<p>Flood susceptibility mapping is essential for mitigating urban flood risks, yet balancing model interpretability and predictive accuracy remains a challenge in environmental modelling. While ensemble machine learning models offer strong predictive performance, their complexity often limits interpretability. This study introduces the application of Accumulated Local Effects (ALE), an explainable AI method, to enhance transparency in flood susceptibility mapping. ALE was integrated with logistic regression (LR), random forest (RF), and eXtreme Gradient Boosting (XGBoost) to interpret variable contributions while maintaining predictive performance. Ensemble models outperformed logistic regression, achieving 94% accuracy and an AUC of 0.98, and effectively captured non-linear relationships, including rainfall thresholds and the influence of urban infrastructure. Key flood drivers identified across models included annual rainfall, distance to river, elevation, slope, and Normalised Difference Vegetation Index (NDVI). ALE further revealed nuanced local effects and threshold responses in variables, patterns that were oversimplified in the linear structure of logistic regression. Validated against historical flood records, susceptibility maps clearly delineate high-risk zones suitable for targeted mitigation. This study advances flood risk modelling by combining interpretability and accuracy and provides a scalable, explainable framework for urban planning and early warning systems in rapidly developing regions.</p>