Randomized trial evaluates LDST-ChangeNet for change detection in remote sensing imagery, suggesting improved efficiency.
Key Points
This study aims to develop a lightweight model for change detection in remote sensing images to improve efficiency and reduce errors caused by pseudo-changes.
Proposed LDST-ChangeNet, a dual spatiotemporal attention network using a Siamese EfficientNet-B1 encoder.
Implemented differential bi-temporal feature fusion for temporal discrepancy modeling.
Introduced Position Attention Module and Pyramid Pooling Module to enhance feature extraction and representation.
Achieved F1-scores of 90.67% on LEVIR-CD and 91.08% on WHU-CD datasets.
Maintained low parameters: 11.72 M and 10.03 GFLOPs on LEVIR-CD; 11.77 M and 9.12 GFLOPs on WHU-CD.