Abstract
In recent years, deep learning-based methods for remote sensing change detection have developed rapidly. However, radio-metric differences between bitemporal images caused by seasonal or illumination variations often make changed regions difficult to detect accurately. Motivated by its multiscale segmentation-oriented feature representations, we introduce the frozen Segment Anything Model 2 (SAM2) image encoder into the framework. Based on this encoder, we propose a SAM2-based spatiotemporal-aware change detection network (S2TANet). S2TANet integrates two components: a spatiotemporal-aware (STA) adapter and a dynamic adaptive scanning change (DASC) decoder. The STA adapter extracts difference features from bitemporal images and enhances them through spatial attention. In a single forward pass, the attention-guided difference features are fed back into the bitemporal branches. This operation enhances feature consistency in unchanged regions while enlarging feature discrepancies in changed regions, thereby mitigating radiometric interference. The DASC decoder is composed of multiple DASC blocks. Each block employs a dynamic adaptive scanning mechanism. Specifically, an offset prediction network (OPN) predicts offsets for scanning reference points from local features. The predicted offsets displace the scanning reference points to locate patches of interest and adapt the scanning strategy. This content-dependent scanning enhances the detection of irregular changes under complex backgrounds. Extensive experiments on SYSU-CD, SECOND-CD, and LEVIR-CD show that S2TANet outperforms the strongest competing methods. The corresponding F1-score gains are 1.87, 2.00, and 0.43 percentage points, respectively.
| Original language | English |
|---|---|
| Journal | IEEE Sensors Journal |
| DOIs | |
| Publication status | Accepted/In press - 2026 |
Keywords
- Change detection (CD)
- Segment Anything Model 2 (SAM2)
- adapter
- deep learning
- remote sensing
Fingerprint
Dive into the research topics of 'S2TANet: SAM2-based Spatiotemporal-Aware Network for Remote Sensing Change Detection'. Together they form a unique fingerprint.Cite this
- APA
- Author
- BIBTEX
- Harvard
- Standard
- RIS
- Vancouver