From Data to Decision: Interpretable Machine Learning for Predicting Flood Susceptibility in Gdańsk, Poland

Abstract

Flood susceptibility prediction is complex due to the multifaceted interactions among hydrological, meteorological, and urbanisation factors, further exacerbated by climate change. This study addresses these complexities by investigating flood susceptibility in rapidly urbanising regions prone to extreme weather events, focusing on Gdańsk, Poland. Three popular ML techniques, Support Vector Machine (SVM), Random Forest (RF), and Artificial Neural Networks (ANN), were evaluated for handling complex, nonlinear data using a dataset of 265 urban flood episodes. An ensemble filter feature selection (EFFS) approach was introduced to overcome the single-method feature selection limitations, optimising the selection of factors contributing to flood susceptibility. Additionally, the study incorporates explainable artificial intelligence (XAI), namely, the Shapley Additive exPlanations (SHAP) model, to enhance the transparency and interpretability of the modelling results. The models’ performance was evaluated using various statistical measures on a testing dataset. The ANN model demonstrated a superior performance, outperforming the RF and the SVM. SHAP analysis identified rainwater collectors, land surface temperature (LST), digital elevation model (DEM), soil, river buffers, and normalized difference vegetation index (NDVI) as contributors to flood susceptibility, making them more understandable and actionable for stakeholders. The findings highlight the need for tailored flood management strategies, offering a novel approach to urban flood forecasting that emphasises predictive power and model explainability.