Scholarlink

<p>Prediction of landslide displacements during creep stages is an important issue for landslide hazard and risk management and for establishing early warning systems to reduce loss of human life and economic damage. Predictions of landslide displacements were usually based on monitoring results of different elements that impact landslide displacements, such as precipitation, groundwater level, water level fluctuations at the toe of landslides, as well as existing displacement rates, using phenomenological approaches and statistical methods. In this study, we proposed a novel hybrid modelling framework that explicitly incorporates multi-time-scale components (MTSCs) of rainfall and reservoir water level into a hybrid convolutional neural network–long short-term memory (CNN-LSTM) model optimized by swarm intelligence algorithms, based on monitoring data at the Taping landslide, located near Quchi Town, Wushan County, Chongqing, China. The Taping landslide is a huge prehistoric landslide reactivated after the establishment of the Three Gorges Project and monitored by a GNSS system. The time series values of rainfall and reservoir water levels are decomposed using complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) to obtain intrinsic mode functions (IMFs). Fast Fourier transform (FFT) spectral analysis is applied to each IMF to identify dominant frequencies and convert them into corresponding time periods. IMFs are then classified into four time-scale ranges: 0–15&#xa0;days, 15–30&#xa0;days, 30–60&#xa0;days, and &gt; 60&#xa0;days. IMFs within the same range are reconstructed into MTSCs that represent the temporal variability of rainfall and reservoir water levels. The landslide displacement series is decomposed using an EMD-based method and statistically analyzed using <i>t</i>-tests to separate high-frequency noise, cyclic components, and long-term trends. These refined components serve as inputs to a CNN-LSTM model, whose hyperparameters are optimized using six swarm intelligence algorithms. It was found that the proposed framework has high predictive accuracy and effectively captures both short- and long-term landslide behavior and can be used in landslide early warning systems.</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

Prediction of enhanced creep landslide displacement by analyzing multi-time-scale displacement impact factors using convolutional neural network (CNN)

  • 期刊论文

  • 发表日期：

    2026年1月6日

查看全文
  • Zhongdi Rong,
  • Shengyi Cong,
  • Liang Tang,
  • Pengfei Chen,
  • Yonghui Wang,
  • Meihong Ma,
  • Željko Arbanas

摘要

Prediction of landslide displacements during creep stages is an important issue for landslide hazard and risk management and for establishing early warning systems to reduce loss of human life and economic damage. Predictions of landslide displacements were usually based on monitoring results of different elements that impact landslide displacements, such as precipitation, groundwater level, water level fluctuations at the toe of landslides, as well as existing displacement rates, using phenomenological approaches and statistical methods. In this study, we proposed a novel hybrid modelling framework that explicitly incorporates multi-time-scale components (MTSCs) of rainfall and reservoir water level into a hybrid convolutional neural network–long short-term memory (CNN-LSTM) model optimized by swarm intelligence algorithms, based on monitoring data at the Taping landslide, located near Quchi Town, Wushan County, Chongqing, China. The Taping landslide is a huge prehistoric landslide reactivated after the establishment of the Three Gorges Project and monitored by a GNSS system. The time series values of rainfall and reservoir water levels are decomposed using complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) to obtain intrinsic mode functions (IMFs). Fast Fourier transform (FFT) spectral analysis is applied to each IMF to identify dominant frequencies and convert them into corresponding time periods. IMFs are then classified into four time-scale ranges: 0–15 days, 15–30 days, 30–60 days, and > 60 days. IMFs within the same range are reconstructed into MTSCs that represent the temporal variability of rainfall and reservoir water levels. The landslide displacement series is decomposed using an EMD-based method and statistically analyzed using t-tests to separate high-frequency noise, cyclic components, and long-term trends. These refined components serve as inputs to a CNN-LSTM model, whose hyperparameters are optimized using six swarm intelligence algorithms. It was found that the proposed framework has high predictive accuracy and effectively captures both short- and long-term landslide behavior and can be used in landslide early warning systems.