Wang et al. (2026) An Architecture-Feature-Enhanced Decision Framework for Deep Learning-Based Prediction of Extreme and Imbalanced Precipitation

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

Identification

• Journal: Water
• Year: 2026
• Date: 2026-01-08
• Authors: Yu Wang, Zhicheng Yue, Zhinan Li, Yujia Liu
• DOI: 10.3390/w18020176

Research Groups

Not specified in the provided text.

Short Summary

This study proposes an integrated framework using CNN-LSTM and Transformer architectures with feature engineering and augmentation for precipitation forecasting, finding that the Transformer with physics-informed augmentation achieves an F1-score of 0.1429 for balanced precision and recall, while CNN-LSTM offers superior recall (up to 0.90) for extreme event detection.

Objective

• To develop and evaluate an integrated "architecture-feature-augmentation" framework to improve precipitation forecasting accuracy, particularly addressing atmospheric stochasticity and class imbalance in rainfall datasets.

Study Configuration

• Spatial Scale: Not specified in the provided text.
• Temporal Scale: Not specified in the provided text.

Methodology and Data

• Models used: CNN-LSTM, Transformer.
• Data sources: Not specified in the provided text.
• Techniques: Feature engineering, physics-informed augmentation, synthetic oversampling (SMOTE, GAN).

Main Results

• The Transformer model, combined with feature engineering and physics-informed augmentation, achieved a peak F1-score of 0.1429, representing the optimal configuration for balancing precision and recall in precipitation forecasting.
• CNN-LSTM demonstrated superior robustness in detecting extreme events, consistently maintaining high recall rates (up to 0.90) across various scenarios.
• Feature engineering was identified as a critical modulator, significantly improving CNN-LSTM's baseline metrics and enabling the Transformer to reach its maximum predictive capacity.
• Physics-informed augmentation provided the most consistent performance gains, especially in multi-class contexts, outperforming synthetic oversampling techniques like SMOTE and GAN.
• The Transformer, augmented by physical constraints, is recommended for high-precision requirements, while CNN-LSTM with synthetic augmentation is suggested for early warning systems prioritizing recall.

Contributions

• Proposes and systematically evaluates an integrated "architecture-feature-augmentation" framework for precipitation forecasting.
• Delineates distinct performance profiles for CNN-LSTM and Transformer architectures in the context of precipitation forecasting.
• Provides empirical guidance for selecting optimal model architectures and augmentation strategies based on specific forecasting priorities (e.g., precision vs. recall).
• Advances extreme weather preparedness and strategic water resource management through improved forecasting methodologies.

Funding

Not specified in the provided text.

Citation

@article{Wang2026ArchitectureFeatureEnhanced,
  author = {Wang, Yu and Sun, Yingna and Yue, Zhicheng and Li, Zhinan and Liu, Yujia},
  title = {An Architecture-Feature-Enhanced Decision Framework for Deep Learning-Based Prediction of Extreme and Imbalanced Precipitation},
  journal = {Water},
  year = {2026},
  doi = {10.3390/w18020176},
  url = {https://doi.org/10.3390/w18020176}
}