arbitrarily defined set of boundary values consisting of the regression 

 winds along the grid perimeter. The ship observation locations usually 

 lie randoKly within the gridblocks. The difference between the observed 

 wind and the regression wind (initial guess) was translated to the 

 nearest gridpoint and used to correct the initial guess there, which 

 then became an internal boundary point. The Poisson's equation vras 

 solved by a relaxation procedure throughout which the boundary values 

 remained unchanged. 



In general the windfield which satisfied the Poisson's equation con- 

 tained unreal data and required a certain amount of smoothing to elim- 

 inate small-scale wiggles from the analysis. The degree of smoothing 

 can be variable and several values of the smoothing operator were tested. 

 The goal of the objective-wind-analysis tests was to determine the 

 proper smoothing factors that resulted in minimum error. Twenty- six 

 separate wind fields were analyzed by the conditional relaxation 

 analysis method with the regression winds serving as the initial guess 

 and the Laplacian of the initial guess serving as a forcing function. 

 Each analysis vras smoothed and verified several times. The minimum 

 analysis error occurred in the totally unsmoothed analysis (Fig. 2) as 

 would be expected but the field was quite erratic. A comparison of the 

 regression vrind errors with the objective-analysis wind errors showed 

 that the direct use of wind observations does reduce the wind specifi- 

 cation errors but only near the location of the wind observations. 

 There is but slight or no improvement elsewhere. 



Three objective-analysis wind fields were obtained corresponding 

 to the same times as the regression wind fields. The main discernible 

 difference between them was that the regression winds specified larger 

 areas of strong winds in the vicinity of the cyclone centers than did 

 the objective-analysis winds. Otherv/ise they were quite similar. 



The simplest method for specifying wind over the ocean is to compute 

 the geostrophic or the gradient wind. As a third approach to see if 

 these might be better than either the regression or the objective- 

 analysis winds, geostrophic and gradient winds were computed from the 

 sea-level pressure gridpoint data at Ships B, C, D, and E and compared 

 with the ship observations. The winds were computed from centered 

 finite-difference approximations over two JKWP grid intervals at each 

 of the toux gridpoints surrounding the ships. By curvilinear inter- 

 polation a value vras then computed at the ships' locations. The 

 results of this showed that the geostrophic and gradient wind errors 

 vrere about 2 kts. higher in all categories than the regression winds. 

 Thus, they were judged inferior to the regression winds. 



To evaluate the relative quality of regression winds with objective- 

 analysis winds one should consider not only the verification scores from 

 tests but also the gridpoint wind fields. Since strong winds are con- 

 sidered more significant for ocean wave forecasting although they 

 usually cover only a small area of the map, care must be taken in 

 interpreting the verification scores. The scores are space averages 



