Scholarlink

<p>Rainfall-induced landslide early warning systems are commonly based on empirical rainfall thresholds that do not explicitly represent the underlying hydro-mechanical processes governing slope failure. This limitation reduces their reliability, particularly under complex and nonlinear rainfall conditions. This study presents a hybrid computational framework that integrates physically based modeling, probabilistic analysis, and machine learning to derive robust and transferable rainfall thresholds. A set of rainfall scenarios was systematically generated by varying rainfall intensity, duration, temporal patterns, and return periods, including cumulative rainfall over 3-, 5-, and 7-day windows. For each scenario, coupled seepage–slope stability simulations were performed to compute spatial distributions of the factor of safety (F.S.). These results were transformed into a spatial probability of failure (P.F.) using a spatial aggregation approach, enabling probabilistic characterization of slope instability. The resulting dataset was mapped into a rainfall parameter space to construct a physics-informed Critical Rainfall Envelope (CRE), which defines rainfall thresholds based on probabilistic contours and historical landslide records. The results demonstrate that ensemble machine learning models were developed as surrogate predictors of P.F. from rainfall descriptors, particularly XGBoost, which were developed to efficiently predict P.F., showing strong agreement with physically based results. The results indicate that cumulative rainfall and temporal evolution are dominant controls on slope instability, and that the 5-day rainfall duration provides the most stable and representative threshold condition. The proposed framework establishes a direct link between hydro-mechanical processes, probabilistic failure, and operational rainfall thresholds, providing a robust, computationally efficient solution for landslide early warning in data-limited regions.</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

A hybrid physics–machine learning framework for probabilistic landslide prediction based on critical rainfall envelopes

  • 期刊论文

  • 发表日期：

    2026年7月17日

查看全文
  • Thawatchai Chalermpornchai,
  • Wen-Yi Hung,
  • Katavut Vichai,
  • Bunpoat Kunsuwan,
  • Warakorn Mairaing,
  • Suraparb Keawsawasvong

摘要

Rainfall-induced landslide early warning systems are commonly based on empirical rainfall thresholds that do not explicitly represent the underlying hydro-mechanical processes governing slope failure. This limitation reduces their reliability, particularly under complex and nonlinear rainfall conditions. This study presents a hybrid computational framework that integrates physically based modeling, probabilistic analysis, and machine learning to derive robust and transferable rainfall thresholds. A set of rainfall scenarios was systematically generated by varying rainfall intensity, duration, temporal patterns, and return periods, including cumulative rainfall over 3-, 5-, and 7-day windows. For each scenario, coupled seepage–slope stability simulations were performed to compute spatial distributions of the factor of safety (F.S.). These results were transformed into a spatial probability of failure (P.F.) using a spatial aggregation approach, enabling probabilistic characterization of slope instability. The resulting dataset was mapped into a rainfall parameter space to construct a physics-informed Critical Rainfall Envelope (CRE), which defines rainfall thresholds based on probabilistic contours and historical landslide records. The results demonstrate that ensemble machine learning models were developed as surrogate predictors of P.F. from rainfall descriptors, particularly XGBoost, which were developed to efficiently predict P.F., showing strong agreement with physically based results. The results indicate that cumulative rainfall and temporal evolution are dominant controls on slope instability, and that the 5-day rainfall duration provides the most stable and representative threshold condition. The proposed framework establishes a direct link between hydro-mechanical processes, probabilistic failure, and operational rainfall thresholds, providing a robust, computationally efficient solution for landslide early warning in data-limited regions.