Impact of data space augmentation strategy on model accuracy and generalization in thin-section rock classification

Area of Science:

• Geoscience
• Machine Learning
• Image Analysis

Background:

• Machine learning (ML) is increasingly vital in geoscience and energy sectors.
• Data augmentation is a common ML practice, but its impact on geoscience models needs systematic evaluation.
• Reliability of ML outcomes is critical for geological data computing.

Purpose of the Study:

• Investigate the effects of static and dynamic data augmentation on convolutional model performance in geological applications.
• Assess the influence of augmentation on model generalization using microscopic rock imagery.
• Provide insights into optimizing ML pipelines for geoscientific prediction tasks.

Main Methods:

• Tested 133 augmentation configurations across 691 scenarios.
• Evaluated five convolutional models (three pretrained, two trained from scratch).
• Utilized datasets of realistic microscopic rock thin section images.

Main Results:

• Data augmentation showed a significant and highly variable impact on model performance.
• Many augmentation methods reduced classification accuracy; specific configurations improved it.
• Linear and nonlinear image mixing enhanced generalization, particularly with limited data.
• Static augmentation sometimes outperformed dynamic approaches in low-data scenarios.

Conclusions:

• Data augmentation can mitigate data limitations in geoscientific ML but requires careful, tailored selection.
• Uncritical augmentation use can distort results and reduce model reliability.
• Specific augmentation techniques, like image mixing, are beneficial for generalization in geoscience.