An initial assessment of the possibilities of fish catch prediction using Gaussian processes

Abstract

The fishing and aquaculture industry is one of the largest industries of Norway. Enhanced knowledge of the distribution of fish in the ocean is important for an economical and sustainable fishing industry. This study investigates the possibilities of using Gaussian processes for regression within fish catch prediction. A dataset that combines catch reports from the Norwegian shipping company Havfisk ASA with a multitude of ocean-related data is created, and analysed in this thesis. Stochastic variational inference for Gaussian process models is used for the regression, as the method allows the use of Gaussian processes for regression on large datasets. The aim of this study is to assess the suitability of the dataset for fish catch predictions, in addition to evaluating the predictions from the Gaussian process model. Different investigations were performed within; time dependency, clustering analysis, data transformation and feature selection. The investigations indicates that Gaussian processes for regression do reveal a structure in the dataset and that the collected dataset is suitable for fish catch prediction.