James 



The results were that the mean absolute error was re- 

 duced by 20 to 4o percent, with the larger reduction occtirring with 

 the higher number of observations. The results of these tests are 

 shown by Table H. 



TABLE II COMPARISON OF SEA SURFACE TEMPERATURE ACCURACIES 

 FOR RAUDCM AND SPECI FIED DATA DISTRIBUTIOH 



Mean Absolute Error ( F) 

 Percent Data Random Specified 



10 1.30 .69 



5 l.TO 1.08 



3.5 1.89 1.39 



2 2.95 2.81 



The conclusion from these tests is that if more tise is 

 made of specified, reliable temperature observations instead of ran- 

 dom ship of opportunity data the accuracy of the analyses will ap- 

 proach the perfect data curve of Figure 8. Routine ART flights for 

 instance would provide data of stxfficient quantity and quality, and 

 of a specified nature, to permit analyses of less than 1°F error. 



^* Area A - Computer Tests 



Tests similar to those described above using manuall ana- 

 lyzed temperature charts were also made using computer analyses. 

 IHiis was done for two reasons: 



(1) To verify that the fanri.ly of curves shown in Fig- 

 ure 8 are not biased by analysts skills. 



(2) To ascertain whether objective analyses introduce 

 any special requirements or lead to difference emphasis on data 

 requirements . 



The original data ini>ut to the computer analyses was the 

 same as used in the manual analyses and shown in Figure 1. Similar 

 randomly distributed data inputs of 10, 3> 3*5 and. 2 percent were 

 analyzed and the mean absolute error computed. In order to maintain 

 the same degree of pre-knowledge as in the manual analyses a flat 

 field was \ised. This means that every grid point was considered 

 zero except those at which a true observation existed. In the case 

 of the test of 2 percent data this means there were 392 grid points 

 of zero and eight observed values. 



The compater program raaJces multiple passes over the grid 

 relaxing the grid values at non-observed points until the grid field 

 shows certain minimum differences between grid values. A descrip- 

 tion of the computer program from which this test program was drawn 

 is given by Thonrpson (1966). In this case no smoothing was applied. 



140 



