all ten amplitude versus frequency estimates are considered. Since the 

 power law form of the spectrum Is known, this functional form Is used In 

 an Iterative regression procedure described In Appendix A and performed 

 In SUBROUTINE POWFIT. The regression coefficients vary widely along the 

 profile, and therefore It Is desirable. In order to aid In the detection 

 procedure, to smooth these estimates. To save calculation time, this Is 

 done by averaging 91 envelope estimates at each frequency before enter- 

 ing the regression routine. This process tends to smear the boundaries 

 somewhat, but makes detection of provinces more reliable. 



With smoothed estimates of the amplitude spectrum available contin- 

 uously along a profile, the final stage of processing Is to detect sig- 

 nificant chances In the. estimated spectra and on this basis Impose prov- 

 ince boundaries. The regression coefficients a and b represent the 

 antilog of the y Intercept, and the slope, respectively, of the spectrum 

 projected In log-log space. The spectral slope, b (which Is related to 

 the so-called Fractal dimension) Is a worthwhile parameter for province 

 detection. A simple algorithm Is run across the slope estimates to 

 detect significant, rapid changes (boundaries). 



The regression coefficient a Is correlated to b and therefore does 

 not represent an Independent parameter for detection. An alternative Is 

 to look at the total, band-limited RMS energy of the estimated spectra 

 for significant shifts In total energy. These RMS parameters are easily 

 calculated using Parseval's Formula (Integrating the power spectrum) 

 applied to the estimated spectra. In addition to detecting rapid 

 changes as with the slope, the RMS is contoured to form segments of some 

 minimum size. Although these detectors have proven fairly reliable. It 

 Is often necessary to Interpret certain boundaries where the various 



138 



