Trainable Noise Model as an XAI evaluation method: application on Sobol for remote sensing image segmentation

Image segmentation lacks explainability research: While XAI methods have been developed for image classification tasks, image segmentation—a fundamental computer vision task—has received "comparatively less attention" in explainability, particularly in critical applications like remote sensing and medical imaging.

New evaluation method needed for segmentation explanations: The paper proposes a "trainable noise model" as a quantitative way to evaluate XAI methods for segmentation tasks, addressing a gap where most existing XAI evaluation approaches focus on classification rather than segmentation.

Gradient-free methods offer an alternative approach: The research adapts the Sobol XAI method (a gradient-free approach) for segmentation, suggesting that explainability doesn't require traditional gradient-based techniques and may offer different advantages for understanding AI decisions.

Explainability is critical for real-world deployment: The paper emphasizes that transparency in deep learning models is "essential" for adoption in mission-critical domains including autonomous driving, medical image analysis, and remote sensing—areas where understanding AI decision-making directly impacts safety and trust.

Quantitative evaluation of explanations is measurable: The proposed noise-based evaluation method provides an objective metric where explanation quality can be assessed numerically, moving beyond subjective interpretations of whether an AI's reasoning is actually understandable.