Augmentation and Evaluation of Geological Tabular Data: Geo-TabGAN Model and Its Applications

Abstract

Data augmentation plays a crucial role in data-driven geoscience research by minimizing sampling costs and improving the generalization and predictive accuracy of models utilized in mineral exploration and oil and gas development. Although geoscience data are predominantly structured in tabular form, research focused on the augmentation of such structured data remains in its nascent stages. This study seeks to address two fundamental questions: 1) is it feasible to generate realistic synthetic data when only a limited amount of labeled data are available? and 2) what criteria can be established to evaluate synthetic data to ensure it contributes positively to model performance? To this end, we introduce the geological tabular data generative adversarial network (Geo-TabGAN) model and propose a comprehensive evaluation framework. Experimental results derived from core analysis data of the Bayan Obo deposit in Inner Mongolia indicate that the integration of synthetic data led to improvements exceeding 5% in the average accuracy, precision, recall, $F1$ score, and Matthews correlation coefficient (MCC) across three classifiers. This methodology significantly enhances the efficacy of big data analysis and predictive modeling within the geoscience domain.