Image-Based identification of Insect Species and Subspecies Using Machine Techniques.

Abstract

Accurate identification of hoverflies (family Syrphidae) is critical to ecological research, given their dual role as pollinators and biological control agents. However, traditional identification techniques—based on morphological traits—can be time-consuming, require expert knowledge, and are often inaccessible to non-specialists. This study explores the development of an image-based hoverfly identification system that leverages deep learning to automate classification tasks. We collected and labeled images of hoverflies, representing 25 hoverfly species, using smartphones, and categorized them into three views: head, dorsal, and full body. These images were systematically organized and labeled with a five-part filename convention, then used to train a convolutional neural network modeled after a VGG-like architecture. The system achieved strong classification performance, with subfamily accuracy reaching 89.7%, gender 80.2%, dorsal view 69.8%, and species 67.1%. The findings suggest that deep learning models, when paired with carefully constructed datasets, can serve as powerful tools for insect classification. The proposed system offers a promising solution for supporting biodiversity assessments, agricultural diagnostics, and educational outreach—democratizing species identification for both professionals and citizen scientists.