| Abstract |
In surficial lithologic mapping, geologists use remotely sensed data prior to fieldwork, however, the utility of these datasets are limited due to vegetation cover. Thus, the use of other sources of information about chemical and physical properties of rocks such as geochemical data (e.g., from stream sediment samples) and airborne geophysical data (e.g., radiometric data) becomes important. In this study, two clustering algorithms, partition around medoids (PAM) and model-based clustering (Mclust) were performed in stream sediment geochemical (SSG) and airborne-gamma-ray (AGR) data as well as in SSG and AGR together to map surficial lithologies in vegetation-covered areas in Central Part of British Columbia Province-Canada. Prior to clustering two approaches, conventional and compositional (CoDa), were applied to SSG and AGR data in order to study the influences of closure problems within the data. In SSG data analysis, clustering was applied using all 13 elements and selected nine elements. In addition, two types of data integration was done SSG all element and AGR (Reference Data I); and SSG selected element and AGR (Reference Data II). Overall accuracy and kappa coefficient was computed for the results and, two references were used to assess accuracy of the classification which is simplified existing lithological map (Reference Data I) and the lithological map based on the interpretation of airborne magnetic data (Reference Data II). The integration of SSG and AGR data produces better results than those using both SSG and AGR data separately. The percentage accuracies of integration data compare to their separated data increase quite significant up to 17% and 0.15 for overall accuracy and kappa coefficient, respectively. In addition, Mclust produces better classifications for lithilogical mapping relatively to PAM clustering base on both qualitative and quantitative assessments. Qualitatively, from visual evaluation, the patterns of Mclust results are more similar to lithological patterns in the existing lithological map than PAM clustering. Quantitatively, assessments results in each separated data (SSG or AGR) show up to 5% and 0.7 differences for overall accuracy and kappa coefficient, respectively, whereas for the integrated data (SSG and AGR) produces non-significant difference results (1% and 0% differences for overall accuracy and kappa coefficient, respectively). Therefore, Mclust could be applied to integrate and classify SSG and AGR data for lithological mapping in regional scale. |