James 



(4) Absolute Data Reqwlrements - Area A 



So far this report has dealt with quantities of data 

 relative to the original base of 400 perfect observations. It was 

 found that analyses of as little as 3«5 percent of these data re- 

 vealed useful information concerning the isotherm pattern. With 2 

 percent data input it was not possible to show the detail known to 

 be present in the isotherm pattern. 



For several reasons the 2 percent data input is felt 

 to be equivalent to the present data availability utilized in ASWEPS. 

 The mean absolute error for the analysis of 2 percent of present data 

 is 3«^ F. This value is close to the mean value of the mean absol- 

 ute errors found in nine evaluations of present manual sea surface 

 temperature analyses. These evaluations included 1786 individual 

 verifying temperature observations and the mean absolute error aver- 

 aged for all nine evaluations was 3.61 F. Reports by Ja-mes (1966), 

 Tuttell (1963), Shank (1966), Carman (1965) and James (I965), de- 

 scribe these evaluations. 



The second reason for assuming the 2 percent data in- 

 put is typical of nresent data availability is that 2 percent of the 

 original data is eight observations and the average mxmber of obser- 

 vations found per day in an equivalent area during I966 was 7 '8 

 observations . 



On this basis 2 percent of the data used in the tests 

 Is considered equivalent to the present data input of ASWEPS. Thus 

 from Figure 8, to reduce the analysis error to below 2^ would re- 

 quire 8 percent of the present type data or four times as much as 

 presently available. This would be 32 observations per five degree 

 square. Hie same analysis reliability could be obtained, of course, 

 by doubling the data input but utilizing only accurate data as used 

 in Case III. TSiis would require only I6 observations. Of course 

 the increase of data will be mostly sMp of opportunity injection 

 reports for a time and gradually the better instrumented reports 

 will predominate. Thus, the true curve representing analysis 

 accxjracy with increasing data will slice across the three curves 

 in Figure 8, approaching the Case III cttrve in tim.e. 



( 5 ) Non-Random Data 



Figure 8 shows the errors that result from analyses 

 of various quantities and qualities of random distributions of data. 

 If ship of opportunity data were rejected and only ART or buoy data 

 utilized then the distribution of data could be pre-selected. To 

 ascertain the value of specifying the data distribution instead of 

 accepting a random field several ART tracks and buoy arrays were 

 specified so as to give varying quantities of perfect data. 



L 



139 



