Scholarlink

<p>Remote sensing change detection (RSCD) has progressed from task-specific convolutional architectures to foundation-integrated and multimodal paradigms. However, existing surveys largely categorize methods by architectural families and do not explicitly capture paradigm level transitions supported by quantitative evidence. This paper presents a hierarchical taxonomy that organizes RSCD into three evolutionary pillars: (i) architectural evolution, (ii) foundation and Multimodal Paradigm Shift, and (iii) Data-centric, Explainable, and Operational CD. To substantiate this taxonomy, we conduct a harmonized quantitative meta-analysis on dominant benchmarks, particularly LEVIR-CD and WHU-CD, enabling controlled cross-paradigm evaluation. Our analysis shows that architectural models consistently plateau at 90–93% F1, whereas foundation-adapted approaches reach up to 95.77% (LEVIR-CD) and 97.23% (WHU-CD). Notably, recent multimodal systems exceed 100&#xa0;M parameters but yield limited proportional accuracy gains. In contrast, knowledge distillation achieves state-of-the-art performance with moderate model complexity, offering the most favorable efficiency-accuracy trade-off. Temporal trend analysis further indicates performance saturation on current benchmarks and diminishing returns from parameter scaling. Overall, RSCD is transitioning from architecture-driven optimization toward representation reuse, parameter-efficient adaptation, and deployment-aware design. Future advances will depend on cross-domain generalization, label-efficient learning, multimodal robustness, and operational scalability.</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

From architectural evolution to foundation driven paradigms: a hierarchical and quantitative survey of remote sensing change detection

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月16日

查看全文
  • Varun Chandra Kola,
  • Sujatha Pothula,
  • Sathya M

摘要

Remote sensing change detection (RSCD) has progressed from task-specific convolutional architectures to foundation-integrated and multimodal paradigms. However, existing surveys largely categorize methods by architectural families and do not explicitly capture paradigm level transitions supported by quantitative evidence. This paper presents a hierarchical taxonomy that organizes RSCD into three evolutionary pillars: (i) architectural evolution, (ii) foundation and Multimodal Paradigm Shift, and (iii) Data-centric, Explainable, and Operational CD. To substantiate this taxonomy, we conduct a harmonized quantitative meta-analysis on dominant benchmarks, particularly LEVIR-CD and WHU-CD, enabling controlled cross-paradigm evaluation. Our analysis shows that architectural models consistently plateau at 90–93% F1, whereas foundation-adapted approaches reach up to 95.77% (LEVIR-CD) and 97.23% (WHU-CD). Notably, recent multimodal systems exceed 100 M parameters but yield limited proportional accuracy gains. In contrast, knowledge distillation achieves state-of-the-art performance with moderate model complexity, offering the most favorable efficiency-accuracy trade-off. Temporal trend analysis further indicates performance saturation on current benchmarks and diminishing returns from parameter scaling. Overall, RSCD is transitioning from architecture-driven optimization toward representation reuse, parameter-efficient adaptation, and deployment-aware design. Future advances will depend on cross-domain generalization, label-efficient learning, multimodal robustness, and operational scalability.