About one-third of the Earth's land surface is located in arid or semi-arid regions, often in areas suffering severely from the negative impacts of desertification and population pressure. Reliable hydrological forecasts across spatial and temporal scales are crucial in order to achieve water security protection from excess and lack of water for people and ecosystems in these areas. At short temporal scales, flash floods are extremely dangerous hazards accounting, for example, for more than 80% of all flood-related deaths in the USA. Forecasting of these floods requires a connected spatially-distributed hydro-meteorological modelling system which accounts for the specific meteorological and hydrological characteristics of semi-arid watersheds, e.g. summertime convective rainfall and channel transmission losses. The spatially highly heterogeneous nature of the precipitation and the nonlinear response behaviour of the system demand the explicit accounting and propagation of uncertainties into the model predictions. This short paper presents the results of a multi-year study in which such a system was developed for flash-flood forecasting in the semi-arid southwestern USA. In particular, we discuss our effort to understand and estimate underlying uncertainties in such a modelling system. To achieve this we use the GLUE approach to uncertainty analysis, in combination with a variance-based global sensitivity analysis technique. In general, the level of uncertainty found was very high and largely dominated by uncertainty in the radar rainfall estimates. Regarding the model parameters, uncertainties in the hillslope model parameter values had a greater impact on the predictions than the uncertainties in the channel parameters, at least for relatively small basins.