Seeing the species through the trees: Using Random Forest classification trees to identify species-specific whistle types.

Abstract

Acoustic identification of delphinid species is hampered by high variability in whistle characteristics. It is possible that not every whistle contains species-specific information and that there are “species-specific” whistle types. Random forest analysis was used to examine whistles of 8 species recorded in the eastern tropical Pacific Ocean (Delphinus species, Globicephala macrorhynchus, Pseudorca crassidens, Stenella attenuata, S. coeruleoalba, S. longirostris, Steno bredanensis, Tursiops truncatus). Fifty-one variables were measured from 2176 whistles. The number of trees within a random forest that “voted” for the predicted species was used as a measure of the strength of classification. A whistle was considered strongly classified if the predicted species received at least 40% of the votes, even if the prediction was incorrect. The percent of whistles that were strongly classified ranged from 33% (S. longirostris) to 73% (G. macrorhynchus). Overall, 62% of strong whistles were correctly classified, ranging from 22% (S. longirostris) to 87% (P. crassidens). Overall correct classification for weakly classified whistles was 33% and ranged from 17% (Delphinus spp. and G. macrorhynchus) to 55% (S. bredanensis). Results suggest that while there may be “species-specific” whistle types, the distinctiveness of these whistles and the frequency with which they are produced varies among species.