SynSection: Sedimentology-driven data generation for deep learning applications in carbonates petrography
The analysis of carbonate rocks through petrographic methods has long posed significant challenges in geological sciences, particularly regarding the systematic description and quantification of thin sections. Recent developments in artificial intelligence have suggested promising avenues for automa...
Uloženo v:
| Vydáno v: | Marine and petroleum geology Ročník 182; s. 107490 |
|---|---|
| Hlavní autoři: | , , , , , , |
| Médium: | Journal Article |
| Jazyk: | angličtina |
| Vydáno: |
Elsevier Ltd
01.12.2025
|
| Témata: | |
| ISSN: | 0264-8172 |
| On-line přístup: | Získat plný text |
| Tagy: |
Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
|
| Shrnutí: | The analysis of carbonate rocks through petrographic methods has long posed significant challenges in geological sciences, particularly regarding the systematic description and quantification of thin sections. Recent developments in artificial intelligence have suggested promising avenues for automation; however, the field remains constrained by limited training data availability, as manual annotation requires considerable time investment and can result in variable interpretations. This study addresses these challenges by introducing SynSection, a method for generating synthetic pairs of carbonate thin section images with corresponding labels for both classification and segmentation tasks. This approach has been carefully designed to incorporate sedimentological heuristics, ensuring the geological validity of generated samples through the integration of three primary components: 2D grain packing, groundmass generation using texture synthesis, and image blending for image composition. The advantage of this methodology lies in its ability to exponentially expand a limited set of manually labeled images, thereby reducing the resources required for dataset creation while maintaining consistency in annotation. The method was evaluated through the generation of 100,000 synthetic image-annotation pairs, implementing a transfer learning strategy in which models were initially trained on synthetic data before being fine-tuned on real thin section images. This approach demonstrated significant improvements in both classification and segmentation tasks compared to conventional training methods that rely solely on limited real data. In classification tasks, the transfer learning strategy enhanced performance metrics by 7.2%, achieving a top accuracy score of 94.9%. The impact of the method was particularly notable in segmentation tasks, where the binary grain segmentation model achieved an Intersection over Union (IoU) of 77.2% and a point counting determination score of 93.2%. The multiclass segmentation achieved a mean Intersection over Union (mIoU) of 55.3% and a point counting score of 89.6%, reflecting the inherent complexity of grain type differentiation. These results, which are consistent with medical imaging applications, demonstrate the method’s potential to optimize carbonate thin section analysis by reducing time allocated to descriptive tasks. The approach generates diverse training data that maintains geological validity, establishing a quantitative framework for automated petrographic analysis that aligns with standardized classification criteria and expert-defined features.
•An innovative method for petrography using synthetic data and deep learning.•A multiscale approach including image classification and segmentation.•Key findings for improving deep learning workflows with small labeled databases.•Rigorous analysis and interpretation to assess method benefits and limitations. |
|---|---|
| ISSN: | 0264-8172 |
| DOI: | 10.1016/j.marpetgeo.2025.107490 |