Hierarchical Deep Learning for Carbonate Grain Classification in Thin-Section Images
2026 11th International Conference on Computer Science and Engineering (UBMK), İstanbul, Türkiye, 16 - 20 Eylül 2026, ss.1046-1049, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/ubmk71847.2026.11745104
- Basıldığı Şehir: İstanbul
- Basıldığı Ülke: Türkiye
- Sayfa Sayıları: ss.1046-1049
- Kocaeli Üniversitesi Adresli: Evet
Özet
Carbonate grain classification from thin-section images is difficult because the visual differences between classes are often small and the class distribution is strongly imbalanced. In the dataset used here, 2,642 annotated grains from 18 plain polarised light micrographs are grouped into four classes: peloid, ooid, broken ooid, and intraclast. Peloids account for 87.1% of the grains, while broken ooids account for only 1.1%. This paper presents a hierarchy-aware ResNet-18 model that follows the geological decision structure of the task: peloid versus non-peloid, ooid-like grain versus intraclast, and intact ooid versus broken ooid. Each decision head is trained with stage-specific Focal Loss, targeted oversampling is applied to broken ooids, and a staged training schedule is used to reduce majority-class dominance. On a held-out test set of 529 grains, the full model achieves 83.5% Balanced Accuracy and 92.1% Overall Accuracy. Broken-ooid recall increases from 16.7% in the hierarchical model without oversampling to 83.3% in the full configuration.