Less Biased Approach for Sandstone Pore Segmentation

Reservoir characterization is a fundamental process in estimating reserves within petroleum and hydrogeological systems, where the precise determination of pore space dictates the validity of fluid flow models. Although X-Ray Computerized Tomography (XRCT) has become the standard non-destructive evaluation method for visualizing the internal structure of rocks, the data interpretation process still faces challenges during the image segmentation stage. Conventional methods, such as greyscale thresholding, often result in inconsistent segmentation because they rely on the subjective interpretation of the operator. This study evaluates the application of the Segment Anything Model (SAM), a computer vision foundation model developed by Meta AI, to perform automated pore segmentation on Ruhr Formation sandstone samples. A dataset of 800 XRCT images was split at the image level into train (80%), validation (10%), and test (10%) sets prior to any processing, ensuring no spatial leakage between sets. SAM’s performance was then comparatively tested against five greyscale thresholding techniques. Experimental results demonstrate SAM’s superiority over the best of the five thresholding methods, achieving a Mean Intersection over Union (mIoU) of 0.4523 and a Dice Score of 0.6226. Further variance analysis reveals that SAM produces more consistent segmentation results than most greyscale thresholding methods, with an IoU variance of 0.0005 and a Dice Score variance of 0.0005. These findings indicate that SAM can transform traditional petrophysical workflows into a more objective and precise process, ultimately improving the accuracy of reserve estimations in subsurface resource exploration.