A High-Resolution Landscape Dataset for Concept-Based XAI With Application to Species Distribution Models
Researchers have introduced a novel approach to enhance the interpretability of complex deep learning models used for species distribution modeling (SDMs). This method employs concept-based Explainable AI (XAI) techniques, specifically Robust TCAV, to quantify the influence of landscape concepts on model predictions. To support this, a new open-access dataset of landscape concepts derived from drone imagery has been released, featuring 653 patches across 15 distinct concepts. The approach was demonstrated on aquatic insects, showing that it can validate SDMs against expert knowledge, uncover new ecological hypotheses, and provide landscape-level information valuable for policy and management. AI
IMPACT Enhances interpretability of AI models in ecological research, potentially aiding conservation policy and management.
- arXiv
- explainable AI
- vision transformer
- convolutional neural network
- lidar
- A High-Resolution Landscape Dataset for Concept-Based XAI With Application to Species Distribution Models
- species distribution model
- Robust TCAV
- Testing with Concept Activation Vectors
- Plecoptera
- Trichoptera
- Augustin De La Brosse