Multi-angle, cross-domain fusion strategy enhances automated insect identification and hierarchical categorization: a case study on assassin bugs (Hemiptera: Reduviidae).

Abstract

Automated insect identification systems hold significant value for biodiversity monitoring, pest management, citizen science initiatives and systematic studies, particularly in an era of declining expertise in insect taxonomy. However, current deep learning approaches often rely on standardized specimen photos from limited-angles and simplified backgrounds, limiting their generalization and effectiveness in diverse practical scenarios. Here, we address this limitation using assassin bugs (Hemiptera: Reduviidae) as a model system-a highly diverse group with complex morphological variation. We developed a comprehensive, high-quality dataset of 11 915 expert-validated images from 92 species across 48 genera and six subfamilies, integrating three image types: standard, turntable-captured and ecological. Using ConvNeXt-B architecture, we systematically evaluated the performance of classification and hierarchical categorization to higher taxonomic ranks across different training and testing scenarios. Multi-angle fusion increased species recognition accuracy by 5.72% and F1-score by 0.061 and increased the hierarchical categorization of unseen species to correct genera by 13.53% and the F1-score by 0.129. Incorporating ecological images further enhanced model performance by 13% for both tasks. Grad-CAM visualization revealed that multi-angle, cross-domain fusion guides the model to focus on taxonomically diagnostic traits, yielding hierarchical feature representations aligned with classical taxonomy across lab and wild environments. Our results demonstrate that integrating diverse viewing angles and ecological contexts substantially improves model cross-domain adaptation, providing a practical framework for developing reliable automated insect identification tools for diverse practical scenarios. Furthermore, our model's ability to accurately categorize unseen taxa into higher taxonomic ranks holds significant potential for future applications in systematics, including facilitate specimen sorting and generating testable phylogenetic hypotheses.