在地球科学领域,岩石微观观测数据的采集过程繁琐且效率低下,这不仅增加了研究成本,降低了可靠性,同时也限制了数据的开源共享。此外,由于岩性的多样性和观测手段的差异,单一数据集的规模通常较小,这对于依赖大规模数据集的深度学习框架而言是一大挑战。为此,本研究探索迁移学习如何促进不同岩性间的信息共享,并通过此机制提高矿物组分识别与智能解释任务的模型性能。通过采集不同区域、岩性、矿物组分和偏光模式下的铸体薄片样本,本文深入研究了深度学习模型在不同观测对象和手段下的迁移学习机制,并聚焦于探索地质信息的深层表征。研究成果不但揭示了迁移学习在促进地质学领域信息共享与模型性能提升中的关键作用,还为自动化和智能化地质认识融合奠定了基础。实验结果显示,通过迁移学习,本文模型在智能解释任务中的准确率显著提高,从53.3%提高至98.73%,而在矿物组分识别任务中,准确率也实现了近10%的提升。这些成果证明了迁移学习在地质学领域内解决实际问题和提高模型泛化能力、性能和稳定性方面的巨大潜力。

In earth sciences the rock microscopic data collection process is both labor-intensive and inefficient, which have a negative impact on research cost/reliability and open data sharing. Additionally, rock heterogeneity and variation in data collection methods typically result in small-scale datasets—this poses a significant challenge to deep learning frameworks that rely on large-scale datasets for training. To address this issue, we investigate how transfer learning can facilitate information sharing across different rock types and enhance model performance in tasks such as mineral identification and intelligent interpretation. By compiling thin section image datasets, taking in diverse rock sampling regions, rock types, mineral compositions, viewed under varying viewing modes, we delve into the mechanisms of transfer learning across different observational targets and methods, focusing on the deep representation of geological information. Our findings not only highlight the pivotal role of transfer learning in promoting information sharing and improving model performance within the field of geosciences, but also lay a foundation for the automatic and intelligent integration of geological insights. According to experimental results, transfer learning led to significant accuracy improvement, from 53.3% to 98.73%, in intelligent interpretation task, and a nearly 10% improvement in mineral identification task. These results convincingly showcase the great potential of transfer learning in addressing practical problems in geology as well as enhancing model generalization, model performance, and model stability.