Scholarlink

<p>Remote sensing change detection (CD) in complex urban scenes faces two persistent challenges: distinguishing structural changes from environmental interference (pseudo-changes) and maintaining geometric boundary integrity. While recent Transformer-based approaches excel at capturing global semantic context, they often dilute high-frequency spatial details, resulting in blob-like change maps with blurred edges. To address these limitations, this paper proposes an Edge-Aware Bi-Temporal Feature Gating (EBFG) network. Unlike standard hybrid architectures that simply concatenate features, EBFG introduces a&#xa0;Background-Suppressing Fusion Module that explicitly constructs a&#xa0;Feature Commonality Map to gate out irrelevant variations caused by shadows and illumination before deep encoding. Furthermore, we design a&#xa0;Gradient-Guided Spatial Gating mechanism in the decoder, which dynamically re-weights multi-scale features to recover fine-grained object boundaries that are typically lost during downsampling. To enforce structural fidelity, the network is optimized using a&#xa0;composite loss function that integrates pixel-wise classification with gradient-based edge supervision. Extensive experiments on three benchmark datasets (LEVIR-CD, BCDD-CD, and SYSU-CD) demonstrate that EBFG achieves state-of-the-art performance, yielding F1 scores of 91.73%, 95.14%, and 82.91%, respectively. Qualitative results confirm that the proposed method significantly reduces false positives in shadow-covered regions and produces sharper, more geometrically accurate change maps compared to existing Transformer-based baselines.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

EBFG-CD: Edge-Aware Bi-Temporal Feature Gating for Remote Sensing Change Detection

  • 期刊论文

  • 发表日期：

    2026年7月6日

查看全文
  • Muhammad Awais Anjum,
  • Di Zhuang,
  • Lamei Zhang,
  • Bin Zou

摘要

Remote sensing change detection (CD) in complex urban scenes faces two persistent challenges: distinguishing structural changes from environmental interference (pseudo-changes) and maintaining geometric boundary integrity. While recent Transformer-based approaches excel at capturing global semantic context, they often dilute high-frequency spatial details, resulting in blob-like change maps with blurred edges. To address these limitations, this paper proposes an Edge-Aware Bi-Temporal Feature Gating (EBFG) network. Unlike standard hybrid architectures that simply concatenate features, EBFG introduces a Background-Suppressing Fusion Module that explicitly constructs a Feature Commonality Map to gate out irrelevant variations caused by shadows and illumination before deep encoding. Furthermore, we design a Gradient-Guided Spatial Gating mechanism in the decoder, which dynamically re-weights multi-scale features to recover fine-grained object boundaries that are typically lost during downsampling. To enforce structural fidelity, the network is optimized using a composite loss function that integrates pixel-wise classification with gradient-based edge supervision. Extensive experiments on three benchmark datasets (LEVIR-CD, BCDD-CD, and SYSU-CD) demonstrate that EBFG achieves state-of-the-art performance, yielding F1 scores of 91.73%, 95.14%, and 82.91%, respectively. Qualitative results confirm that the proposed method significantly reduces false positives in shadow-covered regions and produces sharper, more geometrically accurate change maps compared to existing Transformer-based baselines.