Liu et al. (2026) A Graph‐Based Deep Learning Approach for Daily Flash Flood Susceptibility Modeling in China

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

Identification

• Journal: Water Resources Research
• Year: 2026
• Date: 2026-03-31
• Authors: Jun Liu, Gang Zhao, Junnan Xiong, Tsuyoshi Kinouchi
• DOI: 10.1029/2025wr041360

Research Groups

Not explicitly mentioned in the abstract.

Short Summary

This study proposes a novel graph-based deep learning model (LTG) for daily-scale spatiotemporal flash flood susceptibility (FFS) simulation in China, addressing limitations of traditional models by integrating temporal dependencies, spatio-temporal interactions, and spatial dependencies between catchments. The LTG model significantly outperforms baseline models, achieving an AUC of 0.911 and CSI of 0.719, and effectively captures seasonal FFS variations and spatial hydrological dependencies.

Objective

• To develop and evaluate a graph-based deep learning model (LTG) for spatiotemporal flash flood susceptibility (FFS) simulation at a daily scale in China, considering catchment topology and addressing the limitations of traditional deep learning models in capturing intra-annual temporal variations and spatial interactions between nearby catchments.

Study Configuration

• Spatial Scale: China, focusing on catchments and river networks.
• Temporal Scale: Daily scale, with analysis of intra-annual and seasonal variations.

Methodology and Data

• Models used: LTG model, which integrates Long Short-Term Memory Networks (LSTMs), Temporal Graph Convolutional Networks (TGCNs), and Graph Convolutional Networks (GCNs). Baseline models were used for comparison (specific names not provided).
• Data sources: Not explicitly mentioned in the abstract (e.g., rainfall, topographic, land use data).

Main Results

• The proposed LTG model outperforms baseline models in spatiotemporal FFS simulation.
• Achieved an Area Under the ROC Curve (AUC) of 0.911.
• Achieved a Critical Success Index (CSI) of 0.719.
• The daily-scale simulation by the LTG model exhibits a higher ability to capture seasonal variations compared to yearly-scale modeling.
• Demonstrated a significant intra-annual standard deviation of 0.263 for FFS changes.
• The model effectively considers spatial clustering along river networks and upstream-downstream dependence.
• The model enhances its inferential ability by leveraging information from nearby catchments.

Contributions

• Introduces a novel graph-based deep learning model (LTG) for spatiotemporal flash flood susceptibility (FFS) simulation.
• Addresses the critical gaps in traditional deep learning models by incorporating intra-annual temporal variations and spatial interactions between catchments.
• Provides a daily-scale FFS simulation capability that considers catchment topology and dynamic spatiotemporal dependencies.
• Demonstrates superior performance over existing baseline models in FFS prediction and capturing seasonal variations.
• Offers insights into FFS dynamics along river networks, including spatial clustering and upstream-downstream dependencies.

Funding

Not explicitly mentioned in the abstract.

Citation

@article{Liu2026GraphBased,
  author = {Liu, Jun and Zhao, Gang and Xiong, Junnan and Kinouchi, Tsuyoshi},
  title = {A Graph‐Based Deep Learning Approach for Daily Flash Flood Susceptibility Modeling in China},
  journal = {Water Resources Research},
  year = {2026},
  doi = {10.1029/2025wr041360},
  url = {https://doi.org/10.1029/2025wr041360}
}