picking procedure. In general, however, one can treat estimates from 

 large provinces of persistent geological processes as more reliable than 

 those generated in a relatively small area. 



One source of error in the model can be quantified, and that is the 

 residual error from the regression. Osing standard statistical tech- 

 niques (see, for example, Draper and Smith, 1981), the difference in 

 individual amplitude estimates from the regression model estimates can 

 be expressed as a root mean square. Errors in this study averaged e^ - 

 - .03 and ^iqo « " * 'OlSm for a single spectrum. These errors reflect 

 many of the errors associated with the model, although they cannot be 

 decomposed into component sources. 



Another important factor determining the level of error in the 

 model is the number of estimates used in generating a composite spec- 

 trum. In the ease of an anisotropic area, a variety of azimuths dis- 

 tributed about the compass allows a better estimate to be made. In gen- 

 eral, the ensembling of N time series composed of signal with noise, 

 results in a decrease of noise (as RMS) of 1/ lOT. In the case of our 

 spectral model, the signals are the derived amplitude spectra along an 

 azimuth. A simple method of improving the prediction capability and 

 accuracy of this model Is to ensemble-average the amplitude versus fre- 

 quency estimates from several proximal and near-parallel tracks. The 

 multlbeam sonar provides exactly this capability and future developments 

 should take advantage of it. 



Figure 7-1 Illustrates this reduction of estimation error through 

 ensemble-averaging of multlbeam sonar derived spectra. For this exam- 

 ple, sixteen parallel beams (profiles) from the SASS multlbeam system 

 were analyzed for the statistically homogeneous province of the Gorda 



109 



