Background: Global insect populations are declining at an alarming rate, with estimates suggesting a 40% reduction in biomass over the past three decades. Traditional morphological identification methods are costly, time-intensive, and require scarce taxonomic expertise, creating a critical monitoring bottleneck.
Objective: This study develops and evaluates a deep learning–based insect species recognition framework capable of automated, scalable, and accurate biodiversity monitoring from digital imagery.
Methods: We assembled a curated dataset of 125,000 annotated insect images spanning 3,830 species across five major orders. Five state-of-the-art deep learning architectures—ResNet-50, EfficientNet-B4, Vision Transformer (ViT), YOLOv8, and ConvNeXt-B—were trained, fine-tuned, and benchmarked under standardized conditions. Performance was assessed using classification accuracy, precision, recall, and detection efficiency metrics.
Results: The Vision Transformer achieved the highest classification accuracy (95.8%), precision (95.3%), and recall (95.6%), outperforming all CNN-based baselines. EfficientNet-B4 offered the best accuracy-to-computational-cost ratio, while YOLOv8 demonstrated superior real-time detection throughput at 47 frames per second. Lepidoptera achieved the highest per-order recognition rate (96.1%), whereas Orthoptera posed the greatest challenge (90.2%) owing to cryptic coloration.
Conclusion: AI-driven frameworks substantially enhance the scalability and precision of insect biodiversity monitoring. Integration with citizen science platforms and IoT sensor networks is recommended for continental-scale deployment. The models and annotated dataset are openly available to support future research.