<p>Compound floods, in which heavy rainfall coincides with high tides, increasingly threaten low-lying coastal cities. Effective emergency response requires knowing not only where flooding will occur but what drives it. Explainable machine learning is increasingly used to attribute flood drivers, yet the reliability of these attributions is rarely tested. Here we evaluate eight model–explainer pairings—four classifiers (logistic regression, random forest, XGBoost and a multilayer perceptron) combined with SHAP and LIME—for segment-scale driver attribution, assessing each on accuracy, physical consistency and stability. The results show that the tree ensembles were the best predictors (PR-AUC 0.883 for XGBoost and 0.641 for random forest, versus 0.530 and 0.276 for the MLP and logistic regression). SHAP attributed both forcings in the physically expected direction and reproduced consistently across random seeds. XGBoost–SHAP, the best pairing on every criterion, was used to map the dominant driver and revealed a spatially structured driver field within two compound events, where 81–98% of flooded segments were rainfall-dominated and up to 15% tide-dominated—showing that the driver mix is not fixed but needs to be diagnosed locally. By delivering reliable, location-specific driver diagnosis, the framework could inform where flood-emergency resources should be pre-positioned.</p>