Identifying significant factors and optimal sites for commercial salmon farming in northern Norway. An integrated GIS and machine learning approach using random forest.

Abstract

This study presents a data-driven modelling approach to identify important factors influencing the growth- and mortality rate for farmed salmon in northern Norway. Furthermore, a model is trained to determine the best fish farming sites and identify optimal areas with the best geographical conditions. Aquaculture site production and location data from 323 salmon farming sites (all licensed aquaculture sites) in northern Norway were obtained from the Directory of Fisheries. Two dependent variables, growth- and mortality rate, were calculated based on the monthly increase in biomass and mortality. These variables were combined with state-of-the-art environmental- and exploratory socio-economic data obtained from the institute of marine research (IMR), the Norwegian Meteorological Institute, Delft University of Technology, Norwegian Coastal Administration, and Statistics Norway. Using random forest regression and recursive feature elimination, a data-driven ensemble approach identified significant variables. Prediction of optimal sites for salmon farming in northern Norway was done with a species distribution modelling approach using random forest classification. The important factors affecting salmon growth were specific feeding rate, temperature, and total biomass. The important factors influencing salmon mortality were temperature and total biomass. The predicted optimal areas were inside Vefsnfjorden, Ranfjorden, Sørfjorden and Glomfjorden, small areas near the coast and around the small islands stretching from Gladstad to Narvik. Areas near the coast of Lofoten, Værøy, Røst, Vesterålen, Sortland and Senja. Further north, some dispersed regions were predicted as optimal outside Tromsø and Sørøya. Also large areas around Varangerhalvøya, Olderdalen/Kåfjorden, Lille Altafjorden and near the shore on both sides of Stjernøysundet. The results clearly show that space is a scares resource and that there is an urge to evaluate the regulations and legislations concerning aquaculture in Norway. Especially the minimum distances between the fairways and aquaculture locations. The incorporation of machine learning approaches in GIS-based MCE analysis is suggested to help planners and decision-makers make informed and sustainable decisions about sea-area use.