Towards Explainable AI for Binary Change Detection in Satellite Images

The rapid growth of urban areas and human activities makes automatic change detection in remote sensing essential for tasks such as monitoring urban growth, environmental changes, and disasters. Deep learning methods have shown great promise for these tasks, yet they often lack transparency, making it challenging to build trust in their results. This study addresses model interpretability in change detection by combining the predictive accuracy of deep learning with explainable artificial intelligence (XAI) techniques. A Siamese U-Net model, paired with Grad-CAM and Grad-CAM++ visualizations, is utilized to produce intuitive visual explanations of the model’s predictions. Validated three change detection datasets, namely CDD, LEVIR-CD, and SyntheWorld, which contain high resolution optical satellite images from various urban and natural environments, the presented approach enhances the interpretability of deep learning-based change detection by providing insights into how the model classifies each pixel as "change" or "no change", and aims to improve confidence in its outputs.