Rock type classification using MGM peak-fitting analysis of hyperspectral reflectance spectra (0.35--2.50mum), Sudbury Igneous Complex, Ontario

Automated methods of determining rock type are needed to speed and standardize the rock core logging process. The feasibility of a rock classification system that uses peak-fitting results of hyperspectral reflectance spectra was investigated. The first step in this research was to obtain operational peak-fitting software capable of reproducing published results of the original Modified Gaussian Model (MGM) software program (MGM-Brown), a peak-fitting approach well documented in the geological literature. The spectra of three pure minerals were processed with PeakFit, a commercially available peak-fitting program, and with MGM-IDL, a version of MGM-Brown written in IDL. The fit results obtained from PeakFit and MGM-IDL were compared to published MGM-Brown results. Both programs were found to closely reproduce the MGM-Brown results. However, PeakFit provides additional outputs and a more user-friendly interface than MGM-IDL and was chosen for the remaining analyses. Hyperspectral reflectance spectra (0.35-2.50mum) were measured for 81 rock samples collected primarily from mines near Sudbury, Ontario, and were analyzed with PeakFit using a modified Gaussian distribution for peak shape. The dataset was split into a development set, with which to develop the rock classification system, and a test set to verify how well the classification system works. Twenty-five absorption peaks were identified from the development set spectra, and each peak was attributed to an absorption cause. Three decision tree, rock classification systems were created based on the absorption peaks, and the decision trees were tested using the independent test data. When classifying test samples according to names assigned by INCO Limited geologists, sixty percent (60%) of samples were correctly classified. Lower accuracy (40%) was achieved when samples were classified using decision trees for rock types based on thin section (TS) mineralogy. For all decision trees, the classification accuracy varies greatly among rock types. Higher accuracies were obtained for amphibolites, quartz diorites and granites (INCO categories) and for tonalites and quartz-rich granitoids (TS categories), whereas lower accuracies were obtained for gabbros and norites (INCO categories) and for quartz monzonites, gabbros and quartz gabbros (TS categories). One problem common to all rock types is that sample size is small. This meant that the development set samples of a given rock type often did not exhibit the full range of spectral and mineralogical variability characteristic of the rock type, and that the development and test samples differed substantially, both mineralogically and spectrally. Despite this issue, classification accuracy of 60% is significantly higher than what would be expected due to chance alone (<10% accuracy), which suggests that further research involving an enlarged sample set is warranted.