A High-Resolution Landscape Dataset for Concept-Based XAI With Application to Species Distribution Models

This zenodo repository contains a multimodal, high-resolution dataset of 2,103 patches (653 concept-specific and 1,450 random) extracted from multispectral and LiDAR drone data. It is designed to support concept-based XAI (e.g., TCAV) for modeling species distribution at fine-scale.

Dataset Structure

The dataset is organized by concept class. Each directory represents a specific landscape element (concept) and follows a standardized 3-modality structure:

├───[Concept_Name]│   ├───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)

Concept Classes

Vegetation: Hedge (Hedgerows), IsoTree (Isolated Trees), Wood (Woodlands).

Agriculture: Cereal, Maize, Wheat, PermG (Permanent Grassland), TempG (Temporary Grassland).

Farming Systems: Organic (Organic crops), Convent (Conventional crops).

Water & Wetlands: LinW (Linear Water), SurfW (Surface Water), Wet (Wetlands).

Infrastructure: Build (Buildings), Road (Roads).

Baseline: random_images (1,450 randomly sampled background patches).

Study Sites and Acquisition

The data were acquired in April 2024 using a Trinity F90+ drone equipped with MicaSense Dual MX and Qube240 sensors. To ensure a robust representation of diverse agricultural landscapes, data were collected across five heterogeneous study sites in France, ranging from extensive dairy farming systems to highly intensive cropping systems.

Data Specifications

Spatial Resolution: 8 cm/pixel.

Patch Size: 512 × 512 pixels.

Input Channels: 7 total bands (5 multispectral + 2 LiDAR-derived elevation models).