James 



Figure 6 is simply a model to illustrate the three error cases applied 

 and was not part of the tests. 



(3) Results of Error Tests 



Figure 7 shows the analysis of 10 percent of original 

 data which have heen randomly modified hy the Case I error function. 

 Ibis figure illustrates what happens when kO observations of the 

 quality presently utilized in synoptic analyses are analyzed instead 

 of kO perfect observations as in Figure 2. One of the two tongues 

 is located relatively well but the cold water and the western tongue 

 are misaligned. Gradients are also over-emphasized in some areas. 



Analyses of 5> 3*5 and 2 percent modified Case I data 

 are not shown but revealed a poorer fit to the true isotherm pattern 

 than did analyses of the ssune quantity of perfect data. As would 

 be expected analyses of the seune quantities of data for the Case II 

 and Case III errors produce isotherm patterns worse than those from 

 perfect data and better than the ansilyses of Case I data. All anal- 

 yses show that at 2 percent of data the isotherm pattern is not 

 reliable but this qualitative information was not the goal of this 

 study. 



Mean absolute errors were computed for all tests and 

 it is these values that provide quantitative answers to some of the 

 questions posed in the objective. The mean absolute errors were 

 plotted versus data input with a family of curves resulting as shown 

 by Figure 8. For purposes of comparison assume that 2 percent data 

 is our present data input; this will be justified in a later section. 

 The following conclusions can be made from Figure 8: 



(a) In all cases an increase in data produces a de- 

 crease in the mean absolute error or an increase in analysis reli- 

 ability. The reliability of the analyses can be increased consider- 

 ably by simply doubling the present input data. Indications are 

 that it would be desirable to triple or perhaps quadruple the pres- 

 ent data input. 



(b) QSie same improvement in SST analyses can be ob- 

 tained by quadrupling the data input using present type SST reports, 

 or by only doubling the input but using more reliable data. This 

 could be accomplished, for instance, by using only ART or NSRT data 

 instead of the present predominantly ship of opportunity injection 

 reports. 



(c) Increases in reliability beyond 8 percent (quad- 

 rupling the data) may not be desirable owing to the small improve- 

 ment per data increase. 



136 



