面向标签稀缺场景的DINO特征迁移月球岩石薄片图像分类

doi:10.11871/jfdc.issn.2096-742X.2025.04.011

数据与计算发展前沿 ›› 2025, Vol. 7 ›› Issue (4): 129-142.

CSTR: 32002.14.jfdc.CN10-1649/TP.2025.04.011

doi: 10.11871/jfdc.issn.2096-742X.2025.04.011

• 专刊：空间科学大数据智能算法模型与工具 • 上一篇下一篇

面向标签稀缺场景的DINO特征迁移月球岩石薄片图像分类

戴旻昊¹(),董俊烽²,陈剑²,吕英波²,张立^1,^*(),凌宗成²

1.山东大学，机电与信息工程学院，山东威海 264209
2.山东大学，空间科学与技术学院，山东威海 264209

收稿日期:2025-04-29 出版日期:2025-08-20 发布日期:2025-08-21
通讯作者: 张立
作者简介:戴旻昊，山东大学，本科，目前就读于计算机科学与技术专业。
本文负责本文初稿撰写、模型构建和开展实验。
DAI Minhao, is currently pursuing the B.E. degree in Computer Science and Technology at the Shandong University.
In this paper, he is responsible for drafting the paper, building the model and conducting the experiments.E-mail: 3221416220@qq.com|张立，山东大学，副教授。任国家深空探测遥感测绘工作委员会委员、山东宇航学会理事，近年来主要从事机器学习/深度学习在行星数据中应用等方面的教学与科研工作。
本文负责制定研究计划，论文修改。
ZHANG Li, is an associate professor at Shandong University. He is a member of the National Committee for Remote Sensing and Mapping of Deep Space Exploration, and council member of Shandong Society of Astronautics. In recent years, he has mainly been engaged in teaching and research work related to the application of machine learning and deep learning techniques to planetary data.
In this paper, he is responsible for formulating the research plan and revising the manuscript.
E-mail: zhangliwh@sdu.edu.cn
基金资助:
国家重点研发计划(2022YFF0711400);中国科学院国家空间科学中心开放课题(NSSDC2302001)

Lunar Rock Thin Section Image Classification in Label-Scarce Scenarios via DINO-Based Feature Transfer

DAI Minhao¹(),DONG Junfeng²,CHEN Jian²,LYU Yingbo²,ZHANG Li^1,^*(),LING Zongcheng²

1. School of Space Science and Technology, Shandong University, Weihai, Shandong 264209, China
2. School of Mechanical, Electrical & Information Engineering, Shandong University, Weihai, Shandong 264209, China

Received:2025-04-29 Online:2025-08-20 Published:2025-08-21
Contact: ZHANG Li

摘要/Abstract

摘要：

【应用背景】月球岩石薄片图像蕴含丰富的地质演化信息，但受限于样本稀缺、不均衡与高昂的标注成本，传统依赖监督学习的分类方法面临应用瓶颈。【方法】为此，本文提出一种基于DINO模型进行自监督对比学习，并将提取的特征搭配上不同分类器的图像特征提取与分类框架，旨在无标签条件下实现岩石图像的自动识别与分析。构建了月球岩石图像数据集，并以对比学习进行特征建模，结合多种分类器进行分类评估。【结果】实验结果显示，自监督模型提取的特征在KNN、MLP等分类器上表现优异，最高分类准确率从45.11%升至91.56%，并在样本数量差距较大的情况下未表现出类别不平衡问题。t-SNE可视化与混淆矩阵分析进一步证实了模型在特征聚类与类别判别方面的有效性。当前模型整体表现出良好的鲁棒性和泛化能力。【结论】本研究为月球岩石图像的自动化解译提供了一种可行路径，服务于月球地质演化研究和相关的深空探测任务。

关键词: 月球岩石, 岩石薄片图像, 自监督学习, 自动分类, DINO

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

戴旻昊, 董俊烽, 陈剑, 吕英波, 张立, 凌宗成. 面向标签稀缺场景的DINO特征迁移月球岩石薄片图像分类[J]. 数据与计算发展前沿, 2025, 7(4): 129-142.

DAI Minhao, DONG Junfeng, CHEN Jian, LYU Yingbo, ZHANG Li, LING Zongcheng. Lunar Rock Thin Section Image Classification in Label-Scarce Scenarios via DINO-Based Feature Transfer[J]. Frontiers of Data and Computing, 2025, 7(4): 129-142, https://cstr.cn/32002.14.jfdc.CN10-1649/TP.2025.04.011.

