GeoProtoNet: texture-aware few-shot lithology classification from rock imagery using frozen DINOv2 features

Abstract Rapid lithology identification on open-pit bench faces is typically performed manually by geologists, yet it is subjective, inconsistent under variable field conditions, and cannot be scaled to the large image volumes required by conventional supervised deep learning. We present GeoProtoNet, a lightweight few-shot learning framework for rock image classification that leverages frozen DINOv2 Vision Transformer features combined with episodic metric learning. Two geology-inspired adaptation modules are introduced: a Geological Texture Attention Module (GeoTAM) that uses Gabor-initialized filters and channel–spatial attention to emphasize grain-scale fabric cues, and Task-Conditional Prototype Calibration (TCPC) that adjusts class prototypes according to support-set variance to reduce intra-class heterogeneity caused by weathering and illumination. By pre-caching DINOv2 patch tokens and training only the 0.6M-parameter head, the framework achieves computational efficiency compatible with real-time mining workflows. Rigorous multi-seed evaluation (five independent seeds, 600 episodes per seed) shows that GeoProtoNet attains 63.56 ± 0.76% accuracy and 63.37 ± 0.56% macro-F1 in 3-way 5-shot meta-test episodes, approaching the performance of modern few-shot methods such as FEAT and ProtoMAML while using 7–8 × fewer trainable parameters. Ablation studies confirm that both GeoTAM and TCPC are necessary and complementary: removing either degrades performance below even a generic linear baseline. Gabor initialization outperforms random initialization by 2.4–2.7 percentage points, validating the geological texture prior. Cross-domain evaluation on independent rock-image datasets without fine-tuning reveals a significant accuracy drop to 46.37 ± 0.95%, indicating that domain shift remains a practical barrier and that field deployment should include local calibration. End-to-end latency is 55.5 ms per image (or 5.0 ms when features are pre-cached), and calibration analysis supports a human-in-the-loop protocol where uncertain predictions are flagged for expert review.