Scholarlink

<p>Accurate riverbed topography reconstruction is essential for channel hydraulics, sediment transport analysis, and terrain pre-processing in two-dimensional hydrodynamic simulations. To improve pointwise riverbed elevation interpolation under sparse measured cross-sections, this study proposes a heterogeneous ensemble framework, RF-XGB-ResDNN, that integrates Random Forest (RF), eXtreme Gradient Boosting (XGBoost), and an improved residual deep neural network (ResDNN). A dedicated data preparation workflow was developed, including interpolation-point generation, river-area grid processing, and dataset construction, through which 12 input variables were designed to describe the spatial position and terrain constraints of each interpolation point. The three base learners were independently trained, and their outputs were fused using a grid search-based weighted soft voting strategy. The framework was validated in the Grand Canal reach in Cangzhou, China, and further tested for transferability in the Kedu River. Results showed that RF-XGB-ResDNN achieved the best performance among all compared models. On the main study area, it outperformed conventional interpolation methods, reducing MAE and RMSE by 68.2%–81.5% and 70.6%–82.3%, respectively, and reached an R² of 0.953. In the transferability test, the ensemble model also achieved the highest accuracy, with MAE = 0.181&#xa0;m, RMSE = 0.245&#xa0;m, and R² = 0.927. The results indicate that the proposed framework can effectively combine the nonlinear representation ability of deep learning with the robustness of tree-based models, providing a reliable approach for riverbed topography reconstruction in hydraulic applications.</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

A Novel Deep Learning Ensemble Framework for Riverbed Topography Reconstruction in 2D Hydrodynamic Simulations

  • 期刊论文

  • 发表日期：

    2026年6月3日

查看全文
  • Qiang Liu,
  • Jiaxin Zheng,
  • Chuanxing Zheng,
  • Fengjiao Zhao,
  • Lei Yu,
  • Feng Ling,
  • Zhixiang Da,
  • Jijian Lian

摘要

Accurate riverbed topography reconstruction is essential for channel hydraulics, sediment transport analysis, and terrain pre-processing in two-dimensional hydrodynamic simulations. To improve pointwise riverbed elevation interpolation under sparse measured cross-sections, this study proposes a heterogeneous ensemble framework, RF-XGB-ResDNN, that integrates Random Forest (RF), eXtreme Gradient Boosting (XGBoost), and an improved residual deep neural network (ResDNN). A dedicated data preparation workflow was developed, including interpolation-point generation, river-area grid processing, and dataset construction, through which 12 input variables were designed to describe the spatial position and terrain constraints of each interpolation point. The three base learners were independently trained, and their outputs were fused using a grid search-based weighted soft voting strategy. The framework was validated in the Grand Canal reach in Cangzhou, China, and further tested for transferability in the Kedu River. Results showed that RF-XGB-ResDNN achieved the best performance among all compared models. On the main study area, it outperformed conventional interpolation methods, reducing MAE and RMSE by 68.2%–81.5% and 70.6%–82.3%, respectively, and reached an R² of 0.953. In the transferability test, the ensemble model also achieved the highest accuracy, with MAE = 0.181 m, RMSE = 0.245 m, and R² = 0.927. The results indicate that the proposed framework can effectively combine the nonlinear representation ability of deep learning with the robustness of tree-based models, providing a reliable approach for riverbed topography reconstruction in hydraulic applications.