Multiscale Attention U-Net for Rock Particle Segmentation and Grain-Size Analysis
SPE Journal · 11.11.2026 · 31 (11) · 7583-7603
DOI: 10.2118/236909-pa ↗Kurz gefasst
Fachartikel in der Zeitschrift SPE Journal, online erschienen am 14.09.2026. Verfasst von Y. Luo und 3 weiteren Autor:innen. Die Arbeit wurde bisher 1-mal zitiert (Crossref) und nennt 36 Literaturangaben.
Abstract (Originalsprache)
Accurate segmentation of rock thin-section images is essential for quantitative grain-size and morphology analysis, supporting sedimentological interpretation, rock mechanics analysis, drillability evaluation, and reservoir characterization. However, sandstone thin-section images commonly contain weak grain boundaries, particle adhesion, irregular grain morphologies, complex textures, and scale-varying grains, making manual delineation and conventional image-processing methods time-consuming, subjective, and insufficiently robust. To address these challenges, we constructed a manually annotated sandstone thin-section particle segmentation data set using spatially aligned multiview polarized images and developed a task-oriented segmentation framework based on an enhanced U-Net architecture. The framework integrates multiview polarized image fusion, multiscale convolutional feature extraction, and attention-guided refinement. Multiview fusion incorporates complementary optical information from different polarization states, multiscale convolution captures both fine textural details and broader grain contours, and attention-guided refinement enhances weak-boundary and particle-contact regions. To alleviate data scarcity and improve generalization, online dynamic data augmentation was applied only to the training set, including synchronized horizontal and vertical flips for all five polarized views, brightness/contrast/saturation adjustments, and low-level Gaussian noise injection. The proposed method achieved an intersection over union (IoU) of 90.0% and an F1-score of 94.0%, outperforming the baseline U-Net and representative segmentation models, including SegNet, FCN (fully convolutional network), DeepLab v3+, U-Net++, and HRNet. Beyond mask-level prediction, the segmentation outputs were converted into vectorized particle contours, equivalent grain diameters, and characteristic grain-size percentiles, including D10, D50, and D90. Grain-size evaluation across the test set showed good agreement between model-derived and manual results, with mean absolute relative error (MARE) values of 4.93%, 4.61%, and 4.60% for D10, D50, and D90, respectively. To further evaluate downstream granulometric consistency, grain count and the Folk-Ward sorting coefficient (σφ) were also computed from the predicted particle contours, showing good agreement with manual annotations across the test set, with a grain count mean absolute error (MAE) of 3.84 grains and a sorting-coefficient MAE of 0.039. The proposed workflow therefore provides a practical and reproducible tool for transforming multiview polarized light thin-section images into quantitative grain-size information.
Zitieren
APA
Luo, Y., Shi, X., Chen, S., & Hou, X. (2026). Multiscale Attention U-Net for Rock Particle Segmentation and Grain-Size Analysis. SPE Journal, 31(11), 7583-7603. https://doi.org/10.2118/236909-pa
Vancouver
Luo Y, Shi X, Chen S, Hou X. Multiscale Attention U-Net for Rock Particle Segmentation and Grain-Size Analysis. SPE Journal. 2026;31(11):7583-7603. doi:10.2118/236909-pa
BibTeX
@article{Luo2026,
author = {Luo, Y. and Shi, X. and Chen, S. and Hou, X.},
title = {Multiscale Attention U-Net for Rock Particle Segmentation and Grain-Size Analysis},
journal = {SPE Journal},
year = {2026},
volume = {31},
number = {11},
pages = {7583--7603},
doi = {10.2118/236909-pa}
}
EUVIMED