Flash floods are a type of catastrophic disasters which cause significant losses of life and property worldwide. In recent years, machine learning techniques have become powerful tools for evaluating flash flood susceptibility. This research applies stacking and blending ensemble learning approaches to assess the flash flood potential in Jiangxi, China. Four base models – linear regression, K-nearest neighbours, support vector machine, and random forest – are adopted to build the two ensemble models. All models are evaluated by three metrics (accuracy, true positive rate, and the area under the receiver operating characteristic curve) and compared with a Bayesian approach. The results suggest that the blending approach is superior to all the other models, which has then been selected to evaluate the vulnerability of flash floods for all the catchments in Jiangxi. The derived maps of flash flood susceptibility suggest that over half of the province, in terms of either area or the number of catchments, are prone to flash floods, in particular the north, northeast and south. These empirical findings can help to develop plans for disaster prevention and control, as well as improving public knowledge of flash flood hazards.