Abstract
Digital rock physics (DRP), encompassing image preprocessing and property estimation, offers substantial potential for linking geophysical responses to rock microstructures pertinent to engineering within an image-based computational framework. Recent years have witnessed significant advancements in the intelligentization of DRP analysis, but most studies proceed from segmented images to estimate mechanical properties, leading to two potential limitations: (1) the accuracy of image preprocessing based on deep learning relies on a large data scale (2) prior information (e.g., edge features); related to geological activities is often excluded from property estimations. Therefore, we propose a few-shot-based method (i.e., EdgeSegNet) to alleviate data-scale dependence of training networks for image segmentation and extract edge feature, and a multi-task network (i.e., MMOEROCK) introduced by prior information to predict synchronously mechanical properties. Experimental results show that the segmentation and edge detection precision reach approximately 94 % and 98 %, respectively, under a limited training set, outperforming other common networks through three key architectural optimizations: (1) utilize Contrast Limited Adaptive Histogram Equalization (CLAHE) for image contrast enhancement to improve microstructure clarity; (2) employ Fourier Transform and high-pass filtering for high-frequency spatial detail extraction to reveal fine textures and edges; (3) design an Interactive Constrained Module (ICM) to integrate information from segmentation and edge flows to promote inter-task complementarity through interaction and constraints. Meanwhile, the accuracy of mechanical properties prediction constrained by prior information improve by 0.03 on R2-scores compared with those with no constraint, which potentially suggests considering microscopic geometric features linked to geological processes enhances understanding of petrophysical mechanisms.
Paper Information：
Wang, Z., Hou. Z, and Cao. D (2026), Deep-learning-based digital rock physics analysis: from image segmentation and edge detection by few-shot learning to mechanical properties prediction, Geoenergy Science and Engineering, 256, 214133, doi:https://doi.org/10.1016/j.geoen.2025.214133
