Spoilt for choice: An evaluation of machine learning for wildlife classification in tropical forest settings

Read full text Summarise with VIRAT
Authors: Paulina Kukofka, Barbara Fruth, Anna-Marie Broska, Roland Cleva, Alain Pauni Mussa, Nadia Balduccio, Mattia Bessone
Cited by: 0
References: 2

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

In times of species decline and biodiversity loss, large-scale, multi-species surveys are essential for effective conservation planning. Camera traps have become indispensable for wildlife monitoring, providing detailed insights into species abundance and distribution. However, manual processing of large camera trap datasets remains time-consuming and resource-intensive. Although machine learning algorithms offer a scalable solution, their performance on data collected in densely vegetated tropical forests remains to be validated. Here, we evaluated machine learning algorithms for species classification, using 5,886 videos and 1,369 images from 55 camera traps deployed in the rainforest near Salonga National Park, Democratic Republic of the Congo, between September 2023 and March 2024. We compared performance metrics across species and algorithms and assessed how factors such as model confidence, distance, and body size influenced accuracies. Overall classification accuracies were lower than officially reported but comparable across region-specific algorithms. Specialist algorithms outperformed a generalist algorithm for classifications on the species level. Performance varied strongly between species, particularly for rare and small-bodied animals. Labelling confidence-based thresholding improved accuracies but resulted in substantial loss of data. Algorithmic confidence score was the strongest predictor of classification accuracy in region-specific algorithms, while distance negatively affected detection and correct classifications. Our findings underscore the importance of critical validation when applying machine learning algorithms, especially in challenging environments such as tropical forests. Machine learning performance is highly location- and species-dependent, with errors potentially propagating to derived ecological metrics. We outline practical steps to integrate manual validation into machine learning workflows to help practitioners avoid pitfalls and ensure robust and rigorous wildlife monitoring that strengthens conservation efforts.

Fields of Study & Subjects

Wildlife Ecology and Conservation Advanced Neural Network Applications Environmental DNA in Biodiversity Studies

Publication details

Publication date
25 March 2026
Language
EN
Document type
Preprint