Classifying False Alarms in Camera Trap Images using Convolutional Neural Networks

Wildlife trapping cameras often capture false alarms when triggered by blowing vegetation or cloud shadows moving across the ground. Identifying these false alarms and distinguishing them from true capture events (images of actual animals, human, vehicle) requires a substantial amount of personnel time. Here we explore how convolutional neural networks can be used to develop an automated computer screening model for filtering out the false alarms. The models screening threshold can be varied to suit the requirements of the given camera network. For cameras used for real-time public education and outreach, a low screening threshold can be used. Based on using a screening threshold of 0.5 on a specific Tensorflow model, false alarms were classified with an average accuracy of 88.83 ± 4.29% and true capture events with 91.83±2.85% on a dataset of 23,930 images. A high screening threshold should be used for research purposes. By choosing a threshold of 0.97, only 0.37% of true capture events are misclassified and about 50% of false alarms are correctly classified, saving between 5.5 and 11 eight-hour workdays of personnel time. As part of this study, we also explore some of the ramifications of deploying the existing model to classify images from new camera networks.