Abstract
In recent years, artificial intelligence has been increasingly applied in geosciences; however, the scarcity of labeled data limits its effectiveness. Existing data augmentation methods using conditional generative adversarial networks (cGANs) have succeeded in fields such as finance and medicine but often struggle to account for the multiscale features and strong correlations inherent in geoscience data, making direct adaptation challenging. This study introduces an improved cGANs — improved conditional geoscience GAN (ICG-GAN) — specifically designed for augmenting tabular geoscience data. ICG-GAN uses continuous features as conditional inputs to preserve their physical properties and correlations, using a classification-based voting mechanism to predict and impute discrete features. This approach effectively overcomes the limitations of traditional cGANs when learning from sparse samples. To systematically assess the model’s performance, we developed a comprehensive, multidimensional evaluation framework. Using core analysis data as a case study, experimental results show that ICG-GAN more accurately replicates the original data distribution and achieves significant performance improvements compared with the leading CTGAN model across six evaluation metrics. This study presents a novel method for augmenting tabular geoscience data, with broad applications in resource exploration, geologic mapping, and environmental monitoring, while providing insights for enhancing the performance and robustness of data-driven geoscientific models.
Funding Information
  • State Key Laboratory of Bayan Obo Rare Earth Resource Researches and Comprehensive Utilization (GZ-2023-1-LH-003)

This publication has 34 references indexed in Scilit: