PDSDC: Progressive Spatiotemporal Difference Capture Network for Remote Sensing Change Detection
In the field of remote sensing change detection, the dynamic representation of temporal information and the fusion of multisource features have long been key scientific challenges. To address these issues, this article proposes a progressive spatiotemporal difference capture network. This framework...
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/10999080/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | In the field of remote sensing change detection, the dynamic representation of temporal information and the fusion of multisource features have long been key scientific challenges. To address these issues, this article proposes a progressive spatiotemporal difference capture network. This framework effectively mitigates information degradation and multiscale tradeoffs in feature fusion through a multipath optimization mechanism. Specifically, to address information loss and spatial misalignment in bitemporal feature fusion, we propose an innovative linear perception fusion module based on window position encoding. This module employs a dynamic window partitioning strategy to establish geometric correlation constraints between bitemporal features within a local receptive field, achieving pixel-level spatial alignment and channelwise self-attentive fusion of cross-temporal features. Regarding the decoder design, this study introduces a layered progressive decoding architecture, which organically combines large-kernel attention mechanisms with dynamic graph convolution techniques. Specifically, multiple sets of large-kernel adaptive receptive fields are used to capture long-range dependencies while preserving spatial details. For high-level semantic features, a dynamic graph convolutional attention network is constructed, which dynamically establishes topological associations between features through a learnable adjacency matrix, optimizing global semantic consistency through a channel recalibration mechanism. To validate the effectiveness of the proposed method, we conduct systematic experiments on three publicly available benchmark datasets: LEVIR-CD, WHU-CD, and GZ-CD. Experimental results demonstrate that our method outperforms state-of-the-art models in both detection accuracy and efficiency while maintaining real-time inference speed. |
|---|---|
| ISSN: | 1939-1404 2151-1535 |