Applying Deep Learning to Recognize and Classify Insects in Camera-Trap Images: A Fine-Tuned Efficient-Net B3 Classifier

Project description

Our research question is "How can AI-Based Object Detection be Applied to Accurately Recognize and Classify Insects in Images?". The central objective of this prototype is to develop, train, and evaluate a deep learning classification pipeline capable of accurately identifying insect species directly from camera trap imagery.

Specifically, the project builds a multi-class image classifier covering all 22 insect species using a fine-tuned EfficientNet-B3 architecture, implements a two-stage training strategy with head-only warm-up followed by full fine-tuning, and prioritises calibrated confidence outputs suitable for field deployment, where reliable uncertainty estimation is as important as raw accuracy.

Context

Insects and other arthropods occupy a foundational position in terrestrial ecosystems. As primary pollinators, decomposers, and prey species, they underpin the ecological services upon which both biodiversity and modern agriculture depend. Traditional approaches to insect monitoring rely on manual field surveys and a variety of trapping methods, including pan traps, pitfall traps, malaise traps, and light trap. Light traps are primarily suited to night-active insects such as moths, while day-active species are typically sampled using the other mentioned methods.

A key limitation shared by most of the approaches is that they are lethal, specimens must be collected and killed before taxonomic identification can be carried out under laboratory conditions. This makes continuous, large-scale, and non-destructive monitoring inherently difficult to achieve. The emergence of automated computer vision systems offers a compelling alternative: camera-based monitoring stations capable of capturing high-resolution imagery continuously and autonomously, combined with deep learning classifiers that can identify species from those images at scale with minimal human intervention.

Results

The final EfficientNet-B3 classifier achieved an overall accuracy of 98.22% (Table 7) on the test images that were not used in the training of the model, indicating a strong performance across the 22 types of insects we wanted to classify. The model achieved a macro average f1 score of 0.982/1 and a weighted f1 that is the same indicating that a decent performance was achieved both on bigger classes and smaller classes despite the imbalance in the training data.