From Faithfulness to Validity: Explainable artificial intelligence in remote sensing

Explainable artificial intelligence (XAI) has become increasingly central to remote sensing (RS), as learning-based models are embedded in scientific analysis, operational monitoring, and high-stakes decision making. However, despite the growing number of methods and approaches, explainability is often treated as a transferable capability and assessed using generic criteria inherited from computer vision or tabular learning. Such practices neglect the fact that RS data are governed by strong spatial dependence, spectral redundancy, temporal dynamics, and physical sensing constraints, which fundamentally condition the scientific and operational validity of the explanations. This article revisits XAI in RS from a domain-constrained perspective, framing explainability as a conditional construct whose legitimacy depends on its alignment with the spatial, spectral, temporal, and physical structures of Earth observation data. Rather than cataloging techniques, it synthesizes explainability paradigms, their underlying computational assumptions, evaluation practices, and application regimes to reveal recurrent failure modes that arise when generic XAI approaches are applied without regard to RS-specific constraints. By demonstrating why explanations that appear faithful under standard metrics may be scientifically invalid or operationally misleading, this article reframes explainability around explanatory validity rather than method availability. The resulting framework provides a principled basis for reasoning about XAI choices in RS systems and complements existing works by explicitly linking data structures, model behavior, and interpretability requirements.