to provide transparency in the use of ML models for fisheries data. This approach enables domain experts to better understand, trust, and make informed use of the model's findings in the future.