MSA: Mamba Semantic Alignment Networks for Remote Sensing Change Detection
With the rapid advancement of Earth observation technologies, remote sensing change detection (CD) has become a crucial method for monitoring surface changes. It is widely used in areas, such as urban expansion, disaster assessment, and resource detection. Current deep learning-based CD methods typi...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/10946760/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | With the rapid advancement of Earth observation technologies, remote sensing change detection (CD) has become a crucial method for monitoring surface changes. It is widely used in areas, such as urban expansion, disaster assessment, and resource detection. Current deep learning-based CD methods typically extract feature information from remote sensing images through downsampling and then aggregate early features with deeper ones during upsampling. However, directly aggregating these features without addressing spatial misalignment due to resolution changes can compromise the accuracy of change detection. In addition, there is a need to address the challenge of inadequate long-range dependency modeling in image processing. To tackle these challenges, this article proposes Mamba semantic alignment networks (MSA) for remote sensing CD. MSA introduces the semantic offset correction block, which corrects spatial misalignment during feature aggregation by incorporating a learnable semantic offset map, thereby reducing classification errors caused by feature mismatches. Furthermore, MSA incorporates the global dependency enhancement block, leveraging the Mamba architecture and the lossless downsampling and reversibility of wavelet transforms to significantly enhance global feature modeling. We evaluated MSA on three datasets, and the experimental results demonstrate that MSA outperforms mainstream methods across all three datasets. |
|---|---|
| ISSN: | 1939-1404 2151-1535 |