Scholarlink

<p>Landslide susceptibility assessment (LSA) primarily evaluates the spatial likelihood of landslide occurrence, but its effectiveness is constrained by the incompleteness of spatiotemporal data in historical inventories. In particular, the growing frequency of extreme weather and intensified human activities have made traditional approaches—often relying on multi-year averages of dynamic factors—less sensitive to short-term variations, thereby limiting their capacity to capture dynamic changes in landslide conditions. Therefore, to address the lack of spatiotemporal information in historical landslide inventories and to enhance the prediction accuracy of LSA, we propose an LSA framework that integrates multimodal data. First, textual information (e.g., disaster investigation reports, news, government bulletins, and academic papers) was systematically processed using a fine-tuned BERT-BiLSTM-CRF model, from which 7695 landslide events and transformed into 28,472 spatiotemporal triples of landslides, each triple serves as a standardized knowledge unit, uniformly represented as (event, location, time), explicitly describing when and where a specific landslide occurred. Second, the spatiotemporal entities in these triples were used as anchor points to align with the remote sensing change detection algorithm on the Google Earth Engine platform, enabling the identification and delineation of 1745 historical landslides with both occurrence time and spatial extent. Finally, by integrating non-time series data such as topographic–geological factors with time-series meteorological and engineering parameters (e.g., precipitation, temperature, road and river distances), and applying the random forest, the time sensitivity of the LSA in the entire study area was evaluated. The results show that compared with the multi-year average parameters, the construction of time-series parameters improves the LSA prediction F1 score from 78.16 to 83.72%. Combined with the meteorological factors and artificial factors of the dynamic time window, it has the potential for landslide early warning. The proposed framework improves the spatiotemporal resolution of LSA, which provides a strong support for dynamic monitoring, risk assessment, and disaster prevention decision-making.</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 textual and remote sensing data for time-sensitivity of landslide susceptibility assessment

  • 期刊论文

  • 发表日期：

    2026年1月21日

查看全文
  • Qirui Wu,
  • Zhong Xie,
  • Yifan Zhao,
  • Miao Tian,
  • Yueyu Wu,
  • Qinjun Qiu,
  • Qianhe Xiang,
  • Liufeng Tao

摘要

Landslide susceptibility assessment (LSA) primarily evaluates the spatial likelihood of landslide occurrence, but its effectiveness is constrained by the incompleteness of spatiotemporal data in historical inventories. In particular, the growing frequency of extreme weather and intensified human activities have made traditional approaches—often relying on multi-year averages of dynamic factors—less sensitive to short-term variations, thereby limiting their capacity to capture dynamic changes in landslide conditions. Therefore, to address the lack of spatiotemporal information in historical landslide inventories and to enhance the prediction accuracy of LSA, we propose an LSA framework that integrates multimodal data. First, textual information (e.g., disaster investigation reports, news, government bulletins, and academic papers) was systematically processed using a fine-tuned BERT-BiLSTM-CRF model, from which 7695 landslide events and transformed into 28,472 spatiotemporal triples of landslides, each triple serves as a standardized knowledge unit, uniformly represented as (event, location, time), explicitly describing when and where a specific landslide occurred. Second, the spatiotemporal entities in these triples were used as anchor points to align with the remote sensing change detection algorithm on the Google Earth Engine platform, enabling the identification and delineation of 1745 historical landslides with both occurrence time and spatial extent. Finally, by integrating non-time series data such as topographic–geological factors with time-series meteorological and engineering parameters (e.g., precipitation, temperature, road and river distances), and applying the random forest, the time sensitivity of the LSA in the entire study area was evaluated. The results show that compared with the multi-year average parameters, the construction of time-series parameters improves the LSA prediction F1 score from 78.16 to 83.72%. Combined with the meteorological factors and artificial factors of the dynamic time window, it has the potential for landslide early warning. The proposed framework improves the spatiotemporal resolution of LSA, which provides a strong support for dynamic monitoring, risk assessment, and disaster prevention decision-making.