Work

An investigation of lightweight AI models to classify African ungulate species from tracks

Year: 2025

Type: article

Abstract: In this study, we present a method to classify African ungulate species using lightweight artificial intelligence (AI) models that are particularly appropriate for use in typically lower-resourced areas where wildlife is found. Tracks are cryptic ground evidence that record the history of movement of animals and can be used for monitoring, conservation, anti-poaching and the mitigation of human-wildlife conflict. The wildlife species investigated include black rhinoceros (Diceros bicornis) , blue wildebeest (Connochaetes taurinus) , common eland (Taurotragus oryx) , giraffe (Giraffa camelopardalis) , white rhinoceros (Ceratotherium simum) , and zebra (Equus quagga) based on left and right hind track classification. We used two strategies to classify tracks: firstly a bounding box method for object detection, and then secondly with the additional placement of strategic landmarks within the bounding box. We utilised transfer learning techniques on SSD MobileNet v2 FPNLite and YOLO v8 nano with default architectures, and investigated the effect of additional detection layer and anchor modifications for the bounding box method. YOLO v8 nano pose estimation was used to train the placement of strategic landmarks. The addition of strategic landmark points within the bounding box improved the classification based on mean average precision at intersection over union by varying the threshold from 0.50 to 0.95. For strategic landmark points within the bounding box, the mean average precision for black and white rhinoceros was 0.930, and for other ungulates species was 0.844. For the bounding box method, the mean average precision for all species was 0.823. The best performing bounding box method obviated the need for expert and time-consuming human input in model development inherent in a morphometric approach alone. In addition, we investigated the impact of track quality on classification accuracy. We ranked tracks into three classes according to their characteristics, from the outline to the inside of the tracks. The track quality ranking highlighted a decline in model performance as the track quality declined. The field testing analysis attained 98% correct species classification and 95% hind classification within the correct species classification. This approach demonstrated cost-effectiveness and deployment capability in the wild. • Automatic African ungulate species classification from tracks through the application of artificial intelligence. • A non-invasive and cost-effective approach crucial to monitoring biodiversity indicators. • Develop computationally inexpensive lightweight artificial intelligence models to run in local and small computing devices to permit near real-time use in remote wildlife areas. • A multi-disciplinary collaboration of biologists, trackers, engineers and the military, to develop improved technology for monitoring and protecting wildlife.

Source: Ecological Informatics

DOI: 10.1016/j.ecoinf.2025.103393

Authors: Tinao Petso, Rodrigo S. Jamisola, Sky Alibhai, Molaletsa Namoshe, Wazha Mmereki, Zoe Jewell

Institutions: Botswana International University of Science and Technology, Duke University

Language: English

Topic: Advanced Neural Network Applications

Subfield: Computer Vision and Pattern Recognition

Field: Computer Science

Domain: Physical Sciences

Open Access status: gold

Estimated APC paid: $2,770

Funders: Metropolitan Museum of Art, Botswana International University of Science and Technology, Army Research Office