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

A research team led by Augustin de la Brosse has published a groundbreaking paper and dataset that applies concept-based Explainable AI (XAI) to the ecological challenge of predicting species distributions. The work addresses a critical gap: while deep learning models like Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) are increasingly used for Species Distribution Models (SDMs), their complexity makes it difficult for ecologists to extract meaningful insights about *why* a model predicts a species will be found in a certain area. The team's solution is the first implementation of Robust TCAV (Testing with Concept Activation Vectors) for SDMs, a method that quantifies the influence of human-understandable concepts—like 'dense forest' or 'open water'—on a model's predictions.

To enable this method, the researchers created and released a novel, open-access dataset derived from high-resolution drone imagery. It includes 653 meticulously labeled patches across 15 distinct landscape concepts, plus 1,450 random reference patches, totaling 2,103 data points. In a case study on two aquatic insects (Plecoptera and Trichoptera), applying Robust TCAV to CNN and ViT models allowed the team to validate model decisions against expert knowledge and, importantly, uncover novel species-habitat associations that can form new testable hypotheses for ecologists. The publicly available code and dataset provide a new toolkit for making powerful but opaque ecological AI models transparent and actionable for conservation policy and land management.