迁移学习及其在固体地球科学中的应用  

Transfer learning and its application in solid Earth geoscience

作　　者：林秋怡 左仁广[2] LIN Qiuyia;ZUO Renguang(School of Earth Resources,China University of Geosciences(Wuhan),Wuhan 430074,China;State Key Laboratory of Geological Processes and Mineral Resources,China University of Geosciences(Wuhan),Wuhan 430074,China)

机构地区：[1]中国地质大学(武汉)资源学院,武汉430074 [2]中国地质大学(武汉)地质过程与矿产资源国家重点实验室,武汉430074

出　　处：《地质科技通报》2025年第1期346-356,共11页Bulletin of Geological Science and Technology

基　　金：湖北省创新群体项目(2023AFA001);国家自然科学基金项目(42102332)。

摘　　要：随着地球科学进入大数据时代,机器学习成为可发现数据复杂结构与模式的新兴工具。作为机器学习的一个重要子领域,深度学习通过构建多层隐含层的方式,层层递进地学习海量数据内在特征,可达到提高分类或预测效果等目的。然而机器学习模型往往需要海量数据作为支撑,从而限制了其在固体地球科学领域的广泛应用,迁移学习算法的引入为解决这一问题提供了新的方案。迁移学习可通过利用预先学习类似任务的知识来提高新任务的性能,将源域学习到的知识迁移到目标域,可以在一定程度上克服训练数据不足的问题。迁移学习算法为机器学习在固体地球科学领域的应用提供了新的思路。本文简要综述了迁移学习的基本概念和类别,通过分析迁移学习在固体地球科学中的典型应用案例,讨论了现有迁移学习方法在固体地球科学领域中面临的挑战。当前,迁移学习方法已经在岩石矿物自动识别与分类、地球化学异常识别等方面表现出较大潜力,其具备提高模型泛化性能、避免过拟合的能力,在固体地球科学领域具有广阔的应用前景。但目前迁移学习方法应用于固体地球科学领域的研究还相对较少,未来将持续针对源域数据集选择、迁移模型构建、负迁移评估及可解释性不足等问题开展更为深入的研究。[Significance]With the advent of big data in geoscience,machine learning has emerged as a powerful tool that are able to characterize intricate structures and patterns of data,thus rapidly gaining attention in solid Earth geoscience.As a crucial branch of machine learning domain,deep learning leverages large amounts of datasets to construct multilayer hidden layers,enhancing the classification or prediction performances.Nevertheless,one of the significant difficulties for machine learning models in geoscience is the scarcity of available data,which is limited in solid Earth studies.The advent of transfer learning has introduced a novel approach to address this challenge by using limited training data for effective applications.[Progress]As a typical machine learning technique,transfer learning enhances the performance of new tasks within limited data by utilizing preexisting knowledge from similar tasks through pretraining.By transferring knowledge from a source domain to a target domain,it can partially mitigate insufficient data availability so that prediction accuracy can be improved.This study provides an overview of transfer learning's basic concepts and categories,discussing challenges in current geoscience applications,and analyzing typical cases in solid Earth geosciences.Currently,deep transfer learning shows promising potential in automatic identification and of rocks-minerals classification and geochemical anomalies identification.[Conclusions and Prospects]Transfer learning holds considerable promise for enhancing model generalization performance and mitigating overfitting in solid Earth geosciences.However,some challenges still remain,such as identifying suitable source domains to supply relevant knowledge for target domains.Future research should be explored in terms of source domain dataset selection,transfer model construction,negative transfer assessment,and interpretability of transfer learning.

关 键 词：迁移学习 深度学习 固体地球科学 机器学习 

分 类 号：P628[天文地球—地质矿产勘探]