RDSF-Net: Residual Wavelet Mamba-Based Differential Completion and Spatio-Frequency Extraction Remote Sensing Change Detection Network
Remote sensing change detection is a task of identifying and analyzing the area of surface change by comparing remote sensing images from different periods. It is widely used in many fields such as environmental monitoring, urban planning, and agricultural management. Although the remote sensing cha...
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/10960633/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Remote sensing change detection is a task of identifying and analyzing the area of surface change by comparing remote sensing images from different periods. It is widely used in many fields such as environmental monitoring, urban planning, and agricultural management. Although the remote sensing change detection technology has made great progress in recent years, it still faces many thorny problems: first, the complex heterogeneity of ground objects leads to imperfect processing of the change structure information; second, the influence of nonstationary changes due to seasonal factors. To address these problems, we innovatively propose the residual wavelet mamba-based differential completion and spatio-frequency extraction remote sensing change detection network (RDSF) network. The network is designed with residual wavelet transform as the downsampler, which effectively integrates the key directional information and the overall structural information in the original features, and uses convolutional neural network and Mamba as the backbone network for both long-range and short-range feature extraction. Meanwhile, in order to better capture and compare the differences between time points, we innovatively developed a difference completion sensor to ensure the capture of subtle changes by adjusting the selection, comparison, and dynamic weight assignment between features. In addition, we design a multiscale frequency domain approach that uses a combination of spatial and frequency domain enhancement strategies to reveal the deep structure and boundary changes of the features while reducing the noise interference. RDSF-Net has been extensively experimentally validated on three datasets: the LEVIR-CD, the WHU-CD, and the GZ-CD datasets, and achieved better results than the other state-of-the-art datasets in terms of effect metrics and achieved better results than other state-of-the-art methods. |
|---|---|
| ISSN: | 1939-1404 2151-1535 |