| Abstract (English): |
[Objective] Landslide-displacement prediction is critical when evaluating landslide stability. Despite the achievements of time-series methods based on deep-learning paradigms in predicting landslide displacement, the nonstationary, periodic, and trending characteristics of landslide displacement data cause multivariate predictions of current time-series models to easily overfit. Existing studies primarily focus on improving single models, whereas systematic studies pertaining to multimodel integration methods are scarce. This study aims to develop an integrated model that addresses these challenges and improves prediction accuracy. [Methods] Considering the volatility of landslidedisplacement data and the combined characteristics of their periodic and trending displacement components, a landslidedisplacement prediction model combining isolation forest (IF) anomaly detection, ensemble empirical mode decomposition (EEMD), convolutional neural networks (CNNs), and long short-term memory (LSTM) neural networks is proposed. The stepped Baijiabao landslide in the Three Gorges Reservoir area, which is affected by rainfall, is investigated in this study. First, the IF algorithm is introduced to detect anomalies in the original landslide-displacement data. This enables outliers, which can distort the prediction results, to be identified and excluded. Subsequently, EEMD is adopted to decompose the displacement data into intrinsic mode functions (IMFs), which represent the underlying periodic and trend components. This decomposition allows one to analyze the inherent characteristics of the data more comprehensively. Next, a CNN is employed to capture local periodic and trend patterns within the IMFs. CNNs are particularly effective in recognizing spatial patterns and features, thus rendering them suitable for identifying complex patterns in the displacement data. Finally, the overall displacement is predicted using the LSTM model, which is suitable for accommodating sequential data and capturing long-term dependencies. These techniques are integrated to leverage their respective strengths, thereby improving the prediction accuracy. [Result] The results indicate that the root-mean-square error（RMSE）, mean absolute error（MAE）, absolute percentage error in evaluation (MAPE), and determination coefficient（R2）indices of the EEMDCNN-LSTM model for predicting the overall landslide displacement under rainfall conditions are 0.4190, 0.3139, 0.2379, and 0.9997, respectively, which signify improvements in the accuracy of the first three evaluation indices by 32.3%, 25.1%, and 7.3%, respectively, compared with those of existing models. This significant improvement demonstrates the model's robustness in accommodating the complexities of landslide-displacement data under varying conditions. For predictions without rainfall, the RMSE, MAE, MAPE, and R2 indices are 0.4302, 0.2908, 0.2431, and 0.9996, respectively, which signify improvements in the accuracy of the first three indices by 31.2%, 31.7%, and 8.7%, respectively, compared with those of existing models. These results highlight the model's high generalizability across different scenarios, as it can maintain high prediction accuracies regardless of external influencing factors such as rainfall. Compared with conventional LSTM, random forest, and EEMD-LSTM models, the EEMD-CNN-LSTM model offers significant advantages under the influence of rainfall and without rainfall, thus significantly reducing overfitting and improving the prediction accuracy. The hybrid approach effectively captures the intricate patterns in the data, which cannot be achieved by single models. [Conclusion] In summary, the multimodel integration method based on IF anomaly detection, EEMD decomposition, local-feature capturing by the CNN, and overall prediction by LSTM significantly improves the accuracy of landslidedisplacement prediction, particularly under the influence of rainfall. The integrated model not only addresses the overfitting issues typically encountered in time-series prediction models but also enhances the model's robustness and reliability. The combination of IF for anomaly detection ensures that outliers do not skew the prediction results, whereas EEMD facilitates the decomposition of data into meaningful components. The CNN's ability to capture local patterns, coupled with the LSTM's strength in modeling long-term dependencies, enables the establishment of a comprehensive framework that can effectively accommodate the complexities of landslide displacement data.[Significance]This study provides an effective multimodel integration method for landslide-displacement prediction, which addresses the overfitting issues in existing models as well as offers substantial scientific significance and practical value. The proposed model's ability to accurately predict landslide displacement under varying conditions is extremely beneficial to the stability evaluation of landslideprone areas, thereby contributing to disaster prevention and mitigation efforts. The innovation is based on the systematic integration of anomaly detection, data decomposition, and advanced neural-network techniques, which results in a robust framework that outperforms conventional methods. The findings of this study are applicable to real-world scenarios, thereby enhancing the accuracy and reliability of landslide monitoring systems and supporting informed decision-making in hazard management and infrastructure development. [ABSTRACT FROM AUTHOR] |
| Abstract (Chinese): |
滑坡位移预测是滑坡稳定性评价的重要环节。尽管基于深度学习范式的时间序列方法预测滑坡 位移取得了一定的成果, 但由于滑坡位移数据的非平稳性、周期性和趋势性变化特征, 导致当前时间序 列模型的滑坡位移的多变量预测容易过拟合。为解决这一问题, 针对滑坡位移数据的波动性和由周期项 与趋势项位移叠加组成的特性, 提出一种基于孤立森林(Isolation Forest, IF) 异常检测、集成经验模态分解 (Ensemble Empirical Mode Decomposition, EEMD)、卷积神经网络（Convolutional Neural Networks, CNN）和长 短期记忆神经网络(Long Short-Term Memory, LSTM) 相结合的滑坡位移预测模型。选择三峡库区以降雨为 影响因子的阶跃型白家包滑坡为研究对象, 引入IF 算法对滑坡原始位移数据进行异常检测, 使用 EEMD 方法提取滑坡趋势项和周期项位移, 通过CNN 捕捉局部周期项和趋势模式, 并基于LSTM 模型预 测总体位移。结果表明, EEMD-CNN-LSTM 在预测降雨情况时滑坡总体位移的均方根误差（RMSE）、平 均绝对误差（MAE）、评价绝对百分比误差（MAPE）和决定系数（R²）4 种指标分别为0.4190、0.3139、 0.2379 和0.9997, 前3 种精度评价指标较现有模型分别提升32.3%、25.1%、7.3%。相较于传统的LSTM 模 型、随机森林方法和EEMD-LSTM 方法, EEMD-CNN-LSTM 模型在有、无降雨这一外部影响因素下具有显 著优势, 能够较大地降低过拟合, 提高预测的准确性。. [ABSTRACT FROM AUTHOR] |