Scholarlink

<p>Change detection in remote sensing imagery is essential for tracking land surface transformations, informing urban planning, environmental monitoring, and disaster response. Conventional approaches often face limitations in urban environments, where subtle changes, spectral inconsistencies, and computational constraints complicate detection tasks. To address these, a hybrid model Multiscale Parallel Residual Transformer with Skeleton-based Attention Network (MPRT-SAN) is proposed, combining advanced preprocessing, feature extraction, and optimization. Initial image enhancement uses hybrid bilateral guided filtering to preserve edges and reduce noise. Features are extracted via a Three-branch CNN (Three branch CNN) integrated with Principal Component Analysis to capture spatial, temporal, and bitemporal information efficiently. Feature fusion employs the Mantis Search Gold Rush optimization for robust integration while preserving discriminative power. The fused features are processed by MPRT-SAN, extracting hierarchical, context-aware multi-scale representations. White Shark Optimizer fine-tunes the transformer layers to reduce computational complexity and improve precision. This integrated approach overcomes limitations of traditional methods, delivering strong performance across datasets. On LEVIR-CD, the model achieves 99% accuracy, 99% precision, 90% recall, 91% F1-score, 85% IoU, and 93% Cohen’s Kappa. For SYSU-CD, it records 98% accuracy, 97% precision, 91% recall, 93% F1-score, 88% IoU, and 90% Kappa. On the challenging OSCD dataset, it maintains 96% accuracy, 91% precision, 90% recall, 90% F1-score, 80% IoU, and 90% Kappa, demonstrating robustness and effectiveness in complex remote sensing scenarios.</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

Hybrid Three Branch CNN and Transformer with Skeleton Based Attentions Network for Change Detection in Remote Sensing Images

  • 期刊论文

  • 发表日期：

    2025年7月28日

查看全文
  • Namdeo B. Badhe,
  • Vijaykumar P. Yele,
  • Rahul P. Neve,
  • Jisha Tinsu

摘要

Change detection in remote sensing imagery is essential for tracking land surface transformations, informing urban planning, environmental monitoring, and disaster response. Conventional approaches often face limitations in urban environments, where subtle changes, spectral inconsistencies, and computational constraints complicate detection tasks. To address these, a hybrid model Multiscale Parallel Residual Transformer with Skeleton-based Attention Network (MPRT-SAN) is proposed, combining advanced preprocessing, feature extraction, and optimization. Initial image enhancement uses hybrid bilateral guided filtering to preserve edges and reduce noise. Features are extracted via a Three-branch CNN (Three branch CNN) integrated with Principal Component Analysis to capture spatial, temporal, and bitemporal information efficiently. Feature fusion employs the Mantis Search Gold Rush optimization for robust integration while preserving discriminative power. The fused features are processed by MPRT-SAN, extracting hierarchical, context-aware multi-scale representations. White Shark Optimizer fine-tunes the transformer layers to reduce computational complexity and improve precision. This integrated approach overcomes limitations of traditional methods, delivering strong performance across datasets. On LEVIR-CD, the model achieves 99% accuracy, 99% precision, 90% recall, 91% F1-score, 85% IoU, and 93% Cohen’s Kappa. For SYSU-CD, it records 98% accuracy, 97% precision, 91% recall, 93% F1-score, 88% IoU, and 90% Kappa. On the challenging OSCD dataset, it maintains 96% accuracy, 91% precision, 90% recall, 90% F1-score, 80% IoU, and 90% Kappa, demonstrating robustness and effectiveness in complex remote sensing scenarios.