Reliable hydrologic modeling in arid, topographically complex watersheds depends on accurate land-use/land-cover (LULC) representation. This study evaluates how different LULC categorization methods affect simulated runoff for the Wadi Hatta watershed (UAE) using a GIS-driven machine learning framework that combines high-resolution remote sensing with hydrologic modeling. LULC maps were generated in Google Earth Engine using Random Forest (RF) and Support Vector Machine (SVM) classifiers applied to Sentinel-2 (10 m) and Landsat 8/9 (30 m) imageries and compared with the 10 m ESRI predefined LULC dataset. The resulting LULC classifications were converted to SCS Curve Numbers and used in HEC-HMS hydrologic modeling to simulate runoff under a 50-year design storm, under consistent meteorological and physical conditions. Results show that Sentinel-2 + SVM achieved the highest classification accuracy (overall accuracy up to 0.86) and produced the earliest and highest simulated peak discharge (11.4 m³/s), reflecting improved detection of impervious surfaces. In contrast, the Landsat-9 + RF scenario yielded the lowest peak (7.5 m³/s), consistent with a higher proportion of pervious land covers. LULC change analysis between 2017 and 2024 showed increases in forest cover (1.0–3.3%) and built-up areas (6.0–7.9%) driven by afforestation and urban expansion. These results demonstrate that LULC input resolution and classifier selection significantly influence hydrologic model sensitivity and runoff estimates, underscoring the need for carefully selected, high-resolution LULC products in flood risk assessment and water resource planning in data-scarce arid environments.