Detection of Wildlife Species in the Peruvian Amazon using Transfer Learning

Abstract

Wildlife holds an important role within the Amazon biome. However, wildlife identification and documentation methods in the Amazonian wilderness pose considerable challenges for fauna biology and ecology professionals. This complexity arises from the demand for specialized expertise and the substantial investment of time required. This challenge is compounded by the remarkable resemblance between various animal species. In this study, we delve into the feasibility of diverse iterations of the YOLO (You Only Look Once) algorithm in order to detect wildlife species in the Peruvian Amazon. Our assessment covers a spectrum of YOLO versions, including YOLOv5x6, YOLOv5l6, YOLOv7-W6, YOLOv7-E6, YOLOv8I, and YOLOv8x. To empower our models, we embarked on a training journey using a dataset comprising 653 images thoughtfully collected from reputable sources in ecology and tourism marketing. This dataset encompasses six species: Ara ararauna, Ara chloropterus, Ara macao, Opisthocomus hoazin, Pteronura brasiliensis, and Saimiri sciureus. Our efforts show the efficiency of the YOLOv5l6 model, which stands out prominently in all metrics evaluated. This model achieves a Precision rate of 86.1%; Recall of 84.7%, F1-Score measuring 85.39%, and mean Average Precision (mAP) of 88.1%. Noteworthy is the fact that this model also boasts the swiftest training time among its counterparts, with a total 30.71 minutes. These findings offer promising prospects for refining our understanding of Amazonian wildlife species and establishing proactive measures to safeguard those that face potential vulnerability or endangerment. The YOLO algorithm's capabilities underscore the confluence of technology and ecological conservation, providing optimism for the preservation of the Amazon's intricate biodiversity.