Xu et al. (2026) SMOTE-BN-FLA: enhanced Bayesian network for rainfall-induced flood loss estimation and mechanism decoding in data-scarce regions

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-01-08
• Authors: Yan Xu, Jidong Wu
• DOI: 10.1016/j.jhydrol.2026.134928

Research Groups

• Joint International Research Laboratory of Catastrophe Simulation and Systemic Risk Governance, Beijing Normal University, Zhuhai 519087, China
• School of National Safety and Emergency Management, Beijing Normal University, Beijing 100875, China

Short Summary

This study proposes SMOTE-BN-FLA, an integrated framework combining the Synthetic Minority Oversampling Technique (SMOTE) with data-driven Bayesian Networks (BN) for rainfall-induced flood loss estimation. The framework addresses data imbalance and opaque disaster mechanisms, demonstrating superior accuracy and interpretability in identifying loss drivers compared to conventional methods.

Objective

• To develop an enhanced data-driven model (SMOTE-BN-FLA) for accurate rainfall-induced flood loss estimation and uncertainty quantification, particularly in data-scarce regions with imbalanced data distributions.
• To decode the underlying mechanisms and identify dominant drivers and cascading pathways of flood losses for improved disaster policy formulation.

Study Configuration

• Spatial Scale: Regional scale, specifically Fujian Province, China. Data is aggregated at the township or county level.
• Temporal Scale: Validated against specific flood events in 2022 and 2024.

Methodology and Data

• Models used: SMOTE-BN-FLA (Synthetic Minority Oversampling Technique coupled with data-driven Bayesian Networks). Compared against conventional methods: Random Forest and Multiple Linear Regression.
• Data sources: Multivariate indicators covering hazard intensity, environmental sensitivity, and socioeconomic vulnerability. Implied to be historical flood event data and associated regional characteristics.

Main Results

• The SMOTE-BN-FLA model achieved significantly higher accuracy (R² = 0.74, ROC = 0.96) compared to conventional methods.
• It demonstrated greater spatial stability and weaker scale dependence, particularly for extreme loss events.
• The model successfully captured 77 % of events experiencing relative losses greater than 1 %, overcoming the underestimation bias common in conventional approaches.
• Mechanistic analysis identified per capita GDP and cumulative precipitation as dominant drivers of flood loss.
• Cascading pathways revealed synergies between terrain, vegetation, and hazard factors in influencing losses.

Contributions

• Proposes a novel integrated framework, SMOTE-BN-FLA, that effectively addresses class imbalance and quantifies prediction uncertainty in flood loss assessment.
• Provides a highly accurate, spatially stable, and less scale-dependent model for rainfall-induced flood loss estimation, especially for extreme events.
• Overcomes the underestimation bias for significant flood loss events, which is crucial for effective disaster management.
• Offers an interpretable system for decoding complex flood loss mechanisms, identifying dominant drivers and cascading pathways, thus aiding in evidence-based policy formulation.
• Advances rainfall-induced flood risk management by concurrently resolving data scarcity constraints and providing insights into region-specific loss causation patterns.

Funding

• Not explicitly mentioned in the provided text.

Citation

@article{Xu2026SMOTEBNFLA,
  author = {Xu, Yan and Wu, Jidong},
  title = {SMOTE-BN-FLA: enhanced Bayesian network for rainfall-induced flood loss estimation and mechanism decoding in data-scarce regions},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.134928},
  url = {https://doi.org/10.1016/j.jhydrol.2026.134928}
}