MinDet1: A Deep Learning-enabled Approach for Plagioclase Textural Studies

This is a Preprint and has not been peer reviewed. The published version of this Preprint is available: https://doi.org/10.30909/vol.07.01.135151. This is version 2 of this Preprint.

Add a Comment

You must log in to post a comment.


There are no comments or no comments have been made public for this article.


Download Preprint

Supplementary Files

Norbert Toth , John Maclennan


Textural information, such as crystal size distributions (CSDs) or crystal aspect ratios are powerful tools in igneous petrography for interrogating the thermal history of rocks. They facilitate the investigation of crystal nucleation, growth and mixing as well as the cooling rate of the rock. However, they require large volumes of crystal segmentations and measurements that are often obtained with manual methods. Here a deep learning-based computer vision technique, termed instance segmentation, is proposed to automate the pixel-by-pixel detection of each plagioclase crystal in thin section images. Using predictions from a re-trained model the physical properties of the detected crystals, such as size and aspect ratio, can be rapidly generated to provide textural insights. The present segmentations are validated against published results from manual approaches to prove the method's accuracy. The power and efficiency of this automated approach is showcased by analysing an entire sample suite, segmenting over 48,000 crystals in only a matter of days. Widescale use of this method is expected to drive significant developments in the igneous petrography and related fields




Earth Sciences, Other Earth Sciences, Physical Sciences and Mathematics


Deep learning, Segmentation, Petrography, timescales, MCMC


Published: 2023-05-16 01:29

Last Updated: 2024-02-19 09:22

Older Versions

CC BY Attribution 4.0 International