Xia et al. (2025) A Spatiotemporal Pathformer‐Based Deep Learning Framework for Watershed Flood Forecasting

⚠️ Warning: This summary was generated from the abstract only, as the full text was not available.

Identification

• Journal: Water Resources Research
• Year: 2025
• Date: 2025-12-01
• Authors: Tianyu Xia, Yanlai Zhou, Chong‐Yu Xu, Pan Liu, Yuxuan Luo, Fi‐John Chang
• DOI: 10.1029/2025wr040193

Research Groups

Not specified in the abstract.

Short Summary

This study introduces a spatiotemporal Pathformer-based deep learning framework for multi-step-ahead flood forecasting, designed to dynamically adapt to flood magnitude and duration. The model demonstrates superior predictive accuracy and stability compared to LSTM and Transformer models, particularly during extreme flood events, by effectively mitigating time-lag errors and prediction bottlenecks.

Objective

• To develop and evaluate a spatiotemporal Pathformer-based deep learning framework for multi-step-ahead flood forecasting that dynamically adapts to flood magnitude and duration, aiming to overcome limitations of conventional artificial neural networks in capturing complex spatiotemporal dependencies and mitigating system biases and time-lag errors, especially during extreme flood events.

Study Configuration

• Spatial Scale: Jianxi basin
• Temporal Scale: 3-hour resolution for hydrometeorological records; 1- to 7-step forecast horizons.

Methodology and Data

• Models used: Spatiotemporal Pathformer (integrating a dual self-attention mechanism and adaptive path selection), Long Short-Term Memory (LSTM), and Transformer models for comparative analysis. SHapley Additive exPlanations (SHAP) analysis for model interpretability.
• Data sources: 25,341 hydrometeorological records from the Jianxi basin.

Main Results

• The spatiotemporal Pathformer demonstrated superior predictive accuracy and stability across 1- to 7-step forecast horizons compared to LSTM and Transformer models.
• It improved Nash-Sutcliffe Efficiency by 3.0% and 7.4% compared to LSTM and Transformer, respectively.
• Volume Efficiency increased by 3.4% and 9.6% compared to LSTM and Transformer, respectively.
• Root Mean Square Error (RMSE) was reduced by 18.4% and 34.9% compared to LSTM and Transformer, respectively.
• Mean Absolute Error (MAE) was lowered by 17.5% and 36.1% compared to LSTM and Transformer, respectively.
• The model effectively mitigated time-lag errors and prediction bottlenecks, ensuring robust and reliable forecasting, even during extreme flood events.
• SHapley Additive exPlanations analysis enhanced the model's interpretability by revealing key hydrometeorological drivers of its predictions.

Contributions

• Introduction of a novel spatiotemporal Pathformer-based deep learning framework for multi-step-ahead flood forecasting, incorporating a dual self-attention mechanism and adaptive path selection.
• Demonstrated significant improvements in predictive accuracy and stability over established deep learning models (LSTM and Transformer), particularly in handling complex spatiotemporal dependencies and extreme flood events.
• Enhanced model interpretability and trustworthiness through the application of SHapley Additive exPlanations analysis, revealing the underlying drivers of predictions.

Funding

Not specified in the abstract.

Citation

@article{Xia2025Spatiotemporal,
  author = {Xia, Tianyu and Zhou, Yanlai and Xu, Chong‐Yu and Liu, Pan and Luo, Yuxuan and Chang, Fi‐John},
  title = {A Spatiotemporal Pathformer‐Based Deep Learning Framework for Watershed Flood Forecasting},
  journal = {Water Resources Research},
  year = {2025},
  doi = {10.1029/2025wr040193},
  url = {https://doi.org/10.1029/2025wr040193}
}