Modelling interdependencies between fish species in the North Atlantic : a Bayesian machine learning approach to predictive biomass and recruitment models

• Main Author: Larsson, Aron
• Other Authors: Helsingin yliopisto, Bio- ja ympäristötieteellinen tiedekunta, University of Helsinki, Faculty of Biological and Environmental Sciences, Helsingfors universitet, Bio- och miljövetenskapliga fakulteten
• Format: Master Thesis
• Language: English
• Published: Helsingin yliopisto 2021
• Subjects: Bayes; Bayesian networks; machine learning; R; bnlearn; fisheries; recruitment; biomass; prediction; forecasting; Ympäristömuutoksen ja globaalin kestävyyden maisteriohjelma; Master's Programme in Environmental Change and Global Sustainability; Magisterprogrammet i miljöförändringar och global hållbarhet; Ympäristömuutos; Environmental Change; Miljöförändring; atlantic cod; North Atlantic
• Online Access: http://hdl.handle.net/10138/330566

The science of fish stock assessment is one that is very resource and labor intensive, with stock assessment models historically being based on data that causes a model to overestimate the strength of a population, sometimes with drastic consequences. The need of cost-effective assessment models and approaches increases, which is why I looked into using Bayesian modeling and networks as an approach not often used in fisheries science. I wanted to determine if it could be used to predict both recruitment and spawning stock biomass of four fish species in the north Atlantic, cod, haddock, pollock and capelin, based on no other evidence other than the recruitment or biomass data of the other species and if these results could be used to lower the uncertanties of fish stock models. I used data available on the RAM legacy database to produce four different models with the statistical software R, based on four different Bayes algorithms found in the R-package bnlearn, two based on continuous data and two based on discrete data. What I found was that there is much potential in the Bayesian approach to stock prediction and forecasting, as our prediction error percentage ranged between 1 and 40 percent. The best predictions were made when the species used as evidence had a high correlation coefficient with the target species, which was the case with cod and haddock biomass, which had a unusually high correlation of 0.96. As such, this approach could be used to make preliminary models of interactions between a high amount of species in a specific area, where there is data abundantly available and these models could be used to lower the uncertanties of the stock assessments. However, more research into the applicability for this approach to other species and areas needs to be conducted.