Temporal Feature Interaction for Robust Remote Sensing Image Change Detection: A Taxonomy and Cross-Domain Comparative Study
TL;DR
The Temporal Interaction Taxonomy (TIT) is introduced, which characterizes temporal feature interaction by timing, direction, and operator, and provides a useful dimension for interpreting model behavior across the two evaluated datasets.
Abstract
Remote sensing change detection (RSCD) has advanced through convolutional, attention-based, transformer, and hybrid architectures, yet models are commonly compared as whole architectural families rather than by how their temporal streams interact. This study introduces the Temporal Interaction Taxonomy (TIT), which characterizes temporal feature interaction by timing, direction, and operator. Nine representative models, ranging from early-fusion convolutional baselines to hybrid CNN–Transformer designs, were evaluated using faithful Open-CD implementations under a unified protocol on the LEVIR-CD and WHU-CD building change datasets. TIT provided a consistent basis for describing bi-temporal integration across architectures. Cross-dataset performance was model- and direction-dependent: ChangeFormer achieved the highest mIoU in both transfer directions, reaching 59.99% for WHU→LEVIR and 82.58% for LEVIR→WHU. The results suggest an association between interaction design and cross-dataset robustness, but the independent contribution of temporal interaction cannot be separated from other architectural differences. Transfer also showed model-dependent directional asymmetry; its causes could not be isolated because dataset characteristics and model design were not independently controlled. Overall, temporal interaction provides a useful dimension for interpreting model behavior across the two evaluated datasets. These findings are limited to building change detection on LEVIR-CD and WHU-CD and require task-specific validation before extension to other RSCD applications.