Machine learning for characterizing tropical tuna aggregations under Drifting Fish Aggregating Devices (DFADs) from commercial echosounder buoys data

Abstract

The use of echosounder buoys deployed in conjunction with Drifting Fish Aggregating Devices (DFADs) has progressively increased in the tropical tuna purse seine fishery since 2010 as a means of improving fishing efficiency. Given the broad distribution of DFADs, the acoustic data provided by echosounder buoys can provide an alternative to the conventional CPUE index for deriving trends on tropical tuna stocks. This study aims to derive reliable indices of presence of tunas (and abundance) using echosounder buoy data. A novel methodology is presented which utilizes random forest classification to translate the acoustic backscatter from the buoys into metrics of tuna presence and abundance. Training datasets were constructed by cross-referencing acoustic data with logbook and observer data which reported activities on DFADs (tuna catches, new deployments and visits of DFADs) in the Atlantic and Indian Oceans from 2013 to 2018. The analysis showed accuracies of 75 and 85 % for the recognition of the presence/absence of tuna aggregations under DFADs in the Atlantic and Indian Oceans, respectively. The acoustic data recorded at ocean-specific depths (6–45 m in the Atlantic and 30–150 m in the Indian Ocean) and periods (4 a.m.–4 p.m.) were identified by the algorithm as the most important explanatory variables for detecting the presence of tuna. The classification of size categories of tuna aggregations showed a global accuracy of nearly 50 % for both oceans. This study constitutes a milestone towards the use of echosounder buoys data for scientific purposes, including the development of promising fisheries-independent indices of abundance for tropical tunas.