An active learning model to classify animal species in Hong Kong

TL;DR

This work addresses the poor cross-site generalisability of deep learning models for camera-trap species classification. It proposes a Hong Kong–specific classifier built by fine-tuning a ResNet50 pretrained on ImageNet within a MegaDetector-based preprocessing pipeline and further accelerates labeling through an active-learning loop. The approach achieves strong generalization to an independent HK dataset (94.1% accuracy, 93.7% precision, 81.9% recall, 87.4% F1), though performance varies by class due to sample size imbalances. The method reduces labeling effort while delivering reliable species identification, enabling scalable biodiversity monitoring in urbanized regions like Hong Kong.

Abstract

Camera traps are used by ecologists globally as an efficient and non-invasive method to monitor animals. While it is time-consuming to manually label the collected images, recent advances in deep learning and computer vision has made it possible to automating this process [1]. A major obstacle to this is the generalisability of these models when applying these images to independently collected data from other parts of the world [2]. Here, we use a deep active learning workflow [3], and train a model that is applicable to camera trap images collected in Hong Kong.

Figures (2)

• Figure 1: Accuracy, precision and recall of the model as labels are added to training images.
• Figure 2: Confusion matrix generated by applying the active learning model onto an independent dataset using images collected from other sites in Hong Kong.