On the Disagreement in Perturbation-based xAI -- Benchmarking Perturbation Choices for Flood Detection from SAR Images

2026-07-16

AI Summary

This study addresses the inconsistency and even contradiction in explanations produced by perturbation-based explainable artificial intelligence (xAI) methods for SAR image flood detection, which arise from varying choices of perturbation parameters. It systematically evaluates the impact of perturbation region geometry—encompassing size and shape—and perturbation type—defined by replacement strategies—on saliency maps. Employing consistency and fidelity metrics alongside visual inspection and cross-strategy validation, this work reveals, for the first time in the context of SAR-based flood detection, the high sensitivity of xAI explanations to perturbation settings. The findings demonstrate that different perturbation strategies can substantially alter or even reverse interpretability conclusions, underscoring the necessity of treating perturbation parameters as a core design component in explanation pipelines to avoid misinterpretation of model decision mechanisms.

Abstract

Perturbation-based xAI methods are widely used to analyze the behavior and predictions of deep learning models. By altering input regions and measuring the resulting changes in class probabilities with respect to the original image, they assign relevance scores and generate heatmaps that reflect each region's contribution to the prediction. Despite their apparent simplicity, however, perturbation-based methods are sensitive to parameter choices. In this work, we focus on two key parameters of the perturbation pipeline, namely the patch geometry, including the size and shape of the perturbed regions, and the perturbation type, defined by the replacement scheme. Grounded in the use case of flood detection from Synthetic Aperture Radar imagery, we conduct a comprehensive investigation of how relevance estimation changes under different perturbation settings. Beyond visual inspection of the resulting relevance maps, we evaluate their consistency across perturbation strategies and their faithfulness to the model's reasoning. We demonstrate how different perturbation choices can steer the resulting relevance maps, yielding ambiguous and even contradictory explanations. Our findings emphasize the importance of methodological settings in perturbation-based xAI. They underscore the need to carefully inspect and evaluate perturbation choices and to treat them as an integral part when interpreting explanations, ensuring a robust understanding of both the explanations and model predictions.

Problem

• perturbation-based xAI
• flood detection
• SAR images
• explanation consistency
• relevance maps

Innovation

• perturbation-based xAI
• flood detection
• SAR imagery
• explanation faithfulness
• patch geometry