Answers are grounded in this paper and cite their sources.
Remote sensing (RS) classification models have become dominant tools for RS image analysis, achieving state-of-the-art performance in many downstream tasks. However, the internal mechanisms of their feature representation remain underexplored, raising critical questions: Do RS classification models exhibit common tendencies in encoding geospatial features? Do these feature-representation mechanisms align with human visual cognition of geographic scenes? To address these questions, this paper proposes a methodological framework based on multi-order Shapley contextual interactions to interpret RS classification models. We leverage the Shapley value, rooted in cooperative game theory, to quantify the contribution of individual input variables to model outputs and decompose it into multi-order interactions that capture contextual dependencies of varying complexities. A key finding is the identification of a feature representation bottleneck in RS classification models: these models efficiently encode low-order and high-order interactions but struggle to learn intermediate-order interactions, which serve as a critical bridge between basic features and full scene semantics. This bottleneck creates a cognition gap with humans, who rely on intermediate-scale contexts for geographic scene understanding. Under idealized statistical assumptions, our gradient analysis provides a possible theoretical explanation for why learning signals associated with intermediate-order interactions may be attenuated relative to those at extreme orders. We validate this bottleneck across diverse RS tasks (e.g. land cover classification) and architectures (e.g. ResNet and Transformer). Our work not only reveals a recurring limitation under the evaluated settings but also provides actionable insights for designing more interpretable and robust RS models.
No takes yet. Share an insight, caveat, or question.
Jianya et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: