Two-dimensional shallow water models are widely used tools for flood inundation mapping. However, even if High Performance Computing techniques have greatly decreased the computational time needed to run a 2D inundation model, this approach remains unsuitable for applications that require results in a very short time or a large number of model runs. In this paper we test a non-parametric regression model based on least squares support vector machines as a computationally efficient surrogate of the 2D shallow water equations for flood inundation mapping. The methodology is initially applied to a synthetic case study consisting of a straight river reach flowing towards the sea. A coastal urban area is then used as a real test case. Discharge in three streams and tide levels are used as predictor variables to estimate the spatial distribution of maximum water depth and velocity in the study area. The suitability of this regression model for the spatial prediction of flood hazard is evaluated. The results show the potential of the proposed regression technique for fast and accurate computation of flood extent and hazard maps. K E Y W O R D S flood hazard, flood inundation, Iber model, shallow water equations, support vector machine 1 | INTRODUCTIONPhysically-based flood inundation models are widely used tools for simulating river hydraulics and floodplain inundation processes. In these models, the flood propagation is generally described by the two-dimensional shallow water equations (2D-SWEs), which must be solved using an appropriate numerical technique. Models solve either the full or simplified forms of these equations, as the local inertial approximation or the diffusive wave approximation (Neal et al., 2012). In practice, an inevitable compromise between accuracy and efficiency must be found when defining the spatial resolution of the numerical mesh and the time discretization (Chen et al., 2012). Higher accuracy can be obtained by increasing the physical complexity and the spatial and temporal resolution of a model, but at the cost of higher computation time.This conventional physically-based approach is hence generally impractical for applications that require results in a very short time or a large number of model runs. A typical example is a probabilistic analysis such as the propagation of uncertainty in a Monte Carlo context, which involves the evaluation of thousands model runs. A few Monte Carlo probabilistic analyses using 2D-SWE models have been presented in the literature but the number of runs is generally limited to a few hundreds (Cea, Bermúdez, & Puertas, 2011;Fraga et al., 2016), which limits the number of parameters that can be analysed and the statistical significance of the results. Real-time forecasting systems are another example where fast computation is imperative to issue flood warnings with sufficient lead-time. In these kinds of applications, a possible way to reduce significantly the computation time is the development of computationally more efficient surrogates of these models. A fi...