Answers are grounded in this paper and cite their sources.
Existing supervised and semi-supervised methods for SAR image change detection are limited by the scarcity of labeled data and the challenges of modeling spatial-temporal correlations. To address these issues, we propose STSGNet, a graph contrastive learning-based framework that captures spatial-temporal correlations from large amounts of unlabeled data and leverages limited labeled data for fine-tuning, thereby enhancing the model's change detection performance. STSGNet consists of three primary modules: the Spatial-Temporal Semantic module (STS), the Spatial-Temporal Geographic module (STG), and the Dynamic Weighted Fusion module (DWF). The STS module integrates a Bitemporal Semantic Enhancement Module (BSEM) to capture cross-temporal semantic similarities and differences, while the STG module applies a geographic correlation data augmentation strategy and incorporates Geographic Transformer Blocks (GTB) to integrate geographic information, enhancing spatial representation and contextual reasoning. The DWF module effectively integrates the semantic and geographic features, producing comprehensive spatial-temporal feature representations. By extending the model to the complex-valued domain, STSGNet incorporates both amplitude and phase information, enabling a more detailed perception of potential changes. STSGNet comprehends complex inter-dependencies across time and space, enabling accurate predictions of change areas even with limited labeled data. Experiments on five diverse SAR image scenes confirm that STSGNet surpasses baseline methods under both unsupervised and limited supervised conditions, validating its effectiveness
No takes yet. Share an insight, caveat, or question.
Li et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: