Deep learning-based computer vision techniques for automated identification of Ichneumonoidea and other Hymenoptera insects

• Ano de defesa: 2026
• Autor(a) principal: Pinheiro, João Manoel Herrera
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: aprendizado de máquina; arthropod; artrópode; biodiversidade; biodiversity; convolutional neural network; machine learning; rede neural convolucional; taxonomia; taxonomy
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/18/18162/tde-09022026-143242/

Precise taxonomic identification constitutes a prerequisite for effective biodiversity monitoring, ecological research, and biological control strategies. The order Hymenoptera, encompassing over 150,000 described species, includes the hyper-diverse superfamily Ichneumonoidea, which comprises the families Ichneumonidae and Braconidae. The taxonomic impediment, characterized by the vast number of species within these groups that remain undescribed, has direct impacts on ecology, biodiversity conservation, and practical applications such as biological control. To address this challenge, we introduce a curated, high-resolution dataset designed to advance automated taxonomic identification. The dataset comprises 1,739 specimen images focusing primarily on Ichneumonidae and Braconidae, supplemented by representative samples from nine additional hymenopteran families (Andrenidae, Apidae, Bethylidae, Chrysididae, Colletidae, Halictidae, Megachilidae, Pompilidae, and Vespidae) to ensure robust out-group differentiation. Annotation precision was maximized through a semi-automated workflow utilizing the Computer Vision Annotation Tool (CVAT) integrated with the Segment Anything Model (SAM). This methodology enabled the generation of high-fidelity segmentation masks and bounding boxes for diagnostic morphological structures, specifically facilitating family-level identification. To establish a performance benchmark, we evaluated the efficacy of multiple deep learning architectures. For image-level classification, the YOLOv12 model proved optimal, achieving an accuracy exceeding 93%. Subsequently, an object detection pipeline was implemented for the automated detection of the insect body and wings, attaining a mean Average Precision (mAP) of over 90%. Furthermore, Explainable AI (XAI) techniques were employed to interpret the models inference mechanism. These visual analyses confirmed that the network attends to biologically significant featuressuch as specific wing venation patternsrather than confounding background artifacts, thereby validating the taxonomic reliability of the automated predictions. This dataset and its associated benchmarks represent a critical resource for the advancement of computational entomology.