Julien Maítre, Arnaud L. Back, Kévin Bouchard, Léo Paul Bédard · Applied Computing and Geosciences 2026 · 2026
DOI: 10.1016/j.acags.2026.100415
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Automatic identification of mineral grains from unconsolidated sediments is a critical step in mineral exploration, environmental assessment, and engineering geology, yet it remains a time-consuming and error-prone task when performed manually. This paper presents an end-to-end deep learning framework that addresses both the segmentation and classification of individual mineral grains from high-resolution photomicrographic mosaics acquired under plane-polarized Annular Light (AL). Grain isolation is achieved through the Segment Anything Model 2 (SAM 2), combined with a set of geometric filtering rules designed to eliminate incorrect and redundant segments. Individual grain images are subsequently classified using six pre-trained architectures (VGG-19 with batch normalization, ResNet-101 and ResNet-152, DenseNet-161, ResNeXt-101-32x8d, and ViT-B-16) adapted to our task via transfer learning and fine-tuning. The dataset comprises 8 654 grain images covering 24 mineral species, representing, to the best of our knowledge, the largest number of mineral classes addressed in a single deep learning study in this field. Among the evaluated architectures, fine-tuned ViT-B-16 achieves the highest mean grain-level classification accuracy (0.954 ± 0.006), under the individual-grain partitioning protocol, significantly outperforming all five CNN-based architectures in paired statistical comparisons (Holm-adjusted p < 0.01 ), while achieving a top-3 accuracy of 0.981 ± 0.005. An independently evaluated mixed-mineral mosaic containing 170 reference grains was additionally used to assess heterogeneous-input performance. Of these grains, 154 were retained for classification. Conditional classification accuracy on the retained grains was 0.734 ± 0.032, while the strict end-to-end identification rate relative to all 170 reference grains was 0.665 ± 0.029. These results demonstrate strong classification performance for held-out grains under the controlled acquisition conditions of the present study, while the lower mixed-mosaic end-to-end performance highlights the additional difficulty of heterogeneous-input processing. Broader generalization to independent physical samples remains to be evaluated.
No comments yet — start the discussion below.