Hierarchical Insect Classification Dataset from Camera Trap Imagery [2021-2023] [dataset]

Description Expert-curated dataset of approximately one million insect images extracted from 1,801 camera-trap video recordings, annotated with a five-level, 34-class taxonomic hierarchy for deep learning-based biodiversity monitoring research. Data Composition Total images: ~939,722 cropped still images Source videos: 1,801 motion-activated recordings from FAIR-Device camera traps Collection period: July-August 2021 and August-September 2023 Location: Four grassland sites in Lower Saxony, Germany Taxonomic Structure The dataset uses an asymmetric five-level hierarchical annotation (L1-L5) reflecting biological taxonomy: L1: Insecta, Arachnida (root classes) L2: Orders (Coleoptera, Diptera, Hymenoptera, Lepidoptera, Dermaptera, Orthoptera, Mecoptera) L3-F5: Families and superfamilies (34 terminal classes total) Key Characteristics Highly imbalanced (long-tailed) distribution across classes (3 to 703× variation in sample numbers) Variable image resolution (156 px² - 300,300 px²) Mixed specimen poses due to non-lethal camera trap methodology Includes partially-labeled data where species-level identification was not possible Intended Use Designed for training and evaluating hierarchical deep learning classification models for automated insect biodiversity monitoring, particularly for handling imbalanced datasets and variable-depth taxonomic hierarchies. Associated Research Manuscript: "Deep learning-based hierarchical insect classification using camera trap imagery" by Mahfoud et al. (submitted for publication, 2026) Authors' Affiliation: Weihenstephan-Triesdorf University of Applied Sciences, Germany Associated code: https://github.com/smAIL-WS/HierarchicalInsectClassification Funding Acknowledgement This work was funded by the Free State of Bavaria (Hightech Agenda Bavaria) and the German Federal Ministry of Food and Agriculture (BMEL) through the MonViA project.