Answers are grounded in this paper and cite their sources.
In high-resolution remote sensing (RS) image build ing change detection (BCD) tasks, false detections caused by ra diometric differences and geometric deformations have long con strained detection accuracy and model robustness. To address this challenge, we propose a dynamic weights and cross-temporal modeling network (DWCTNet) that achieves accurate perception of building changes in complex scenes by constructing a multi stage feature optimization and spatiotemporal relationship mod eling framework. Specifically, DWCTNet utilizes a Siamese backbone network to extract multi-scale features from multi temporal images. Subsequently, a dynamic weight module (DWM) is designed to establish feature correlations between temporal phases through a cross-temporal joint attention mechanism while integrating multi-scale spatial attention to achieve dynamic weighting of local and global features. Furthermore, a cross temporal interaction module (CTIM) is proposed. This module integrates a bidirectional cross-attention mechanism with a de formable convolutional network to enhance its adaptability to building displacement and deformation. Additionally, a hierar chical decoder gradually fuses shallow detail features and deep semantic features to generate a high-resolution change detection map. The experimental results on the WHU-CD, GZ-CD, and LEVIR-CD datasets demonstrate that DWCTNet attains F1 scores of 93.66%, 88.64%, and 91.35%, respectively, with IoU metrics of 88.07%, 79.61%, and 83.39%, thereby demonstrating superiority over existing methods.
No takes yet. Share an insight, caveat, or question.
Wei Mao (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: