Development of Fragility Curves for Coastal Roads Exposed to Tsunami-Induced Debris Flows

This paper presents logistic regression-based fragility models for classifying tsunami damage to flexible pavement roads across three damage levels (DL1, DL2, and DL3). Using a hybrid dataset combining observed and synthetically generated samples to address class imbalance, models were trained on preprocessed, class-balanced data with standardized scaling and cross-validation. Models trained on augmented data showed a significant out-of-bag error reduction (0.41 versus 0.64 for observed data) and improved classification accuracy (94.6% versus 60.3%). The DL1 model demonstrated strong discrimination (receiver operating characteristic curve = 0.85) with high precision and recall, while DL2 and DL3 showed reduced performance due to data sparsity. Calibration analysis confirmed reliable predictions for DL1, with decreasing confidence for higher damage levels. Meanwhile, the DL1 model shows strong predictive performance, limited observed data, and reliance on synthetic augmentation reduce accuracy and confidence for DL2 and DL3. Therefore, predictions for moderate and severe damage should be interpreted cautiously. Nevertheless, the smooth, monotonic fragility curves displayed clear damage level separation and physical acceptability, providing a robust foundational framework for tsunami-induced road fragility assessment. Future work with richer datasets is needed to enhance model reliability across all damage levels.