The advent of big data and machine learning offers great promise for addressing conservation and management questions in the oceans. Yet, few applications of machine learning exist to mitigate the overexploitation of marine resources. Tropical tuna purse seine fisheries (TTPSF) are distributed worldwide and account for two thirds of the global tuna catch. In these fisheries, the use of Drifting Fish Aggregating Devices (DFADs)— n-made floating objects massively deployed by fishers to increase their tuna catches—results in the incidental catch of non-target species, termed bycatch. We explored the possibility of applying machine learning on echosounder buoys attached to DFADs, representing an unprecedented source of big data, for identifying high bycatch risk at DFADs. We trained random forests algorithms to differentiate between high and low bycatch occurrence based on matched echosounder and onboard observer data for the same DFADs (representing sample sizes of 838 and 2144 in the Atlantic and the Indian Ocean, respectively). Algorithms showed a better performance in the Atlantic Ocean (accuracy of 0.66 versus 0.58 in the Indian Ocean) and were best at detecting the “high bycatch” occurrence class. These results unravel the potential of machine learning applied to fishers' buoys data for bycatch reduction and improved selectivity in one of the largest fisheries worldwide.