Scholarlink

<p>Under the combined influences of global extreme climate events, agricultural expansion, and industrial development, the North Henan Plain, China (NHP) has experienced long-term and pronounced land deformation. This study proposes a two-stage “mechanism identification–predictive modeling” framework to investigate the geohazard issues. First, Multi-temporal synthetic aperture radar interferometry (MT-InSAR) was utilized to extract long-term land displacement time-series and mean rates for 2017–2022 in NHP; subsequently, machine learning models (XGBoost and RF) combined with explainable SHapley Additive exPlanations (SHAP) were employed to quantify the dominant driving factors that acted on the land deformation in NHP, revealing that groundwater storage changes (GWSC), annual precipitation variations, and road density are the primary contributors, with GWSC exerting a particularly pronounced influence in subsidence zones. On this basis, we developed an LSTM model (CMI-LSTM) that integrates dynamic temporal features (60 epochs of MT-InSAR-derived displacement and GWSC), static factors (soil type, overburden thickness, road density), and spatial location information for training and prediction of the deformation across the NHP. Results demonstrate that CMI-LSTM achieves markedly higher predictive accuracy than both an LSTM model based only on displacement time series without additional factors and a CMI-RF model (same multi-inputs as the CMI-LSTM), yielding an overall RMSE of 9.6&#xa0;mm, MAE of 7.0&#xa0;mm, and R² of 0.9836, with 76.9% of predictions falling within ± 10&#xa0;mm. This framework effectively captures the coupled control of hydrogeological and anthropogenic factors on land deformation, significantly enhancing short- and medium-term predictive capabilities, and offers a transferable methodological pathway for subsidence hazard assessment and groundwater management in groundwater-dependent agricultural regions worldwide.</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

Integrating Explainable Machine Learning and LSTM with Multiple Factors for Mechanism Identification and Prediction of Land Deformation in the North Henan Plain

  • 期刊论文

  • 发表日期：

    2026年1月16日

查看全文
  • Jiabei Wang,
  • Jiyuan Hu,
  • Zheng Zhou,
  • Jiayao Wang,
  • Jie Meng,
  • Deng Pan,
  • Pengyu Li,
  • Lijun Wang,
  • Jianhui Xiang,
  • Jingru Ma

摘要

Under the combined influences of global extreme climate events, agricultural expansion, and industrial development, the North Henan Plain, China (NHP) has experienced long-term and pronounced land deformation. This study proposes a two-stage “mechanism identification–predictive modeling” framework to investigate the geohazard issues. First, Multi-temporal synthetic aperture radar interferometry (MT-InSAR) was utilized to extract long-term land displacement time-series and mean rates for 2017–2022 in NHP; subsequently, machine learning models (XGBoost and RF) combined with explainable SHapley Additive exPlanations (SHAP) were employed to quantify the dominant driving factors that acted on the land deformation in NHP, revealing that groundwater storage changes (GWSC), annual precipitation variations, and road density are the primary contributors, with GWSC exerting a particularly pronounced influence in subsidence zones. On this basis, we developed an LSTM model (CMI-LSTM) that integrates dynamic temporal features (60 epochs of MT-InSAR-derived displacement and GWSC), static factors (soil type, overburden thickness, road density), and spatial location information for training and prediction of the deformation across the NHP. Results demonstrate that CMI-LSTM achieves markedly higher predictive accuracy than both an LSTM model based only on displacement time series without additional factors and a CMI-RF model (same multi-inputs as the CMI-LSTM), yielding an overall RMSE of 9.6 mm, MAE of 7.0 mm, and R² of 0.9836, with 76.9% of predictions falling within ± 10 mm. This framework effectively captures the coupled control of hydrogeological and anthropogenic factors on land deformation, significantly enhancing short- and medium-term predictive capabilities, and offers a transferable methodological pathway for subsidence hazard assessment and groundwater management in groundwater-dependent agricultural regions worldwide.