Automated Identification of Vespid Wasps Using Transformer-Based Deep Learning Models: A Biodiversity Informatics Approach from Sri Lanka

Abstract Background: Effective biodiversity monitoring and conservation depend on accurate species identification, but traditional morphological taxonomy of insects remains labor-intensive and prone to misclassification, especially among visually similar taxa. Vespidae, a family that includes ecologically important wasps, exemplifies this challenge due to subtle differences between species. In South Asia, where biodiversity hotspots face increasing human pressures, fast and reliable insect identification tools are crucial for conservation planning and ecological assessment. Deep learning methods, particularly transformer-based architectures, show promise in automating species recognition from digital images. However, their use in fine-grained insect classification, especially for underrepresented taxa in biodiversity databases, remains limited. This study evaluated the performance of transfer learning with pre-trained deep learning models for species-level identification of vespid wasps from field-collected specimens in conservation contexts. Results: A curated dataset of 300 images representing 14 vespid species collected from field surveys was assembled. Three advanced deep learning architectures were systematically tested: ResNet-50, Vision Transformer, and Swin Transformer. The model based on the Vision Transformer demonstrated the best performance, achieving 78% accuracy and an F1-score of 0.76, surpassing both ResNet-50 and Swin Transformer. This represents the first successful application of transformer-based architectures for species-level wasp identification within the South Asian biodiversity context. Performance analysis showed that transformer models excel at capturing subtle morphological features necessary for distinguishing closely related species, highlighting their effectiveness for fine-grained taxonomic classification tasks. Conclusions: Transformer-based deep learning models provide a scalable and efficient approach to automate insect identification, with significant implications for conservation biology. These tools can expedite biodiversity surveys, enable real-time monitoring in ecological fieldwork, facilitate rapid environmental impact assessments, and enhance data collection in resource-constrained settings. By decreasing dependence on specialized taxonomic expertise, such automated systems make biodiversity monitoring more accessible and strengthen evidence-based conservation decisions in threatened ecosystems.