Sanity Check Dataset for Concept-Based XAI

This repository contains the dataset used for performing sanity checks on the landscape concepts available in this repository (https://zenodo.org/records/18936778). These data are used to validate the internal consistency and distinctiveness of the concepts before they are applied to interpret Species Distribution Models (SDMs).

Purpose

To ensure that our learned concepts are statistically meaningful and ecologically consistent, we use these patches for two primary validation steps:
1.  Concept Separability: Comparing positive concept patches (e.g., `Wet`) against their negative counterparts or background noise (e.g., `NoWet`) to ensure the model can distinguish the target feature.
2.  Robust TCAV Self-Influence: Verifying that a concept (e.g., `Road`) shows a strong positive influence on its own classification, confirming that the CAV accurately represents the intended landscape element.

Dataset Structure

The dataset is organized into pairs of "Concept" and "No[Concept]" folders. Each directory follows a standardized 3-modality structure:

├───[Concept_Name] / [NoConcept_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)

Validation Pairs

- Water & Wetlands: `Wet` vs. `NoWet`
- Infrastructure: `Road` vs. `NoRoad`
- Vegetation: `Hedge` vs. `NoHedge`
- Agriculture: `Wheat` vs. `NoWheat`, `TempG` (Temporary Grassland) vs. `NoTempG`
- Farming practices: `Organic` vs. `Convent` (Conventional crops)

Data Specifications

- Input Modalities: 7-band stacked tensors (5 multispectral + 2 LiDAR-derived elevation models).
- Spatial Resolution: 8 cm/pixel.
- Acquisition: Data were collected in April 2024 via a Trinity F90+ drone across five heterogeneous study sites in France.