Smart camera traps and computer vision improve detections of small fauna

AbstractLimited data on species' distributions are common for small animals, impeding conservation and management. Small animals, especially ectothermic taxa, are often difficult to detect, and therefore require increased time and resources to survey effectively. The rise of conservation technology has enabled researchers to monitor animals in a range of ecosystems and for longer periods than traditional methods (e.g., live trapping), increasing the quality of data and the cost‐effectiveness of wildlife monitoring practices. We used DeakinCams, custom‐built smart camera traps, to address three aims: (1) To survey small animals, including ectotherms, and evaluate the performance of a customized computer vision object detector trained on the SAWIT dataset for automating object classification; (2) At the same field sites and using commercially available camera traps, we evaluated how well MegaDetector—a freely available object detection model—detected images containing animals; and (3) we evaluated the complementarity of these two different approaches to wildlife monitoring. We collected 85,870 videos from the DeakinCams and 50,888 images from the commercial cameras. For object detection with DeakinCams data, SAWIT yielded 98% Precision but 47% recall, and for species classification, SAWIT performance varied by taxa, with 0% Precision and Recall for birds and 26% Precision and 14% Recall for spiders. For object detections with camera trap images, MegaDetector returned 99% Precision and 98% Recall. We found that only the DeakinCams detected nocturnal ectotherms and invertebrates. Making use of more diverse datasets for training models as well as advances in machine learning will likely improve the performance of models like YOLO in novel environments. Our results support the need for continued cross‐disciplinary collaboration to ensure that large environmental datasets are available to train and test existing and emerging machine learning algorithms.