Scholarlink

<p>Reliable flood vulnerability assessment is critical for risk mitigation, yet mainland China faces a persistent bottleneck of historical depth-damage data scarcity. This study proposes an innovative methodology leveraging Large Language Models (LLMs) integrated with Retrieval-Augmented Generation (RAG) to develop flood vulnerability curves for buildings in China. A multi-source data acquisition framework was implemented, combining: (1) sub-component cost ratios derived from 646 empirical records in the Glodon construction database, (2) indoor property inventories from 881 field survey responses, and (3) market values extracted from over 42,000 e-commerce price records. The LLM framework, calibrated through domain-specific knowledge retrieval and three advanced prompting paradigms, Zero-Shot Chain-of-Thought, Self-Consistency Aggregation, and Tree of Thoughts, dynamically generates economic loss ratios for building-component and indoor properties across varying inundation depths. The framework provide a new view for empirically synthesizing depth-damage curves across diverse structural archetypes. The resulting curves are benchmarked against established flood depth-damage references to evaluate plausibility rather than as empirically verified national curves. The results indicate that building-component vulnerability is heavily dictated by typology and the indoor property loss trajectories reveal distinct spatial dependencies for different types of buildings. In general, the Tree of Thoughts paradigm yielded the more consistent loss trajectories among different building types and the Zero-Shot Chain-of-Thought led to more conservative overestimations. This framework provides a scalable, continuously updatable tool for urban flood risk assessment, effectively bridging the empirical data gap for flood vulnerability in mainland China.</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

Integrating multi-source data and large language models in flood vulnerability curve analysis for buildings in mainland China

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月4日

查看全文
  • X. Z. Cui,
  • H. F. Huang,
  • S. H. Shang,
  • B. Bi,
  • Z. D. Duan

摘要

Reliable flood vulnerability assessment is critical for risk mitigation, yet mainland China faces a persistent bottleneck of historical depth-damage data scarcity. This study proposes an innovative methodology leveraging Large Language Models (LLMs) integrated with Retrieval-Augmented Generation (RAG) to develop flood vulnerability curves for buildings in China. A multi-source data acquisition framework was implemented, combining: (1) sub-component cost ratios derived from 646 empirical records in the Glodon construction database, (2) indoor property inventories from 881 field survey responses, and (3) market values extracted from over 42,000 e-commerce price records. The LLM framework, calibrated through domain-specific knowledge retrieval and three advanced prompting paradigms, Zero-Shot Chain-of-Thought, Self-Consistency Aggregation, and Tree of Thoughts, dynamically generates economic loss ratios for building-component and indoor properties across varying inundation depths. The framework provide a new view for empirically synthesizing depth-damage curves across diverse structural archetypes. The resulting curves are benchmarked against established flood depth-damage references to evaluate plausibility rather than as empirically verified national curves. The results indicate that building-component vulnerability is heavily dictated by typology and the indoor property loss trajectories reveal distinct spatial dependencies for different types of buildings. In general, the Tree of Thoughts paradigm yielded the more consistent loss trajectories among different building types and the Zero-Shot Chain-of-Thought led to more conservative overestimations. This framework provides a scalable, continuously updatable tool for urban flood risk assessment, effectively bridging the empirical data gap for flood vulnerability in mainland China.