Scholarlink

<p>Rainfall-induced landslides frequently exhibit step-like, nonlinear, and delayed displacement patterns. However, in many settings, only rainfall and surface displacement records are available because of limited subsurface monitoring capabilities, making accurate forecasting challenging to achieve. This study proposes the Time-Lag and Lead-Follow Analysis-Based Hybrid Dynamic Landslide Cumulative Displacement Prediction (TL-LF-HLDP) framework, which integrates multiscale decomposition, lag-aware feature selection, and deep time-varying regression. Displacement and hydrometeorological time series are first decomposed into trend and periodic components using the Gray Wolf Optimizer tuned Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise(GWO–ICEEMDAN). Time-lagged cross-correlation (TLCC) is then applied to capture delayed relationships between rainfall and slope responses. A dynamic lagged feature selection (DLFS) module further refines the input set, which is subsequently passed to a deep, sparse, time-varying autoregressive network (DDS-TVAR) for prediction. The framework is validated on two real-world case studies: the rainfall-induced Hehaotun landslide in Guangxi, southern China, and the reservoir-driven Baijiabao landslide. Across six stations concerning these two landslides, TL-LF-HLDP reduces the MAE and RMSE by 38% and 32%, respectively, on average, when computed as stationwise means relative to the best competing baseline at each station. The results underscore the potential of the framework for providing early landslide warnings, particularly in data-scarce environments.</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 multiscale forecasting framework for rainfall-induced landslide displacement: case studies from Guangxi, China

  • 期刊论文

  • 发表日期：

    2026年3月2日

查看全文
  • Fan Zhang,
  • Yuanfa Ji,
  • Xiaoming Liu,
  • Siyuan Liu,
  • Xiyan Sun

摘要

Rainfall-induced landslides frequently exhibit step-like, nonlinear, and delayed displacement patterns. However, in many settings, only rainfall and surface displacement records are available because of limited subsurface monitoring capabilities, making accurate forecasting challenging to achieve. This study proposes the Time-Lag and Lead-Follow Analysis-Based Hybrid Dynamic Landslide Cumulative Displacement Prediction (TL-LF-HLDP) framework, which integrates multiscale decomposition, lag-aware feature selection, and deep time-varying regression. Displacement and hydrometeorological time series are first decomposed into trend and periodic components using the Gray Wolf Optimizer tuned Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise(GWO–ICEEMDAN). Time-lagged cross-correlation (TLCC) is then applied to capture delayed relationships between rainfall and slope responses. A dynamic lagged feature selection (DLFS) module further refines the input set, which is subsequently passed to a deep, sparse, time-varying autoregressive network (DDS-TVAR) for prediction. The framework is validated on two real-world case studies: the rainfall-induced Hehaotun landslide in Guangxi, southern China, and the reservoir-driven Baijiabao landslide. Across six stations concerning these two landslides, TL-LF-HLDP reduces the MAE and RMSE by 38% and 32%, respectively, on average, when computed as stationwise means relative to the best competing baseline at each station. The results underscore the potential of the framework for providing early landslide warnings, particularly in data-scarce environments.