196-5 Deep Learning-Driven Strain Quantification in Metasedimentary Rocks: A U-Net-Based Segmentation Workflow

Session: Strain and Displacement: Patterns, Gradients, Partitioning, and Reconstructions (Posters)

Presenting Author:

Authors:

Abstract:

Strain analysis in metasedimentary rocks is a fundamental technique for reconstructing deformation histories, but conventional methods such as, manual tracing of elliptical clasts are time-consuming, prone to user bias, and difficult to scale across large datasets. This study introduces a semi-automated workflow that leverages deep learning and computer vision to improve the efficiency and reproducibility of strain quantification. We developed a U-Net-based convolutional neural network (CNN) trained on petrographic thin section images from the Galice and Mariposa Formations, located in the Klamath Mountains and Western Sierra Nevada Metamorphic Province, respectively. The model was trained using Labkit-generated ground truth masks, enabling segmentation of images into three mineralogical classes: clasts, matrix, and lithics.

The CNN achieved strong performance metrics, with F1-scores exceeding 92% and Intersection over Union (IoU) values above 86% across all classes.  The dataset included 6 high-resolution thin section images, manually labeled and augmented into 1,200 patches for training a U-Net segmentation model. The model achieved strong performance across all classes, with F1-scores of 94% for clasts, 93% for matrix, and 92% for other minerals, and corresponding IoU values of up to 89%, confirming robust pixel-level accuracy. The algorithm could detect around 700 grains in six samples that were tested after model validation. The fully automated pipeline reduced analysis time from several hours to under two minutes per thin section while eliminating subjectivity inherent to manual methods. All trained models, annotated datasets, and processing scripts are openly available to support reproducibility and research applications in structural geology and digital petrography.

Geological Society of America Abstracts with Program. Vol. 57, No. 6, 2025

doi: 10.1130/abs/2025AM-9419

© Copyright 2025 The Geological Society of America (GSA), all rights reserved.