<p>Flood susceptibility assessments are essential for urban planning and disaster risk management, especially in rapidly urbanising regions with evolving land use patterns. While many studies assess flood risk at a single period, few explore how susceptibility evolves alongside long-term land use change. This study fills that gap by conducting a machine learning-based comparison of flood susceptibility in Kuala Lumpur for the years 2000 and 2020. Random Forest (RF) and Logistic Regression (LR) were used to model flood-prone areas based on climate, terrain, hydrological, and anthropogenic variables. Results indicate notable shifts in susceptibility, driven by urban expansion and vegetation loss. The RF model achieved high accuracy (96% in 2000 and 94% in 2020), with balanced precision and recall. The LR model showed slightly lower performance (93% and 89%, respectively), with a tendency to underpredict risk zones. Validation against historical flood records showed strong spatial alignment. SHapley Additive eXplanations (SHAP) were applied to enhance interpretability, revealing that the importance of factors like elevation, NDVI, and drainage density varied between decades. TWI, NDBI, and river proximity maintained moderate influence. A threshold-based Modified Normalised Difference Water Index (MNDWI) method was introduced, improving surface water and flood point detection and offering scalability to other regions. This study provides novel insights into the spatiotemporal dynamics of flood susceptibility. It emphasises the importance of integrating temporal land cover change into flood modelling. These findings will support adaptive flood risk management and urban resilience strategies, especially in fast-developing Southeast Asian cities.</p> Graphical Abstract <p>Based on the graphical snapshot, this study presents a spatiotemporal flood susceptibility assessment in the rapidly urbanising city of Kuala Lumpur over a two-decade period (2000–2020), employing machine learning and explainable artificial intelligence (XAI) to understand the influence of land use and topographic dynamics on flood risk. The graphical abstract illustrates the workflow comprising data-driven modelling, spatial mapping, feature attribution, and temporal land use analysis. Flood susceptibility maps generated for 2000 and 2020 indicate a spatial expansion of high-risk zones, particularly in low-lying and urbanised areas. Two models, Random Forest (RF) and Logistic Regression (LR), were applied and compared. RF achieved superior performance, with 96% and 94% accuracy for 2000 and 2020, respectively. The SHapley Additive exPlanations (SHAP) framework was used to interpret model predictions, revealing changes in the importance of flood-driving factors. Elevation and vegetation cover emerged as dominant predictors, while the land use/land cover (LULC) change analysis revealed a 2.77% increase in impervious surfaces and a 4.26% decline in vegetation, highlighting the adverse impacts of urban expansion and reduced green cover on flood risk. The Modified Normalised Difference Water Index (MNDWI) was validated for flood point detection and found effective across spatial contexts. This integrative approach offers actionable insights for sustainable urban planning and adaptive flood risk management. The study’s findings support data-informed policymaking to address increasing flood susceptibility under dynamic land transformation conditions.</p>