Extending CAM-based XAI methods for Remote Sensing Imagery Segmentation

AI models analyzing satellite imagery currently operate as "black boxes": The paper identifies that deep learning models used for satellite image analysis lack transparency in how they process data and make decisions, which has hindered their adoption by experts and organizations that need to understand and trust these systems.

Explainability for image segmentation is an underdeveloped area: While explainable AI (XAI) techniques exist for image classification tasks, the paper notes that methods for interpreting image segmentation—a critical capability for applications like identifying buildings from satellite imagery—remain "largely unexplored," representing a significant gap in responsible AI development.

New evaluation methods are needed to properly assess XAI reliability: The authors argue that conventional methods for evaluating XAI techniques are not robust enough, proposing instead an entropy-based metric to better measure model uncertainty and provide more reliable assessment of whether AI explanations are trustworthy.

Adapting existing XAI methods can improve satellite imagery analysis: By extending classification-focused XAI algorithms to work with multi-class segmentation tasks, the paper demonstrates a practical pathway toward making satellite-based AI systems more interpretable and suitable for real-world deployment in applications like urban planning and infrastructure monitoring.