Abstract:

[Background] Lunar rock thin-section images contain rich information about geological evolution. However, due to limited sample availability, dataset imbalance, and high annotation costs, traditional supervised learning-based classification methods face significant application challenges. [Methods] To address this, this paper proposes a self-supervised contrastive learning framework based on the DINO model, which extracts image features and integrates them with various classifiers to enable automatic recognition and analysis without requiring labeled data. A lunar rock image dataset was constructed, and contrastive learning was employed for feature modeling, followed by evaluation using multiple classifiers. [Results] Experimental results demonstrate that features extracted by the self-supervised model achieve outstanding performance with classifiers such as KNN and MLP, reaching a maximum classification accuracy of 91.56% from 45.11%, and it did not exhibit the problem of class imbalance even when the sample sizes were significantly different. The t-SNE visualization and confusion matrix analysis further confirm the model's effectiveness in feature clustering and category discrimination. The model exhibits strong robustness and generalization capabilities. [Conclusion] This study provides a feasible approach for automated interpretation of lunar rock images, supporting research on lunar geological evolution and related deep-space exploration missions.

Key words: lunar rocks, rock thin section images, self-supervised learning, automated classification, DINO