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

Researchers have released an open-access dataset of high-resolution landscape imagery designed to make deep-learning species distribution models more interpretable through concept-based explainable AI, according to a paper posted on arXiv [1]. Species distribution models are a cornerstone of conservation policy and invasive species management, but the rising complexity of deep-learning approaches has made it harder to extract ecological insights from their predictions [1][2]. To address that tension, a team led by Augustin De La Brosse proposes applying concept-based Explainable AI to the field for the first time [1][2]. The work relies on Robust TCAV — Testing with Concept Activation Vectors — to measure how strongly specific landscape concepts influence a model’s output [1][2]. The methodology requires a dedicated concept dataset, which the authors constructed from high-resolution multispectral and LiDAR drone imagery [1][2]. The resulting collection contains 653 patches spanning 15 distinct landscape concepts, alongside 1,450 random reference patches, and is intended to be reusable across a wide range of species [1][2]. The paper demonstrates the approach through a case study of two aquatic insect groups, Plecoptera and Trichoptera, using two Convolutional Neural Networks and one Vision Transformer [1][2]. The authors report that concept-based XAI both validates model behavior against expert knowledge and surfaces novel associations that can generate new ecological hypotheses [1][2]. They add that the landscape-level information produced by Robust TCAV holds potential value for policy-making and land management [1][2]. The preprint was first submitted to arXiv on 14 April 2026 as version one, weighing 11,284 KB, and was revised on 18 June 2026 with a second version of 17,194 KB [1]. arXiv, which was launched in 1991, is an open-access repository that hosts e-prints across mathematics, physics, computer science, and related fields after moderation but without peer review [6]. As of November 2024, the platform was receiving roughly 24,000 new articles per month [6]. The authors have made both the code and the landscape dataset publicly available [1][2]. The paper appears under the Computer Vision and Pattern Recognition category on arXiv, where readers can access a range of experimental community tools through the arXivLabs framework, including citation explorers and code-finding services [1][4][5].