Wang et al. (2026) Coupling strategies of snowmelt runoff model and machine learning in the Lhasa River Basin

Identification

• Journal: Journal of Hydrology Regional Studies
• Year: 2026
• Date: 2026-03-30
• Authors: Tao Wang, Dingzhi Peng, Yuwei Gong, Xingtong Chen, Depeng Zuo
• DOI: 10.1016/j.ejrh.2026.103400

Research Groups

• College of Water Sciences, Beijing Normal University, Beijing, China

Short Summary

This study developed two coupling strategies between the Snowmelt Runoff Model (SRM) and the Transformer machine learning model to enhance streamflow simulation accuracy and interpretability in the Lhasa River Basin. The residual coupling strategy significantly improved simulation accuracy (NSE: 0.97), and SHAP analysis identified precipitation as the main driving factor, with temperature, solar radiation, relative humidity, and land use types exhibiting complex, threshold-dependent influences on streamflow.

Objective

• To enhance the accuracy and interpretability of streamflow simulation in high-altitude regions by proposing a coupled framework of the Snowmelt Runoff Model (SRM) and Machine Learning (ML), specifically the Transformer model.
• To develop and evaluate two coupling strategies: one using SRM residuals as Transformer features for error correction, and another using SRM's intermediate runoff components as input features for physical constraints.
• To understand the driving mechanisms of streamflow by analyzing the comprehensive impacts of meteorological factors and land use on streamflow in the Lhasa River Basin (LRB) using SHapley Additive exPlanations (SHAP).

Study Configuration

• Spatial Scale: Lhasa River Basin (LRB), approximately 3.25 × 10⁴ km², located in the central-southern part of the Qinghai-Tibet Plateau, with elevations between 4000 m and 5500 m.
• Temporal Scale: Daily streamflow simulation and analysis from 2003 to 2015.

Methodology and Data

• Models used:
  • Snowmelt Runoff Model (SRM): A physically-based hydrological model using the degree-day method, optimized with Differential Evolution (DE) and L-BFGS-B algorithms.
  • Transformer: A neural network architecture based on a self-attention mechanism for sequential data processing.
  • Long Short-Term Memory (LSTM): A recurrent neural network used as a benchmark.
  • SHapley Additive exPlanations (SHAP): A game-theoretic model interpretability method to quantify feature contributions.
• Data sources:
  • Observed daily streamflow data (2003–2015) from the Tibet Bureau of Hydrology.
  • Daily temperature data at Lhasa from the Tibet Meteorological Bureau.
  • 0.1° daily gridded precipitation CHM_PRE V2 dataset.
  • MODIS snow cover data MOD10A2 (500 m resolution).
  • 30 m China Land Cover Dataset (CLCD) for land use.

Main Results

• The residual coupling strategy (SRM residuals + Transformer) achieved the highest streamflow simulation accuracy, with Nash-Sutcliffe Efficiency (NSE) of 0.97, R² of 0.97, and Kling-Gupta Efficiency (KGE) of 0.95 on the testing set, significantly outperforming standalone SRM (NSE: 0.68) and slightly surpassing standalone Transformer (NSE: 0.96) and LSTM (NSE: 0.94).
• SHAP analysis identified precipitation as the most important driving factor for streamflow in the Lhasa River Basin.
• Temperature showed a significant positive promoting effect on streamflow when exceeding 11.2 °C.
• Solar radiation exhibited a pronounced positive contribution when it exceeded 245.69 W/m².
• Relative humidity showed a similar positive contribution when above 45%.
• Among land use types, barren land and snow cover contributed positively to streamflow when their proportions were high (bare land threshold ~10%, snow cover threshold ~26%), while wetlands showed inhibitory effects.
• Land use changes from 2003 to 2015 included a significant decrease in glacier area (-18%), expansion of water bodies (+60%) and wetlands (+201%), and a sharp decline in farmland (-93%), reflecting climate change impacts and ecological restoration efforts.
• Complex nonlinear interactions and threshold dependencies were observed between climatic factors and land use types, influencing streamflow.

Contributions

• Proposed and validated a novel hybrid modeling framework integrating the physical Snowmelt Runoff Model (SRM) with the Transformer machine learning architecture for enhanced streamflow prediction in high-altitude cold regions.
• Demonstrated that coupling strategies, particularly the residual coupling, significantly improve simulation accuracy compared to standalone physical or machine learning models, addressing challenges in data-scarce and complex hydrological environments.
• Provided an interpretable analysis of streamflow drivers in the Lhasa River Basin using SHAP, quantifying the contributions and interaction mechanisms of meteorological factors (precipitation, temperature, solar radiation, relative humidity) and land use types (barren, wetland, grassland, snow cover).
• Revealed critical thresholds for meteorological factors (e.g., temperature > 11.2 °C, solar radiation > 245.69 W/m², relative humidity > 45%) and land cover proportions (e.g., barren > 10%, snow cover > 26%) that significantly influence streamflow.
• Offered a valuable tool and insights for regional water resource management, flood control, and ecological protection in the context of climate change and land use dynamics in the Qinghai-Tibet Plateau.

Funding

• National Science and Technology Major Project (2025ZD1204403)

Citation

@article{Wang2026Coupling,
  author = {Wang, Tao and Peng, Dingzhi and Gong, Yuwei and Chen, Xingtong and Zuo, Depeng},
  title = {Coupling strategies of snowmelt runoff model and machine learning in the Lhasa River Basin},
  journal = {Journal of Hydrology Regional Studies},
  year = {2026},
  doi = {10.1016/j.ejrh.2026.103400},
  url = {https://doi.org/10.1016/j.ejrh.2026.103400}
}