| 2 |
2026/07/14 17:24 |
李* |
期刊论文，论文题目“基于数据增强的滑坡预测研究”，类型：研究型论文，暂定摘要：Displacement forecasting of rainfall-triggered landslides is challenging due to the complex interaction between long-term progressive creep and irregular, event-driven fluctuations associated with transient hydrological forcing. To address the difficulty of forecasting under limited-sample conditions, this study proposes a decomposition, prediction, and reconstruction framework, evaluated at three monitoring sites (G1–G3) on the Luojiaping landslide in China. After identifying a structural displacement jump, the post-break displacement increment is separated into trend and fluctuation components using LOWESS smoothing. The trend component is predicted via Holt-type exponential smoothing, while the fluctuation component is modeled using gradient-boosting learners driven by antecedent rainfall and recent fluctuation history. To support fluctuation learning under limited training samples, an Attention-TCN-TimeGAN data augmentation strategy is exclusively applied to the fluctuation-term data. The trend prediction results showed that the trend component was highly predictable, achieving R² values of 0.9639, approximately 1.0000, and 0.9989 at G1, G2, and G3, respectively. At the reconstructed cumulative-displacement level, XGBoost-based GAN augmentation reduced the Root Mean Square Error (RMSE) across all sites, decreasing from 2.627 to 2.384 at G1, from 4.498 to 4.172 at G2, and from 3.326 to 3.161 at G3. The results indicate that while final cumulative displacement prediction is primarily controlled by the trend component, GAN-augmented fluctuation prediction provides useful local correction. This framework may be useful for monitored rainfall-triggered landslides characterized by structural displacement changes and irregular short-term fluctuations.
Keywords: rainfall-triggered landslide; displacement forecasting; LOWESS decomposition; fluctuation prediction; GAN-based data augmentation; gradient boosting
|
CSTR:11738.11.ncdc.Sanxia.2020.106
本作品采用
CC BY 4.0 （知识共享 署名 4.0 国际许可协议）进行许可。
10.12072/ncdc.sanxia.db7422.2026
