Deep learning-based insect classification with YOLOv۱۱-CLS: Applications in agricultural pest management and ecological conservation

Automated insect species classification is critical for sustainable pest management, biodiversity monitoring, and precision agriculture. This study presents a systematic evaluation of five variants of the YOLOv۱۱-cls deep learning models (nano to extra-large) for fine-grained classification of ۵۴ insect species, including agricultural pests, pollinators, and biological control agents. A curated dataset comprising over ۳,۰۰۰ images was employed, with rigorous preprocessing and stratified data splitting into training, validation, and test sets. The models were trained under standardized conditions using transfer learning, advanced data augmentation, and dynamic learning rate scheduling to ensure optimal convergence and generalization. Results demonstrate a clear trade-Off between model size and performance. Larger architectures, such as YOLOv۱۱x-cls, achieved superior accuracy metrics, including ۱.۵ ms/image, making them suitable for resource-constrained environments such as edge computing and mobile applications. Notably, all models attained near-perfect top-۵ accuracy (>۰.۹۹), indicating robust multi-class discrimination capabilities. The integration of these models into a web-based application enables real-time species identification with visual feedback, offering scalable solutions for field researchers, extension agents, and smallholder farmers. Given Iran's ecological diversity and the lack of entomological expertise in many rural areas, this system provides a valuable tool for improving decision-making in crop protection and conservation of beneficial species. Lightweight models are particularly well-suited for deployment in remote regions with limited computational infrastructure. Despite promising results, limitations include reliance on relatively small class-specific image counts (~۶۰ per class on average) and the use of static images that may not fully capture real-world variability. Future work will focus on synthetic data generation, domain adaptation, and integration of multimodal inputs such as hyperspectral imaging to enhance resilience and cross-species generalization. These findings underscore the importance of tailoring model selection to application-specific requirements, balancing accuracy and efficiency in Al-driven workflows for sustainable agriculture and ecological monitoring.

کلیدواژه ها:

Artificial intelligence ، Insect classification ، Precision agriculture ، YOLOv۱۱ ، Biodiversity conservation