Explainable Convolutional Networks for Crater Detection and Lunar Landing Navigation

Explainability is built into AI systems from the start, not added afterward: The paper demonstrates that attention mechanisms can be integrated directly into neural networks during design to make them more transparent, rather than treating explainability as a separate analysis step. This approach helps address expert concerns about AI reliability in critical applications like lunar landing.

Attention mechanisms reveal what AI systems "focus on" during decision-making: By using attention-based networks for crater detection, the researchers can identify which parts of lunar images the AI prioritizes when making navigation decisions, making the system's reasoning process more visible to human operators and engineers.

Explainability can be measured quantitatively: The paper uses Pearson Correlation Coefficient (PCC) to objectively evaluate how well different layers of the network explain their functions, suggesting that transparency in AI isn't purely subjective but can be assessed with concrete metrics.

Explainable AI is essential for high-stakes autonomous systems: The research specifically addresses lunar landing—a mission-critical application where failures have severe consequences. This demonstrates that explainability isn't just a "nice-to-have" feature but a fundamental requirement for AI systems operating in safety-critical environments where human experts need to understand and verify AI decisions.