Zhu et al. (2026) HieraBoost-Q: interpretable karst discharge prediction from multi-site electrical conductivity with SHAP-based mechanism insights

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-02-19
• Authors: Yinxia Zhu, Jie Niu, Qingmin Zhu, Fen Huang, Han Qiu, Dongdong Liu, Pan Wu, Bill X. Hu
• DOI: 10.1016/j.jhydrol.2026.135153

Research Groups

• School of Water Conservancy and Environment, University of Jinan, Jinan 250022, PR China
• College of Resources and Environmental Engineering, Guizhou University, PR China
• Key Laboratory of Karst Georesources and Environment, Ministry of Education, Guiyang 550025, PR China
• Key Laboratory of Karst Dynamics, MNR and GZAR, Institute of Karst Geology, Chinese Academy of Geological Sciences/International Research Center on Karst under the Auspices of UNESCO/National Center for International Research on Karst Dynamic System and Global Change, Guilin 541004, PR China
• Pingguo Guangxi, Karst Ecosystem, National Observation and Research Station, Pingguo 531406, Guangxi, PR China
• Department of Sustainable Earth Systems Sciences, University of Texas at Dallas, Richardson, TX 75080, USA

Short Summary

This study proposes HieraBoost-Q, an interpretable hybrid framework integrating multi-site electrical conductivity and hierarchical XGBoost with SHAP, to improve discharge prediction and elucidate recharge mechanisms in karst catchments. It significantly enhances prediction accuracy and provides insights into the spatiotemporal dynamics of karst recharge.

Objective

• To develop an interpretable and bias-aware hybrid framework (HieraBoost-Q) for improved discharge prediction in complex karst catchments.
• To elucidate karst recharge mechanisms using SHAP-based interpretation of multi-site electrical conductivity data.

Study Configuration

• Spatial Scale: A karst catchment in Southwest China.
• Temporal Scale: Captures instantaneous (<1 hour) and delayed (10–12 hours) hydrological signals for event-scale prediction.

Methodology and Data

• Models used: HieraBoost-Q (integrates hierarchical XGBoost modeling and SHAP-based interpretation), four-stage bias-correction chain.
• Data sources: Multi-site electrical conductivity (EC), rainfall (implied as a key influencing factor).

Main Results

• HieraBoost-Q substantially improved discharge prediction performance over raw XGBoost, reducing RMSE from 9.767 L s⁻¹ to 3.603 L s⁻¹ and increasing R² from 0.907 to 0.987.
• The framework robustly captured delayed and weak signals, particularly under low-flow conditions.
• SHAP analysis revealed that upstream boreholes with strong hydraulic connectivity are dominant predictors.
• Rainfall exerts a nonlinear, state-dependent modulation on EC contributions, forming a dual response structure with instantaneous (<1 hour) and delayed (10–12 hours) signals.
• Interaction effects showed that rainfall intensity can trigger reversals in feature contributions across event stages.
• Multi-site EC provides earlier and more stable event-scale precursors compared to rainfall alone.

Contributions

• Introduces HieraBoost-Q, an interpretable and bias-aware hybrid framework for enhanced karst discharge prediction.
• Demonstrates the effectiveness of multi-site electrical conductivity as low-cost, scalable surrogate indicators for capturing both rapid and slow karst recharge dynamics.
• Provides SHAP-based insights into complex karst recharge mechanisms, including the dominance of upstream hydraulic connectivity and the dual, state-dependent response of rainfall.
• Offers a practical tool for flood warning, water resources management, and advanced karst hydrological studies.

Funding

• Not specified in the provided text.

Citation

@article{Zhu2026HieraBoostQ,
  author = {Zhu, Yinxia and Niu, Jie and Zhu, Qingmin and Huang, Fen and Qiu, Han and Liu, Dongdong and Wu, Pan and Hu, Bill X.},
  title = {HieraBoost-Q: interpretable karst discharge prediction from multi-site electrical conductivity with SHAP-based mechanism insights},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.135153},
  url = {https://doi.org/10.1016/j.jhydrol.2026.135153}
}