ИСТИНА

In the northwestern Pacific Ocean, there are two ecotypes of killer whales: residents (fish-eaters) and transients (marine-mammal eaters). All prior efforts to distinguish between them morphologically were limited to descriptive variations. In the present study, both ecotypes’ images were classified by machine learning (neural network). A total of 1084 images have been selected from photo-identification studies over 15 years off Eastern Kamchatka and the Commander Islands. Image processing (rotating, aligning, and cropping) was performed manually before their loading into the model. The model was based on transfer learning of the MobileNetV2 in Edge Impulse environment. The Edge Impulse platform was used for training and testing machine learning algorithms. During testing, the model had an accuracy of 91.12% where residents’ classification had 89.6 % accuracy with 6.6% error and 3.8% uncertain while Transients’ classification had 92.6 % accuracy with 3.7% error and 3.7% uncertain. Those results indicate that the neural network can learn to differentiate ecotypes even on this small dataset and this can be applied to differentiate killer whale ecotypes even with a small-scale dataset, which emphasizes their long-term reproductive isolation. Furthermore, machine learning can be used as another tool for assessing population variation and recognition of ecological and evolutionary processes in living cetaceans.