Deep Learning-based Model for Wildlife Species Classification

Abstract

In this paper, we describe that motion-activated cameras are nowadays widely used in Ecological parks as well as wildlife sanctuaries. These cameras capture the images whenever any motion is observed by the sensors. They are also capable of capturing infrared images therefore providing millions of images, which was earlier a very expensive as well practically infeasible task. However, extracting useful information from these images about any wildlife species is still a time-consuming and labour-intensive task. We demonstrate that deep learning models can be used for extracting this information near human-level accuracy. We trained VGG16 ConvNet architecture using transfer learning on a dataset of 33,511 images of 19 species from the Ladakh region of India and achieved training and testing accuracy of 89.12% overall. The pipeline developed here has wide application in wildlife monitoring across different national parks.