Scholarlink

<p>In order to respond to emergencies, estimate damage, and distribute resources effectively, remote sensing photography must be able to detect changes induced by disasters rapidly and precisely. Problems with bi-temporal change detection include the need for masks that are precise at the boundary and can be easily integrated into GIS workflows, as well as significant seasonal/illumination variability and residual co-registration problems. In this study, the DisasterChangeNet framework is presented. It can accept optical and Synthetic Aperture Radar (SAR) inputs and is built to detect changes in disaster-related data utilising pre- and post-event observations. A lightweight radiometric harmonisation is performed, dual-stream representations are learnt with cross-temporal interaction to avoid misalignment-driven false alarms, and gated multi-scale fusion is applied for strong semantic change separation in order to handle nuisance variation. For mapping to be feasible, it is essential to improve the localisation accuracy of thin structures and sharp edges. Roads, riverbanks, and building outlines are some of the many items that undergo this process in a dedicated border refining step. Every experiment follows the same protocol, whether it’s the LEVIR-CD benchmark for urban building change, the SEN12-FLOOD for multimodal flood mapping, or the xBD for large-scale building damage assessment. In comparison to well-known CNN baselines (FC-EF, FC-Siam-Conc) and a robust densely connected Siamese decoder (SNUNet-CD), DisasterChangeNet not only achieves better overall accuracy (IoU = 0.789, F1 = 0.882, and Kappa = 0.845), but it also enhances boundary fidelity (BFScore = 0.731) and maintains inference performance that is suitable for deployment (≈ 38 ms/image, ~ 26 FPS). The results show that boundary-aware fusion is crucial for precise large-scale catastrophe change mapping.</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

DisasterChangeNet: Boundary-Preserving Deep Change Detection in Remote Sensing for Disaster Management

  • 期刊论文

  • 发表日期：

    2026年8月3日

查看全文
  • V. N. Vinaykumar,
  • K. Anusha,
  • G. S. Nandeesh,
  • R. Varadaraj,
  • N. Naveen,
  • M. S. Lohith

摘要

In order to respond to emergencies, estimate damage, and distribute resources effectively, remote sensing photography must be able to detect changes induced by disasters rapidly and precisely. Problems with bi-temporal change detection include the need for masks that are precise at the boundary and can be easily integrated into GIS workflows, as well as significant seasonal/illumination variability and residual co-registration problems. In this study, the DisasterChangeNet framework is presented. It can accept optical and Synthetic Aperture Radar (SAR) inputs and is built to detect changes in disaster-related data utilising pre- and post-event observations. A lightweight radiometric harmonisation is performed, dual-stream representations are learnt with cross-temporal interaction to avoid misalignment-driven false alarms, and gated multi-scale fusion is applied for strong semantic change separation in order to handle nuisance variation. For mapping to be feasible, it is essential to improve the localisation accuracy of thin structures and sharp edges. Roads, riverbanks, and building outlines are some of the many items that undergo this process in a dedicated border refining step. Every experiment follows the same protocol, whether it’s the LEVIR-CD benchmark for urban building change, the SEN12-FLOOD for multimodal flood mapping, or the xBD for large-scale building damage assessment. In comparison to well-known CNN baselines (FC-EF, FC-Siam-Conc) and a robust densely connected Siamese decoder (SNUNet-CD), DisasterChangeNet not only achieves better overall accuracy (IoU = 0.789, F1 = 0.882, and Kappa = 0.845), but it also enhances boundary fidelity (BFScore = 0.731) and maintains inference performance that is suitable for deployment (≈ 38 ms/image, ~ 26 FPS). The results show that boundary-aware fusion is crucial for precise large-scale catastrophe change mapping.