Zhang et al. (2025) Runoff simulation based on landscape pattern classification and machine learning

Identification

• Journal: Journal of Hydrology Regional Studies
• Year: 2025
• Date: 2025-12-01
• Authors: Xueli Zhang, Lingan Kong, Tianning Xie, Miao Gan, Caihong Hu
• DOI: 10.1016/j.ejrh.2025.102968

Research Groups

• College of Water Conservancy and Transportation, Zhengzhou University, Zhengzhou 450001, China

Short Summary

This study developed a transferable classification-coupling framework integrating landscape structure and machine learning to improve runoff simulation accuracy and stability in the Middle Yellow River Basin. It demonstrated that coupling landscape patterns with meteorological variables significantly enhances model performance, with XGBoost achieving the highest accuracy (e.g., NSE = 0.966, NRMSE = 0.037) and strong generalization to ungauged basins.

Objective

• To examine how integrated landscape structure affects runoff across multiple scales under climate and human pressures.
• To establish a transferable classification–coupling framework integrating landscape structure into hydrological modeling, providing new insights for runoff prediction and watershed management.

Study Configuration

• Spatial Scale: 37 subbasins within the Middle Yellow River Basin, China, covering a total area of 345,000 square kilometers. The subbasins were classified into small- to medium-sized basins (S-MB) and large basins (LB).
• Temporal Scale: Annual data from 2000 to 2020.

Methodology and Data

• Models used:
  • Clustering: AutoClass, k-means (for comparison).
  • Runoff Simulation: Partial Least Squares Regression (PLS), Random Forest (RF), Extreme Gradient Boosting (XGBoost).
  • Classification Transferability: Long Short-Term Memory (LSTM) neural network.
  • Model Interpretation: SHAP (SHapley Additive exPlanations) algorithm.
• Data sources:
  • Topography: 30 meter DEM data from China Geospatial Data Cloud (elevation, slope, aspect).
  • Soil: Soil texture data from the National Earth System Science Data Center (0–200 cm depth).
  • Land Cover/Use & Vegetation: Integrated into 1 × 1 km grid framework.
  • Runoff: Daily runoff data from 37 hydrological stations, aggregated to annual totals, from the National Earth System Science Data Center and the Yellow River Hydrological Yearbook.
  • Precipitation: 2000–2020 data from 93 meteorological stations of the National Meteorological Information Center, interpolated to a 1 × 1 km grid.
  • Meteorological Variables: Daily average, minimum, and maximum temperature, sunshine duration, relative humidity, and wind speed from the National Qinghai–Tibet Plateau Scientific Data Center.
  • Landscape Metrics: 19 metrics including Largest Patch Index (LPI), Mean Euclidean Nearest-Neighbor Distance (ENN-MN), Landscape Shape Index (LSI), Splitting Index (SPLIT), Interspersion and Juxtaposition Index (IJI), Patch Richness Density (PRD), Shannon’s Diversity Index (SHDI), and Patch Richness (PR).

Main Results

• Landscape Pattern Classification: AutoClass clustered the 37 subbasins into two optimal landscape-pattern clusters: C1 (small- to medium-sized basins, S-MB) characterized by grassland dominance, simpler patch shapes, low richness, and weak fragmentation; and C2 (large basins, LB) dominated by farmland and forest, showing higher patch richness, complex shapes, stronger fragmentation, and lower connectivity.
• Improved Simulation Accuracy: Coupling landscape pattern indicators with meteorological variables significantly improved runoff simulation accuracy compared to models using only meteorological factors.
• Superior Model Performance: Among the tested algorithms, XGBoost consistently achieved the highest accuracy and robustness across all scenarios (cross-scale, S-MB, LB).
  • Cross-scale coupled model: XGBoost (NSE = 0.964, NRMSE = 0.027).
  • S-MB coupled model: XGBoost (NSE = 0.966, NRMSE = 0.037).
  • LB coupled model: XGBoost (NSE = 0.820, NRMSE = 0.075).
• Classification-based Modeling Benefits: Classification-based modeling (using AutoClass clusters) reduced input heterogeneity and improved model stability and accuracy compared to cross-scale models, effectively capturing landscape-specific runoff responses.
• Transferability to Ungauged Basins: An LSTM-based classification model achieved 100% test accuracy and successfully classified an ungauged basin (Quanyanshan Basin) as C1. Applying the corresponding C1_XGBoost model resulted in high simulation accuracy (NSE = 0.94, NRMSE = 0.11), demonstrating strong spatial transferability.
• Dominant Drivers: SHAP analysis revealed distinct dominant drivers for runoff between landscape types:
  • S-MB basins: LSI, PR, IJI, Total Summer Precipitation (TSuP), PRD, Annual Precipitation (PRE), and SPLIT.
  • LB basins: LSI, Grassland (GRA) proportion, TSuP, and PRE.

Contributions

• Establishes a novel and transferable classification–coupling framework that integrates landscape structure clustering (using AutoClass and LSTM) with advanced machine learning algorithms (XGBoost) for improved runoff simulation.
• Demonstrates that incorporating comprehensive landscape pattern characteristics significantly enhances the accuracy and robustness of runoff models compared to meteorology-only or traditional models.
• Provides a robust and generalizable methodology for runoff prediction in data-scarce or ungauged basins through the spatial transferability of the classification model.
• Offers new insights into the structural–functional differentiation of runoff generation by identifying distinct dominant landscape and climatic drivers for different basin types via SHAP analysis.
• Proposes targeted, landscape-type-specific recommendations for soil and water resource management in heterogeneous basins, such as maintaining connectivity in S-MB basins and improving vegetation structure in LB basins.

Funding

• Ministry of Science and Technology of the People’s Republic of China: National key research and development program (2023YFC3209303)
• China Scholarship Council
• Qian Kehe Zhicheng [2023] Yiban 206
• Qian Kehe Zhicheng [2024] Yiban 130
• The Science found for distinguished young scholars of Henan Province (232300421017)

Citation

@article{Zhang2025Runoff,
  author = {Zhang, Xueli and Kong, Lingan and Xie, Tianning and Gan, Miao and Hu, Caihong},
  title = {Runoff simulation based on landscape pattern classification and machine learning},
  journal = {Journal of Hydrology Regional Studies},
  year = {2025},
  doi = {10.1016/j.ejrh.2025.102968},
  url = {https://doi.org/10.1016/j.ejrh.2025.102968}
}