北京大学学报（自然科学版） ›› 2026, Vol. 62 ›› Issue (2): 359-366.DOI: 10.13209/j.0479-8023.2025.030

上一篇    下一篇

基于显隐式特征交互的滑坡短期位移预测

田原1,2, 马睿平1,2, 赵文祎3,4,5,†, 张建学1,2, 黄儒豪1,2, 汪翰林1,2, 白宇丹1,2   

  1. 1. 北京大学遥感与地理信息系统研究所, 北京 100871 2. 空间信息集成与3S工程应用北京市重点实验室, 北京 100871 3. 中国地质环境监测院, 北京 100081 4. 自然资源部地质灾害智能监测与风险预警工程技术创新中心, 北京 100081 5. 中国地质大学(北京)信息工程学院, 北京 100083
  • 收稿日期:2025-03-05 修回日期:2025-04-17 出版日期:2026-03-20 发布日期:2026-03-20
  • 基金资助:
    中国地质调查局地质调查项目(DD20230085)和中央高校基本科研业务费资助

Landslide Short-Term Displacement Prediction Based on Explicit and Implicit Feature Interactions

TIAN Yuan1,2, MA Ruiping1,2, ZHAO Wenyi3,4,5,†, ZHANG Jianxue1,2, HUANG Ruhao1,2, WANG Hanlin1,2, BAI Yudan1,2   

  1. 1. Institute of Remote Sensing and Geographical Information Systems, Peking University, Beijing 100871 2. Beijing Key Laboratory of Spatial Information Integration and Its Applications, Beijing 100871 3. China Institute of Geo-Environment Monitoring, Beijing 100081 4. Technology Innovation Center for Geohazard Monitoring and Risk Early Warning, Ministry of Natural Resources, Beijing 100081 5. School of Information Engineering, China University of Geoscience, Beijing 100083
  • Received:2025-03-05 Revised:2025-04-17 Online:2026-03-20 Published:2026-03-20

PDF

摘要：

针对当前采用机器学习或深度学习模型预测滑坡短期位移的方法难以同时保障预测模型的泛化能力和记忆能力且可解释性欠佳的问题, 从显隐式特征交互视角出发, 设计具有一定可解释性的显式特征交互网络(IFIN), 构建融合显隐式特征交互的滑坡位移预测模型(EIFIM)。EIFIM可以迁移学习其他坡面变形规律, 并基于预测坡面的动静态因子预测坡面未来三日的位移。实例验证结果表明, EIFIM的预测效果优于基线模型。同时, 模型输出的可解释特征组合表明模型具有较好的架构合理性。

关键词: 显隐式特征交互, 时间序列, 可解释性, 滑坡位移, 短期预测

Abstract:

Most existing approaches for short-term landslide displacement prediction apply machine learning or deep learning models, which are unable to ensure both excellent generalization and memory capabilities and have limited interpretability. From the perspective of explicit and implicit feature interaction, this paper proposes an Interpretable Explicit Feature Interaction Network (IFIN) and constructs the Explicit and Implicit Feature Interaction-Model (EIFIM) for landslide displacement prediction. Based on the transfer learning method, EIFIM can be trained on deformation pattern dataset including many slopes and then applied to a new single slope to predict its displacement of the next three days based on both static and dynamic factors. Case studies show that prediction performance of EIFIM outperforms common baselines. Moreover, the explainable feature combinations output by the proposed model also indicate its good architectural rationality.

Key words: explicit and implicit feature intersection, time series, interpretability, landslide displacement, short-term prediction

版权所有 © 《北京大学学报（自然科学版）》编辑部
技术支持：北京玛格泰克科技发展有限公司