Revealing Feature Contribution Mechanisms for Remote Sensing Scene Understanding

Deep learning plays a central role in remote sensing scene understanding, making interpretability essential for analyzing and trusting model decisions. Feature contribution analysis is a key interpretability tool, yet existing methods often rely on artificial feature conflicts or feature suppression, which are easily confounded by strong semantic correlations in remote sensing imagery. (In particular, the fixed overhead viewing geometry tightly couples object shape with semantic category, which biases feature contribution estimation and misleads the interpretation of intrinsic model preferences, thus obscuring the genuine feature utilization patterns of models.) To address these limitations, we propose a systematic feature contribution analysis framework that integrates multi-modal feature decoupling with dynamic contribution aggregation. By disentangling shape, texture, and spectrum representations and progressively aggregating them, the proposed method enables unbiased quantification of feature contributions. The framework supports cross-architecture and cross–data set analysis. Extensive experiments reveal clear architectural- and data set–dependent feature preference patterns: convolutional neural networks exhibit an inherent texture bias across remote sensing tasks, while Vision Transformers realize balanced integration of shape, texture, and spectrum in object-level classification and shift to spectral feature dominance in scene-level land cover classification. We further find that the feature preference of remote sensing deep learning models is jointly determined by network inductive biases and data set characteristics rather than a single architectural attribute, offering new insights into remote sensing deep learning models.