Rapid Identification of Flood Inundation Areas and Dominant Drivers in Compound Floods Using Explainable Machine Learning

Compound floods, driven by concurrent heavy rainfall and high tides, increasingly threaten coastal urban areas with severe infrastructure damage and socioeconomic disruption. Rapid identification of inundation extents and their dominant drivers is crucial for enabling timely emergency response and infrastructure protection. Current machine learning (ML) models excel at predicting flood inundation extents, but their 'black-box' nature restricts identifying dominant local drivers. While explainable AI (XAI) techniques are emerging to address this challenge, it is necessary to determine how to best pair an XAI method with a high-performing ML model to ensure both predictive accuracy and robust interpretability. This study systematically evaluates Explainable Machine Learning (EML) models to predict inundation areas and identify dominant drivers. Our EML models were created by pairing two representative XAI techniques, SHAP and LIME, with three distinct types of ML models: a linear model (Logistic Regression), a tree-based ensemble (Random Forest), and a neural network (Multilayer Perceptron). Norfolk, Virginia, USA was selected to train, test, and conduct driver analysis for the models. Results revealed that the RF achieved the best predictive performance (Accuracy = 0.81). Furthermore, of the two XAI techniques evaluated, SHAP-based driver attribution demonstrated greater consistency with real-world conditions because its ability to account for complex driver interactions, such as rainfall and tides, provides a more reliable identification of their influence. By leveraging XAI techniques, ML models can move beyond prediction to guiding informed decision-making and developing more effective flood management strategies.