2002 Denver Annual Meeting (October 27-30, 2002)

Paper No. 14
Presentation Time: 4:45 PM


LANGER, Katherine E., 11764 Melody Drive, Northglenn, CO 80234 and CLOSS, L. Graham, Geology and Geological Engineering, Colorado School of Mines, Golden, CO 80401, bluequartz@aol.com

Data from two drainage geochemistry surveys for gold exploration in southwestern Sumbawa, Indonesia, were used to compare the merits of factor analysis, discriminant analysis, cluster analysis, and neural network techniques for data interpretation. Orientation survey results obtained from 49 sample sites by Moedjiarto in 1993 constituted one data set. Reconnaissance drainage survey results from 261 sample sites conducted by P.T. Newmont Nusa Tenggara in 1987 made up the other data set. Both surveys contain As, Au, Cu, Mo, Pb, Sb, Zn data, as well as lithologic, stream order, and catchment area data. The Batu Hijau porphyry copper-gold deposit, as well as other copper-gold mineral occurrences, is located within the survey areas.

Comparisons of the multivariate data analysis methods were based upon three factors. The first factor was reliability: a method was deemed reliable if it correctly identified nearly all of the catchments with known mineral occurrences. The second was ease-of-use: this criterion relates to the time and experience required to prepare and run each technique. The third factor was cost-effectiveness: this criterion incorporates reliability and consideration of operational costs associated with each technique in terms of both time and money.

Factor and cluster analyses do not require training data to adequately interpret the results. R-mode factor analysis, providing information on element associations per sample site, was the most reliable and cost-effective method, taking the least amount of time to interpret. Cluster analysis, providing information on the sample site associations, was the second most reliable and was estimated to be the third most cost-effective, based upon the description of the steps taken to complete the analysis by Sjoekri (1997). Discriminant analysis and most neural network techniques require training data sets to adequately interpret unknown data. Neither the orientation nor reconnaissance survey contained enough data to develop adequate training sets. Thus, discriminant analysis was the third most reliable, despite its ease of use, as it did not correctly identify most catchments with known mineral occurrences. The neural network technique was the least reliable of the four methods due to the lack of training patterns available resulting in an unstable network.