Change-Guided Difference Interaction Attention Network for Remote Sensing Change Detection
Change Detection (CD) is a fundamental task in remote sensing image analysis, widely applied in fields such as municipal planning and vital signs monitoring. However, many existing methods struggle to extract change-relevant features when faced with dual-temporal remote sensing images characterized...
Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Access |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/11006998/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Change Detection (CD) is a fundamental task in remote sensing image analysis, widely applied in fields such as municipal planning and vital signs monitoring. However, many existing methods struggle to extract change-relevant features when faced with dual-temporal remote sensing images characterized by inconsistent and complex feature distributions, leading to false alarms. Moreover, these methods rely on simple concatenation, differential operations, or the addition of attention mechanisms to fuse dual-temporal features, resulting in the loss of edge information. Additionally, these fusion strategies fail to fully exploit the relationship between dual-temporal features and difference features, which negatively impacts the overall performance of CD. Furthermore, the problem of internal holes remains unresolved. To address these challenges, we propose the change-guided difference interaction attention network (CGDIANet). This network effectively establishes interaction between dual-temporal features through difference interaction attention module (DIAM), enhancing the capability to extract change features. During the feature fusion stage, the edge-enhanced difference fusion module (EEDFM) thoroughly integrates dual-temporal features with difference features, and employs an edge enhancement mechanism to prevent the loss of edge information. In the decoding stage, multi-scale change guidance module (MSCGM) combines prior change information with enhanced multi-scale dilated convolutions, addressing the limited receptive field of traditional CNN decoders, thereby effectively mitigating the internal holes. Experimental results on three public datasets demonstrate that CGDIANet outperforms ten existing advanced methods. |
|---|---|
| ISSN: | 2169-3536 |