Training Dataset for Species Distribution Modeling (SDM)

This repository contains the training, validation, and testing patches used to develop Species Distribution Models (SDMs) for aquatic insects (Plecoptera and Trichoptera). The data comprises multimodal drone-acquired imagery across five heterogeneous study sites in France.

Dataset Structure

The dataset is split into train_patches, val_patches, and test_patches. Each split is organized by study site, following a standardized 3-modality structure:

├───[train/val/test]_patches
│   ├───[site_name]_spring
│   │   ├───image_patches  # 5-band multispectral data (B, G, R, RE, NIR)
│   │   ├───dsm_patches    # Digital Surface Model (Canopy elevation)
│   │   └───dtm_patches    # Digital Terrain Model (Ground elevation)

Study Sites

– cisse: Cissé
– fao: Fao
– louroux: Louroux
– roudoudour: Roudoudour
– timbertiere: Timbertière

Labels and Data Association

A CSV file containing the ground truth labels is provided. To correctly associate a patch with its label, users must parse the file basename of the patches, which contains two key identifiers:
1.  Site Name: Matches the `site` column in the CSV (e.g., “roudoudour”, “fao”).
2.  Unique ID: Matches the `id` column in the CSV.

Each sampling point corresponds to a specific sticky trap location used during the April 2024 emergence period.

Data Specifications

– Input Modalities: 7 total bands (5 multispectral bands + 2 LiDAR-derived elevation models).
– Spatial Resolution: 8 cm/pixel.
– Acquisition: Data were collected in April 2024 using a Trinity F90+ drone (MicaSense Dual MX and Qube240 sensors).
– Site Diversity: The five sites represent a gradient of agricultural intensity, from extensive dairy farming to intensive cropping systems, ensuring high model generalizability.