Automatic Species Identification in Camera-Trap Images

Abstract

At Jasper Ridge Biological Preserve, heat-change triggered cameras have been deployed for wildlife detection for over 10 years. These cameras generated more than 191,000 photos that were subsequently labeled by volunteers. To relieve the intensive human labor required in the labeling process, we developed an automatic photo classification system using convolutional neural networks. In this process, we experimented with two data preprocessing techniques and four CNN-based pipelines, with our best single model achieving an accuracy of 93.05% and an average F-1 score of 84.39% on the test set. Furthermore, we experimented with different ensemble strategies and model combinations. The best-performing ensemble achieves an accuracy of 93.73% and an average F-1 score of 85.73% on the test set. The system is also equipped two additional features. Firstly, for users that aim for higher prediction accuracy, the system has the option to only settle predictions with confidence higher than a user-specified threshold, and separate out unsettled photos for further human verification. Secondly, the system provides suggested labels and model visualization for unsettled photos to ease the human verification process. With this automatic photo classification system, we hope to largely reduce the human labor required to label camera trap images at Jasper Ridge Biological Preserve.