Floods are a devastating natural disaster, causing significant harm to human lives, economic activities, and cultural heritage globally. This problem is exacerbated in semi-arid regions like Beni Mellal city, Morocco, where mountain terrain accelerates water flow into rapidly urbanizing zones. This study enhances urban flood susceptibility prediction in Beni Mellal city by comparing individual machine learning models Random Forest (RF), Support Vector Machine (SVM), CatBoost (CB) and their pairwise hybrid combinations. An integrated geospatial dataset was used. This included a flood inventory developed from Sentinel-2 and Landsat 7 ETM + imagery using NDFI. Twelve additional conditioning factors were derived from sources including Landsat 8 imagery (for NDVI, NDBI, LULC), a Digital Elevation Model, geological and soil maps. The developed models produced detailed flood susceptibility maps. All evaluated machine learning methods demonstrated strong predictive performance. CatBoost (CB) achieved the highest standalone testing performance (AUC: 91.02%), while among the hybrid models, the Random Forest-Support Vector Machine (RF-SVM) ensemble was most effective, achieving a testing AUC of 92.98%. The analysis consistently identified Slope, Geology, and NDVI as the most influential factors driving flood susceptibility. The resulting urban flood susceptibility maps revealed distinct risk classifications among models. This research confirms that hybrid machine learning can significantly improve urban flood susceptibility assessment. The validated models and detailed maps serve as practical tools for decision-makers in Beni Mellal city, supporting targeted mitigation strategies, guiding sustainable urban planning, and improving community safety in similar semi-arid urban environments.