Wind data were collected by the National Oceanic and Atmospheric Administration WP-3D aircraft on low-level (50 and 90 m) crosswind and along-mean-wind tracks of approximately 350 km during the Storm Transfer and Response Experiment in November and December 1980. Observed mesoscale variations in the marine wind fields are characterized by the velocity correlation tensor for three atmospheric regines: cloud streets, open and closed cellular convection, and prefrontal warm air advection. The dominant scale of mesoscale variation in the offshore wind field normal to the mean wind direction in the case of cold continental air flowing over a warmer ocean, producing cloud streets, was 27 km. For this case, the standard deviation in momentum transfer, which was calculated from 2-km subsets of the flight track by the bulk aerodynamic method assuming a constant drag coefficient, was 13% of the synoptic scale (330 km) mean. The dominant scale of mesoscale variation for open cellular convection was 62 km, and the dominant scale for closed cellular convection was 90 km. The standard deviation of mesoscale momentum transfer (scales greater than 2 km; constant drag coefficient) for a 345-km flight track containing both cell types was 26% of the synoptic scale mean. The warm air advection case had no measurable mesoscale variability. For each regime a model of the horizontal velocity correlation tensor, which can be used to estimate a mesoscale variability, is fitted to the observed velocity correlation tensor with velocity component and weather regime dependent coefficients. This general model is consistent with an interpretation of the mesoscale wind field as an ensemble of coherent structures, associated with cloud type, in which the spatial variability of the wind field in each weather regime is associated with physically determined dominant length scales (i.e., cells or rolls), as contrasted with a continuum interpretation of two-dimensional turbulence. To accurately describe regional winds and fluxes at the sea surface, wind speed and temperature data should be averaged over the dominant mesoscale length scale with either a suitable time average or a spatial average, such as can be obtained by scatterometry, or an estimate of the mesoscale variability should be explicitly stated. It is also suggested that enhanced vertical flux in the oceanic mixed layer occurs at length scales of atmospheric boundary layer structures.