Machine-learning aiding sustainable Indian Ocean tuna purse seine fishery

Abstract

Among the various challenges facing tropical tuna purse seine fleet are the need to reduce fuel consumption and carbon footprint, as well as minimising bycatch of vulnerable species. Tools designed for forecasting optimum tuna fishing grounds can contribute to adapting to changes in fish distribution due to climate change, by identifying the location of new suitable fishing grounds, and thus reducing the search time. While information about the high probability to find vulnerable species could result in a bycatch reduction. The present study aims at contributing to a more sustainable and cleaner fishing, i.e. catching the same amount of target tuna with less fuel consumption/emissions and lower bycatch. To achieve this, tropical tuna catches as target species, and silky shark accidental catches as bycatch species have been modelled by machine learning models in the Indian Ocean using as inputs historical catch data of these fleets and environmental data. The resulting models show an accuracy of 0.718 and 0.728 for the SKJ and YFT, being the absences (TPR = 0.996 for SKJ and 0.993 for YFT, respectively) better predicted than the high or low catches. In the case of the BET, which is not the main target species of this fleet, the accuracy is lower than that of the previous species. Regarding the silky shark, the presence/absence model provides an accuracy of 0.842. Even though the model's performance has room for improvement, the present work lays the foundations of a process for forecasting fishing grounds avoiding vulnerable species, by only using as input data forecast environmental data provided in near real time by earth observation programs. In the future these models can be improved as more input data and knowledge about the main environmental conditions influencing these species becomes available.