Collective Anomaly Detection in Fisheries

Abstract

One of the tasks of the Norwegian Directorate of Fisheries is surveillance of ocean fisheries in Norwegian waters. To discourage illegal, unreported and unregulated fisheries, they collect various types of data about fishing activities, in particular data about each catch operation.
However, catch data from fishing activities are by nature unpredictable, and many factors may be causes of variation. This makes it hard to identify single catch operations that are anomalous and perhaps incorrectly reported. In this paper we show how we can use the concept of collective anomalies by looking at collections of catch operation reports and check how they deviate from the expected. We do this by running a machine learning model to predict total catches of trawlers' catch operations, compute the prediction errors from the model, and see how the prediction error distribution of a vessel deviates from the whole set of catch reports. The experiments are promising and we are able to identify deviating vessels in a consistent manner, but the outcomes still need to be evaluated by domain experts.