Reliable uncertainty model calculation in subsurface engineering from pore- and grain-scale to field-scale relies on sufficient data, but subsurface dataset acquisition remains a challenge, particularly in domains where data collection is expensive or time-consuming, such as Computed Topography (CT) imaging for digital rock images. While AI-based data augmentation may assist the model training, it still requires many training images as well as the quality assessment of generated data. Yet, most data quantitative metrics flatten spatial data into vectors; therefore, removing the essential spatial relationships within the data. We evaluate topology-based metrics for quality assessment of AI-based image augmentation, coupled with digital rocks augmentation practice using the Single image Generative Adversarial Network (SinGAN) for binarized (segmented) images. Compared to most traditional dimensionality reduction methods that process images into a flattened vector, we propose topological image analysis for dimensionality reduction while preserving the essential geometric and topological features of the high-dimensional data. To demonstrate our proposed approach, we evaluate the generated images starting from four distinct digital rock samples, sorted sandstone, synthetic sphere pack, limestone, and poorly sorted sandstone, using Minkowski functionals, image graph network-based measures, graph Laplacian-based measures, local trend maps, and a homogeneity-heterogeneity classifier. Our workflow suggests that AI-based digital rock augmentation, combined with topological dimensionality reduction offers a powerful tool for enhanced quality assessment and diagnostic of digital rock augmentation and improved interpretation to support decision-making.