The promise of machine learning in the Risso’s dolphin Grampus griseus photo-identification

Photo-identification (photo-ID) studies are strategic to fill the gap of knowledge of data deficient species such as Risso’s dolphin. Unfortunately, the photo-ID process is very time consuming and strongly depends on the user-ability. Some photo-ID algorithms are available, which can, automatically or semi-automatically, find the closest match between the dolphin in the query and a catalogue of previously sighted dolphins. However the limitation of these algorithms is that in any case they will return a prevision of the dolphin identity, in other words these can not identify the individuals never sighted before, i.e. unknown dolphins. Hence the automation of the photo-ID process through the use of innovative algorithms is still needed. In this paper the opportunity of employing machine learning strategies for the automated photo-ID of Risso’s dolphin is investigated. In particular the performances of RUSBoost algorithm result to be very good in identifying the unknown dolphins, even if in general these depend on the available data for training the model. Experimental results highlight the great potential of machine learning in the automation of photo-ID process, as well as focus on the need of collecting more and more data in order to perform a more effective data analysis.