Automated segmentation of core images

Fossil reefs constitute crucial archives of past environmental changes. Identifying and quantifying sedimentary components of such build-ups is essential for paleoenvironmental interpretation. However, analyzing a significant number of cores demands substantial expertise and resources, making automation a game-changer for efficiency and accuracy.

Based on an IODP reef cores dataset, this project first evaluates classical Deep Learning techniques for image segmentation on real-life core images, emphasizing best practices in an enhanced methodology. It then explores how this approach boosts efficiency for geologists in environmental sciences and subsurface resource industries.
 

We introduce an innovative workflow for semantic segmentation of sedimentological cores using deep learning. It combines three key innovations: the U-Net architecture, a patch-based learning, and an innovative use of weighted loss. A final conditional random fields (CRF) post-processing step enhances geological realism of the segmented images.

The network was trained for 100 epochs, ensuring loss convergence. Considering the accuracy as a performance metric, the model achieved 85.7% accuracy on the training set and 75.8% on the test set. Inaccuracies stem from slight color inconsistencies in the photo set and challenges in segmenting coral classes or components affected by diagenetic alteration or bioerosion.

Given the promising initial results, several approaches can be explored. One option is to enhance the model by incorporating sub-classes, such as different coral morphologies. Alternatively, we could integrate complementary datasets from the same cores (e.g., hyperspectral images, X-ray tomography) for a multi-source approach.