James 



since the area was so small. Also no forcing function was used in 

 order to avoid injecting prior knowledge of the isotherm pattern. 



(1) Results of Computer Tests - Area A 



Figure 9 shows the computer analysis of 10 percent data. 

 Although this analysis lacks the resolution of the manual analysis 

 shown in Figure 2 the major warm tongue is shown as is the cold water 

 intrusion from the northwest. An interesting aspect of computer 

 analysis is shown by Fipures 10 ajid 11. Both figures represent anal- 

 yses of 2 percent of data but one of perfect data (Figure 10) and the 

 other of poorer quality. Case I data (Figure 11). The interesting 

 point is that there is little difference between the analyses in 

 spite of the great difference in the quality of observations. The 

 same similarity was found in comparison of computer analyses of other 

 percents of good and poor data. 



Ihis leads to the conclusion that the computer program han- 

 dles imperfect data better than the mantial, subjective analysis, 

 since in the latter tests more differences are apparent. This char- 

 acteristic of the computer analyses may be due to the relaxation pro- 

 cedures which tends to modify the temperature field around an obser- 

 vation without creating extreme gradients or i>roducing distortions 

 in the field. An analyst on the other hand tends to extend the iso- 

 therms to include high or low values far from the main warm or cold 

 tongue. 



The mean absolute errors for the computer analyses of per- 

 fect and Case I data are shown in Figure 12. For reference the 

 curves for the same manually prepared analyses are repeated from 

 Figure 8. Although the curves differ slightly the computer curves 

 verify those derived from the manual analyses. No particular differ- 

 ences in data requirements appear to result from these tests except 

 that computer analyses can make better use of a. minimum input of 

 poor data than can manual analyses. As expected from the similarity 

 of their patterns, the differences in accuracy between the computer 

 analyses of 2 percent "good" and "poor" data is small. 



Although the manxial analyses show slightly lower errors 

 than the computer products this has no significance. The computer 

 program used was not as sophisticated as that used operationally, 

 since a simplified prograra was adequate to establish the curves. 



(2) Effect of The Degree of Randomness on Results 



In conjunction with the computer tests a test was also run 

 to ascertain the effect of the degree of randomness in the distri- 

 bution of data on the resvilts. 



141 



