Estimation of the extremal behaviour of hurricane wind speeds is complicated by the lack of accurate extreme data at any one site. As a result, wind engineers have developed climatological/physical models from which hurricane events can be simulated and used as a basis for inference. We present an application of non-linear spatial regression techniques to hurricane wind speed data simulated at locations on the Gulf and Atlantic coasts of the United States. Modelling the spatial variation in extremal behaviour provides a means of pooling data, thus increasing the extent of information available for inference at each site. Estimates of distributional models for extremal behaviour and, consequently, return levels are more precise than those obtained by standard methods applied to individual site data. We also adapt these spatial techniques to estimate the distribution of directions of the r–largest hurricane wind speeds at each site. The techniques we describe are preliminary steps to spatially modelling the joint behaviour of wind direction and speeds. We anticipate that this method of estimating return levels for winds in a given direction will yield estimates with greater accuracy than current techniques enable.