Evaluation of automated lithology classification architectures using highly-sampled wireline logs for coal exploration

Wireline logs are a supplemental data source to conventional core logging. The recent explosion of machine learning algorithms has provided researchers with the opportunity to develop advanced statistical tools for automatically classifying lithology from these logs, enabling geologists to rapidly produce first-pass interpretations and validate others, even when core samples are missing or damaged. However, the machine learning algorithms need to be evaluated in the case where wells contain a large number of wireline logs which are highly-sampled. This paper explores different machine learning algorithms and architectures for classifying lithologies (e.g. coal and sandstone) using wireline data from a project area well-known for coal mineralisation: Juandah East, 60km north-west of Wandoan (Queensland, Australia). We used data from seven wells, each containing 19 wireline logs uniformly sampled at 1cm, retrieved from the open Queensland Digital Exploration (QDEX) database. Three popular supervised machine learners, namely the Naïve Bayes classifier, Support Vector Machine, and Artificial Neural Network, were tested under two architectures: committee (one classifier per well log) and singular (one classifier for all well logs). Favourable performance was achieved under both architectures when the base classifier was tuned to maximise a coal-specific performance metric. Results show that the committee architecture increased overall accuracy, generally by increasing accuracy on the dominant lithology class and reducing the classification rate of minor lithology classes. Overall accuracy was further improved by post-processing to remove thin classified intervals (