Leveraging U-Net/CNN-Based Segmentation in Deep-Learning-driven Strain Analysis of Metasedimentary Rocks

Strain analysis is essential for reconstructing the deformation history of rocks; however, traditional methods typically rely on manual tracing of elliptical clasts, which is both time-consuming and susceptible to user bias. Although deep learning has been applied in stress and strain estimations in engineered materials, its application to geological strain analysis has never been attempted. This study introduces a pioneering semi-automated workflow that integrates deep learning and computer vision to address this gap. A U-Net-based Convolutional Neural Network (CNN) was trained on petrographic thin-section images from the Galice and Mariposa Formations, located in the Klamath Mountains and the Western Sierra Nevada Metamorphic Province, respectively. Image segmentation into matrix, clasts, and lithics was performed using ImageJ-generated ground truth masks and real-time data augmentation. An OpenCV-based ellipse-fitting algorithm extracted key geometric parameters, including aspect ratios, centroid positions, and clast orientations. The segmentation model was trained on 1200 image patches, achieving F1-scores across classes above 92% and Intersection over Union (IoU) values exceeding 86% across all classes, which demonstrated high pixel-level classification accuracy. Strain-related patterns were visualized using rose diagrams and axial ratio histograms. This workflow drastically reduces strain analysis time while improving reproducibility and scalability. Although challenges such as textural complications persist, our proposed pipeline offers a robust and transferable framework for quantitative fabric analysis. The trained model, annotated dataset, and processing scripts are publicly available to support continued development and broader application in geologic studies.