Deep learning has transformed image analysis across scientific domains, but its application in geosciences remains limited by scarce training data. Traditional image augmentation techniques, such as translation, rotation, and flipping, fail to capture the complexity and range of subsurface geological feature configurations. Here, we demonstrate that mathematical morphology-based image augmentation significantly improves the performance of deep learning models in data-limited subsurface geological applications, achieving an \(R^2\) of 0.97 compared to 0.84 for standard augmentation methods, with as few as four original training images. Our approach systematically transforms images by mimicking the natural variability of subsurface geological structures, enabling models to learn subsurface-specific invariants rather than merely being location- and orientation-independent. This novel method addresses a fundamental challenge in geoscientific machine learning, where high data acquisition costs have historically restricted model development. By generating geologically plausible synthetic training examples, our method enables robust subsurface geological characterization with minimal original data. Validated on synthetic pore/channel images, this morphology-based augmentation strategy demonstrates potential for extension to other deep learning-based subsurface characterization tasks in resource exploration and development applications.