图/表 12

图1

表1

图2

图3

图4

表2

表3

图5

图6

表4

图7

图8

参考文献 28

[1]	TAYLOR S R. Planetary science: a lunar perspective[M]. Houston, Tex: Lunar and Planetary Institute, 1982.
[2]	WILSHIRE H G, STUART A L, JACKSON E D. Apollo 16 rocks: Petrology and classification[J]. Journal of Geophysical Research, 1973, 78(14): 2379-2392.
[3]	TAYLOR G J. Ancient lunar crust: Origin, composition, and implications[J]. Elements, 2009, 5(1): 17-22.
[4]	NEAL C R. Petrogenesis of mare basalts: A record of lunar volcanism[J]. Geochimica et Cosmochimica Acta, 1992, 56(6): 2177-2211.
[5]	KIRK R L. The competition between thermal contraction and differentiation in the stress history of the Moon[J]. Journal of Geophysical Research: Solid Earth, 1989, 94(B9): 12133-12144.
[6]	LUCEY P G, BLEWETT D T, JOLLIFF B L. Lunar iron and titanium abundance algorithms based on final processing of Clementine ultraviolet-visible images[J]. Journal of Geophysical Research: Planets, 2000, 105(E8): 20297-20305.
[7]	CHEN J, LING Z C, LIU J Z, et al. Digital and global lithologic mapping of the Moon at a 1:2,500,000 scale[J]. Science Bulletin, 2022, 67(20): 2050-2054. doi: 10.1016/j.scib.2022.09.015 pmid: 36546102
[8]	VON ENGELHARDT W, STENGELIN R. Normative composition and classification of lunar igneous rocks and glasses, I: Lunar igneous rocks[J]. Earth and Planetary Science Letters, 1979, 42(2): 213-222.
[9]	JIANG Y, ZHOU J, FENG J, TENG Q. Application of DBSCAN algorithm and mathematical morphology in rock thin section image segmentation[J]. Microcomputer Applications, 2016, 35: 39-41.
[10]	吴继敏. 薄片显微描绘和自动图像分析技术在岩石学定量评价中的应用[J]. 矿物岩石, 1999(02): 26-31.
[11]	SINGH N, SINGH T N, TIWARY A, SARKAR K M. Textural identification of basaltic rock mass using image processing and neural network[J]. Computational Geosciences, 2010, 14(2): 301-310.
[12]	周程阳, 刘伟, 吴天润, 等. 基于混合专家模型的岩石薄片图像分类[J]. 吉林大学学报(理学版), 2024, 62(4): 905-914.
[13]	程国建, 李碧, 万晓龙, 等. 基于SqueezeNet卷积神经网络的岩石薄片图像分类研究[J]. 矿物岩石, 2021, 41(4): 94-101.
[14]	SU C, ZHU K Y. Research progress of intelligent image analysis for petrographic thin section images[J]. Bulletin of Mineralogy, Petrology and Geochemistry, 2023, 42(1): 13-25.
[15]	ZONG K Q, WANG Z C, LI J W, et al. Bulk compositions of the Chang'E-5 lunar soil: Insights into chemical homogeneity, exotic addition, and origin of landing site basalts[J]. Geochimica et Cosmochimica Acta, 2022, 335: 284-296.
[16]	XIA Z P, MIAO B K, CHEN H Y, et al. The petrology and mineralogy of lunar meteorite EET 96008 from Antarctica[J]. Advances in Polar Science, 2013, 25(4): 352-361.
[17]	CARON M, TOUVRON H, MISRA I, et al. Emerging Properties in Self-Supervised Vision Transformers[C]// 2021 IEEE/CVF International Conference on Computer Vision (ICCV). Montreal, QC, Canada: IEEE, 2021: 9630-9640.
[18]	NASA Johnson Space Center. Lunar Sample Compendium[DB/OL]. Houston: NASA Johnson Space Center. https://curator.jsc.nasa.gov/lunar/samplecatalog/.
[19]	APOLLO 15 PRELIMINARY EXAMINATION TE- AM. The Apollo 15 lunar samples: A preliminary description[J]. Science, 1972, 175(4020): 363-375. pmid: 17731350
[20]	CHEN H, XIE L, SHU Q, MIAO B. Northwest Africa 12279: Evidence for the interaction between early lunar mantle melt and anorthositic crust[J]. Journal of Geophysical Research: Planets, 2023, 128: e2023 JE007844.
[21]	DEER W A, HOWIE R A, ZUSSMAN J. An introduction to the rock-forming minerals[M]. 2nd ed. London: Mineralogical Society of Great Britain and Ireland, 2013.
[22]	TATAR A, HAGHIGHI M, ZEINIJAHROMI A. Experiments on image data augmentation techniques for geological rock type classification with convolutional neural networks[J]. Journal of Rock Mechanics and Geotechnical Engineering, 2025, 17(1): 106-125.
[23]	GUI J, CHEN T, ZHANG J, et al. A survey on self-supervised learning: Algorithms, applications, and future trends[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024, 46(12): 9052-9071.
[24]	ABE M, NIIOKA H, MATSUMOTO A, et al. Self-supervised learning for feature extraction from glomerular images and disease classification with minimal annotations[J]. Journal of the American Society of Nephrology, 2025, 36(3): 471-486.
[25]	BALESTRIERO R, IBRAHIM M, SOBAL V, et al. A Cookbook of Self-Supervised Learning[J]. arXiv preprint arXiv:2304.12210, 2023.
[26]	HE K, ZHANG X, REN S, SUN J. Deep residual learning for image recognition[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Las Vegas: IEEE, 2016: 770-778.
[27]	VAN DER MAATEN L, HINTON G. Visualizing data using t-SNE[J]. Journal of Machine Learning Research, 2008, 9: 2579-2605.
[28]	LIU H, HAOCHEN J Z, GAIDON A, et al. Self-supervised Learning is More Robust to Dataset Imbalance[C]// International Conference on Learning Representations, 2022.

岩性类别	英文名称	所属大类	采样任务	图像类型	数量
钛铁矿岩	Ilmenite	玄武岩	Apollo 11-17	正交偏光薄片显微图像	683
撞击熔融岩	Impact Melt	角砾岩	Apollo 11-17	正交偏光薄片显微图像	623
橄榄石岩	Olivine	玄武岩	Apollo 11-17	正交偏光薄片显微图像	323
风化月壤	Regolith	角砾岩	Apollo 11-17	正交偏光薄片显微图像	611

	参数项	值	说明
DINO	主干网络	ViT-S/16	Patch size 为 16 的小型 Vision Transformer，包含12个 Transformer Encoder Block
	学生模型温度	0.1	控制softmax分布的平滑度
	教师模型温度	0.04	稳定教师输出
	优化器	AdamW	加权衰减的Adam优化器
	训练轮数	500 epochs	保证特征学习收敛
	batch_size	64	实验环境下所能使用的最大 batch_size
KNN	k	20	DINO推荐参数值
MLP-Rock	训练轮数	50 epochs
MLP-Rock	学习率	0.0003
ResNet18	训练轮数	30 epochs
ResNet18	学习率	0.001

Classifier	Accuracy	Precision (Macro)	Recall (Macro)	F1 Score (Macro)	Precision (Weighted)	Recall (Weighted)	F1 Score (Weighted)
KNN	0.8822	0.8830	0.8747	0.8779	0.8830	0.8822	0.8817
SVM	0.8689	0.8751	0.8574	0.8645	0.8701	0.8689	0.8681
SVM-Mix	0.8756	0.8765	0.8691	0.8723	0.8757	0.8756	0.8751
ResNet18	0.8911	0.8881	0.8976	0.8917	0.8936	0.8911	0.8913
MLP-Rock	0.9156	0.9165	0.9210	0.9185	0.9155	0.9156	0.9153

	precision	recall	f1-score	support
Ilmenite	0.9220	0.9489	0.9353	137
Impact	0.9333	0.8960	0.9143	125
Olivine	0.9265	0.9692	0.9474	65
Regolith	0.8843	0.8699	0.8770	123

面向标签稀缺场景的DINO特征迁移月球岩石薄片图像分类

Lunar Rock Thin Section Image Classification in Label-Scarce Scenarios via DINO-Based Feature Transfer

RichHTML

PDF

可视化

摘要/Abstract

引用本文

使用本文

图/表 12

参考文献 28

相关文章 0

编辑推荐

Metrics

本文评价