Modeling High-Order Semantic Abstraction and Temporal Interaction for Remote Sensing Change Detection

Jingxing Zhong, Shurui Zhang, Tianbo Ji, Zhen Tian

Abstract: Remote sensing change detection (RSCD) identifies changes by comparing images of the same area over time. Although deep learning mitigates issues like seasonal variation and pseudochanges, its development in RSCD remains largely heuristic and lacks clear architectural guidance. In this article, we use hidden feature visualization to highlight two essential functionalities that an effective RSCD method should possess, and propose targeted solutions. First, the Siamese-designed single-temporal hierarchical feature extractor should accurately distinguish different semantics in RS images without being constrained by a fixed receptive field. To achieve this, we propose a high-order abstracted spatial dynamic graph aggregation method for semantic enhancement, starting from a non-Euclidean space. This approach enables each spatial unit to flexibly attend to features of interest at arbitrary locations without the limitations of a fixed receptive field, while high-order features improve node abstraction for better semantic knowledge aggregation. Second, the dual-temporal feature interaction should effectively exclude pixel-level changes and avoid pseudochange interference. To this end, we revisit the distinct contributions of $QKV$ in cross-attention across dual-temporal inputs and combine matrix subtraction to generate change-sensitive features. To mitigate the computational cost of cascading cross-attention, we leverage the convolution theorem and design an efficient high-order abstraction method based on frequency-domain interaction, allowing us to obtain higher-level change feature perception. The proposed approach, within a hierarchical change generation framework, achieves 57.23% IoU on CLCD, 90.04% on GoogleBuilding, and 81.22% on LEVCD, outperforming nine state-of-the-art methods across all benchmarks.

External IDs:doi:10.1109/jstars.2025.3634406