Forecasting closures on shellfish farms using machine learning

Abstract

Biotoxins and harmful algal blooms (HABs) are damaging to aquaculture operations. Occurrences lead to disrupted operations, fish kills, and significant risks to human health. The conditions leading to blooms are driven by known, but complex processes. Heuristics exist about the drivers but the nonlinearity and opaqueness of relationships make it difficult to resolve using traditional rule-based mathematical models. An alternative approach leverages machine learning to uncover the conditions that lead to the closure of farms. This paper presents a comprehensive framework that combines semi-automated machine learning with ensemble classification approaches to predict site closures. Performance is evaluated on 7 years of site closure data from a shellfish farm in Southwest Portugal, together with publicly available environmental data. The model reports an accuracy of 83% across a challenging forecasting task. The proposed framework provides a pragmatic, scalable, site-specific decision tool to help aquaculture stakeholders mitigate the impacts of HABs.