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

Description Manually-curated hierarchical insect image dataset comprising approximately one million image crops extracted from 1,801 camera-trap video recordings, accompanied by metadata, taxonomy definitions, train/validation/test partitions, trained model weights, and evaluation outputs supporting reproducible biodiversity monitoring research. This archive provides both the image dataset and the associated artefacts required to reproduce the experiments presented in: Mahfoud et al. (2026)"Deep learning-based hierarchical insect classification using camera trap imagery"(submitted for publication) Data Composition Total cropped images: ~939,722 Source videos: 1,801 motion-activated FAIR-Device camera-trap recordings 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 taxonomy consisting of: Level 1 (Root) Insecta Arachnida Level 2 Coleoptera Diptera Hymenoptera Lepidoptera Dermaptera Orthoptera Mecoptera Levels 3-5 Progressively finer taxonomic groupings, including superfamilies, families and functional groups, resulting in: 34 total hierarchy nodes 11 leaf classes used for model training and evaluation The archive includes: ClassificationClasses_IHC.csv Hierarchy_tree.png hierarchy_tree.txt level_name_maps.json to document the taxonomy used in the study. The image dataset retains the original folder labels used during dataset preparation. The accompanying file ClassificationClasses_IHC.csv documents the correspondence between these labels and the scientific taxonomic names used in the manuscript and evaluation outputs. Archive Contents Dataset IHC_dataset.tar.gz Image dataset used for training and evaluation. Metadata insect_images_public.duckdb DuckDB database containing image metadata and hierarchy information. Image paths are stored as portable relative paths to facilitate reproducibility across systems. Experimental Split Files 2026-01-19_portable.tar.gz Contains: Training split Validation split Test split Hierarchy mappings Node-count information Dataset split summary The split files correspond to the published experiment. Only machine-specific file path prefixes were removed to improve portability. Train/validation/test membership, hierarchy mappings, class indices, and sample counts remain unchanged. Model Checkpoint model_Insect_100_96_bz1024_resnet18_OneCycle_2026-01-19_01-09.pth Trained model checkpoint used to generate the results reported in the manuscript. Evaluation Outputs analysis_thr0p6.tar.gz Contains the final evaluation outputs generated using the confidence-thresholded hierarchical inference pipeline, including: Per-level metrics Per-class metrics Coverage statistics Confusion matrices Analysis tables Environment Specification environment.txt Contains the exact Python package versions used during model development, training, and evaluation. This file is intended to support reproducibility of the published experiments. Documentation ClassificationClasses_IHC.csv Hierarchy_tree.png Workflow.png README_ZENODO_IHC.txt Key Characteristics Long-tailed class distribution with substantial class imbalance Variable image resolution Non-lethal camera-trap monitoring methodology Hierarchical taxonomic labels with variable depth Portable train/validation/test splits supplied for exact experiment reproduction Reproducibility This archive includes the dataset, metadata database, hierarchy definitions, train/validation/test partitions, trained model checkpoint, and evaluation outputs required to reproduce the published experiments. The accompanying source code is available at: https://github.com/smAIL-WS/HierarchicalInsectClassification Users wishing to reproduce the complete workflow may: Download the dataset and metadata database. Generate dataset splits using split_train_test.py. Train models using main.py or main_optuna_full.py. Evaluate trained checkpoints using analysis_metrics.py. For exact reproduction of the published results, use the supplied train/validation/test split files, hierarchy files, trained model checkpoint, and environment specification (environment.txt), together with the configuration documented in the GitHub repository README. Funding Acknowledgement We thank the Free State of Bavaria for funding the positions affiliated with the Professorship of Smart Farming within the framework of the Hightech Agenda Bavaria. The camera-trap device used to acquire the raw dataset was developed as part of the joint project "Monitoring of biodiversity in agricultural landscapes" (MonViA) that was funded by the German Federal Ministry of Food and Agriculture (BMEL)