Applying SOM Cluster Analysis With Iterative Refinement to Infer Lithology Units in Eastern Victoria, Australia

Abstract

This study presents a semi-supervised machine learning method for predicting the occurrence of specific surface lithologies over a 330 km × 115 km area in Victoria, Australia. The study area is a geologically complex region within the Lachlan Fold Belt, characterized by orogenic events and surface lithologies that include deep-marine sedimentary turbidites, granitic intrusions, volcanic formations and metamorphic complexes. The approach used a modified Self-Organizing Map algorithm that was enhanced by an iterative multi-step clustering process that used geophysical surveys (magnetic, radiometric, and gravity) with varying signal enhancements as inputs. The clustering results were refined through validation with a lithological database, allowing the algorithm to associate clusters of characteristics in the geophysical survey data with lithological categories. The lithological database comprised both natural rock samples, and synthetic samples derived from published geological maps in order to compensate for severe spatial heterogeneity in the locations of natural samples. It divided the observed and synthetic samples into 11 manually chosen categories that were expected to show distinctive fingerprints in the geophysical survey data: psammitic sedimentary, pelitic sedimentary, chert (quartz-dominant) sedimentary, felsic intrusive, intermediate/mafic/ultramafic intrusive, felsic volcanic, intermediate/mafic/ultramafic volcanic, and regional metamorphic units. Within this simplified set of lithological categories, the output of the algorithm agreed well with a published geological map. The algorithm's performance demonstrates potential for broader applications to spatial lithological prediction, provided that the target rock types are characterized within existing global databases of rock samples and geophysical observations.