An Architecture-Feature-Enhanced Decision Framework for Deep Learning-Based Prediction of Extreme and Imbalanced Precipitation

Accurate precipitation prediction is critical for water security and disaster mitigation, yet remains challenging due to atmospheric complexity and class imbalance in rainfall data. This study introduces an integrated "architecture-feature-augmentation" framework to address these limitations. Through systematic comparison of CNN-LSTM and Trans-former architectures, we identify a fundamental trade-off: CNN-LSTM demonstrates higher enhanceability, achieving 80% recall for heavy rainfall when combined with phys-ics-informed augmentation, while Transformer shows superior inherent sensitivity (75% recall) but greater vulnerability to data distribution shifts. Feature engineering benefits are model-specific, significantly improving CNN-LSTM but often introducing redundancy for Transformer. Notably, oversampling techniques like SMOTE achieve peak F1 scores but with substantial generalization gap (ΔF1 > 0.47), indicating overfitting risks, whereas physics-informed augmentation proves more reliable. We establish a principled decision framework: for robust predictions, use CNN-LSTM with physics-informed augmentation; for peak performance where risks are tolerable, employ CNN-LSTM with SMOTE. These findings provide scientific guidance for extreme weather preparedness and water resource management.