Explainable AI for Earth Observation: Current Methods, Open Challenges, and Opportunities

Deep learning models dominate Earth observation but lack transparency: While deep learning has significantly improved performance in remote sensing applications, these powerful neural networks remain "black boxes" that cannot easily explain their decisions—a critical limitation for responsible deployment in real-world Earth observation tasks.

Explainable AI (XAI) techniques are actively being developed for remote sensing: The paper documents a "wide spectrum" of XAI methods being explored specifically to address interpretability challenges in Earth observation, indicating that the field is actively working to make AI decisions more transparent and understandable.

Explainability matters across diverse Earth observation applications: The paper organizes XAI approaches by different application fields, suggesting that the need for interpretable AI extends across multiple domains (such as land use classification, disaster monitoring, or agricultural assessment), making this a broad concern rather than a niche issue.

The gap between performance and explainability remains unresolved: The paper highlights that despite "significant advances in terms of performance," explainability "remains a major source of criticism," indicating that improving accuracy alone is insufficient—stakeholders increasingly demand to understand how AI systems reach their conclusions about Earth's surface.