Improving the Accuracy of Animal Species Classification in Camera Trap Images Using Transfer Learning

Abstract

Understanding biodiversity, monitoring endangered species, and estimating the possible effect of climate change on particular regions all rely on animal species identification. Closed-circuit television (CCTV) cameras, which can collect huge volumes of video data, are an excellent environmental monitoring tool. However, manually evaluating these massive datasets is time-consuming, difficult, and expensive, emphasizing the need for automated ecological analysis.Deep learning models have transformed computer vision, handling problems such as object and species detection. Their cutting-edge performance qualifies them for this application. The purpose of this work was to create and test machine learning models for distinguishing diverse animal species using camera trap images. On VGG19, GoogLeNet (InceptionV3), ResNet50, and DenseNet121, we used transfer learning. The best multi-classification accuracy was attained by GoogLeNet (87%), followed by ResNet50 (83%), DenseNet (81%), and VGG19 (53%). This evidence suggests that transfer learning outperforms training models from scratch for this task.