Answers are grounded in this paper and cite their sources.
Recently, self-supervised pretraining algorithms have been extensively investigated in the field of remote sensing, yielding significant performance improvements in change detection tasks. However, many existing algorithms inadequately exploit multi-level features, which impedes their ability to precisely extract salient change characteristics and accurately determine the structural integrity of changed targets. Moreover, most mainstream approaches tend to neglect the mining of shallow, frequency-domain details, thereby limiting the model's capacity to discern object boundaries. To address these challenges, we propose a novel multi-dimensional feature alignment self-supervised pretraining framework, termed MDFANet. Specifically, we introduce a Hierarchical Decoupling and Multi-dimensional Feature Alignment (HDMFA) mechanism that facilitates the precise capture of multiple change cues at different levels through decoupled modeling of deep and shallow features and the design of multi-dimensional alignment branches. Additionally, we develop a Shallow Frequency-domain Feature Alignment (SFFA) branch that emphasizes the high-frequency components in the input data, effectively compensating for the inadequate representation of edge features in the spatial domain and enhancing the model's ability to detect complex boundaries. Experimental results on LEVIR-CD, SYSU-CD and EGY-BCD datasets validate the superiority of MDFANet, which outperforms SSLCD, Cmid and other SOTA methods by 1.7–2.3% in F1 score.
No takes yet. Share an insight, caveat, or question.
Wang et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: