Recent Advances on the Use of Supervised Learning Algorithms for Detecting Tuna Aggregations Under FADS from Echosounder Buoys Data

Assessing the accuracy of biomass estimates obtained through echosounder buoys and improving the
current algorithms used for estimating the associated biomass is a key step towards the derivation of
fisheries-independent abundance indices for tropical tuna. Recent results obtained through supervised
learning algorithms on M3I buoys, one of the main buoy models deployed by the French tuna purseseiners,
demonstrate a good accuracy for assessing the presence and absence of tuna under FADs,
regardless of the ocean. However, these algorithms (and buoy model) are less accurate in determining the
size of tuna aggregations. In this paper we investigated possible ways of improving the classification of
tuna aggregation sizes by accounting for the species composition constituting the aggregation. Also, we
inspected how environmental variables (sea-surface temperature and chlorophyll-a) can affect the
accuracy of the biomass estimates. Our results demonstrate that accounting for the species composition of
tuna aggregation, sea-surface temperature and chlorophyll-a does not improve significantly the accuracy of
biomass estimates with this buoy model.