On the Use of Deep Learning Models for Automatic Animal Classification of Native Species in the Amazon

Camera trap image analysis, although critical for habitat and species conservation, is often a manual, time-consuming, and expensive task. Thus, automating this process would allow large-scale research on biodiversity hotspots of large conspicuous mammals and bird species. This paper explores the use of deep learning species-level object detection and classification models for this task, using two state-of-the-art architectures, YOLOv5 and Faster R-CNN, for two species: white-lipped peccary and collared peccary. The dataset contains 7,733 images obtained after data augmentation from the Tiputini Biodiversity Station. The models were trained in 70% of the dataset, validated in 20%, and tested in 10% of the available data. The Faster R-CNN model achieved an average mAP (Mean Average Precision) of 0.26 at a 0.5 Intersection Over Union (IoU) threshold and 0.114 at a 0.5 to 0.95 IoU threshold, which is comparable with the original results of Faster R-CNN on the MS COCO dataset. Whereas, YOLOv5 achieved an average mAP of 0.5525 at a 0.5 IoU threshold, while its average mAP at a 0.5 to 0.95 IoU threshold is 0.37997. Therefore, the YOLOv5 model was shown to be more robust, having lower losses and a higher overall mAP value than Faster-RCNN and YOLOv5 trained on the MS COCO dataset. This is one of the first steps towards developing an automated camera trap analysis tool, allowing a large-scale analysis of population and habitat trends to benefit their conservation. The results suggest that hyperparameter fine-tuning would improve our models and allow us to extend this tool to other native species.