B. Eliminate Spurious Gradient Irregularities 



Temperatures on a "standard scale bathythermograph" (Figure 1) are 

 represented from 28° to 90° Fahrenheit in a graph interval of about 3.3 inches. 

 Thus, an increment of 0.1° Fahrenheit is represented by 0.0053 inches on the 

 standard scale BT slide. The author has concluded from measurements on this 

 scale that a stylus trace thickness varies between 0.01 and 0.02 inches, cover- 

 ing a temperature range of from 0.2° to 0.4° Fahrenheit. Assuming that at 

 this graph scale the value read for a data point may vary within the thickness 

 of a stylus trace, then two different readers may record values for the same 

 depth on the same profile that differ by as much as 0.4° Fahrenheit on a vertical 

 trace . 



Graphical methods (Figure 3) may be used to illustrate the extreme 

 range of error that may be expected when temperature gradient is calculated 

 using the differences between adjacent data points. Recalling that data points 

 are digitized at 0.03 inch intervals on the depth scale, suppose that two 

 different readers were to independently digitize opposite temperature extremes 

 within a trace thickness at consecutive levels on a vertical trace. If the line 

 thickness is 0.02 inches, the two readers will, in effect, digitize graphical 

 temperature slopes that deviate plus or minus 33° from the true slope of the 

 stylus trace, or 66° from each other. It is desirable to eliminate such spurious 

 irregularities from the digital profile, if possible. It may be expected that an 

 algorithm designed to eliminate spurious irregularities in a digitized profile 

 data set will sacrifice real profile irregularities of short interval and small 

 amplitude. It is required to minimize such sacrifice, and it is desirable to 

 provide an estimate of the possible sacrifice that may have been sustained. 



C. Identify Straight Line Subsets in a Data Point Aggregate 



When two different readers digitize points at identical depths from a 

 trace that is straight, the lines computed to fit the resulting data sets will tend 

 to agree more closely, depth for depth, than corresponding points in the original 

 data sets (Figure 4). Moreover, the slope of the straight line computed to fit 

 either data set will more likely provide a better approximation of the slope of 

 the trace than slopes caluclated from the differences between consecutive data 

 points. Accordingly, it would seem desirable to replace the set of points 

 digitized at closely spaced intervals from a straight line trace by vectors selected 

 to represent the line of regression on the data points and the regression interval . 



Of course, it happens that most BT profiles cannot be represented by a 

 straight line. However, a profile can be represented by sets of straight lines, 

 each line selected to fit a series of data points that fall within a negligible 



