Survey of Perturbation Approaches for Explainable ML in the Context of Flood Detection from SAR Images

Abstract

Machine learning and especially deep convolutional networks (ConvNets) are increasingly being used for various image analysis tasks in Earth observation. Despite their strong performance, ConvNets are considered black boxes lacking explainability of their predictions. Methods under the umbrella term “explainable machine learning” or more “explainable AI” (XAI) aim to provide human-interpretable reasoning for why a model made a particular prediction. Amongst them, perturbation techniques explore changes in the prediction when the input is locally distorted. We investigate the influence of different parameter choices on the quality of explanations in the context of flood detection using SAR images.