Petrographic analysis is based on the microscopic description and classification of rocks and is a crucial technique for sedimentary and diagenetic studies. When compared to hand specimens, thin sections provide better and more accurate means for analysis of mineral proportion, distribution, texture, pore space analysis, and cement composition. Most petrographic analysis relies on visual inspection of rock thin sections under a microscope, a task that is laborious even for experienced geologists. Large projects with a tight time frame requiring the analysis of a large amount of thin sections may require multiple petrographers, thereby risking the introduction of inconsistency in the analysis. To address this challenge, we explore the use of deep convolutional neural networks (CNN) as a tool for acceleration and automatization of microfacies classification. We make use of transfer learning based on robust and reliable CNN models trained with a large amount of non-geological images. With a relatively small number of labeled thin sections used in “fine-tuning” training we are able to adapt CNN models that achieve low error levels (<5%) for the classification of microfacies from the same dataset, and moderate results (<40%) for the classification of microfacies of thin sections from different datasets. These alternate datasets differ from the training data on two independent factors: the thin sections are from different formations and are prepared by different laboratories. While becoming widely accepted as a useful tool in the biological and manufacturing disciplines, CNN is currently underutilized in the geoscience community; we foresee an increase of use of such techniques to help accelerate and quantify a wide variety of geological tasks.

中文翻译：

---

使用深度卷积神经网络进行岩相微相分类

岩相分析基于岩石的微观描述和分类，是沉积和成岩研究的关键技术。与手工标本相比，薄片为矿物比例、分布、质地、孔隙空间分析和水泥成分分析提供了更好、更准确的方法。大多数岩相分析依赖于在显微镜下对岩石薄片进行目视检查，即使对于经验丰富的地质学家来说，这项任务也很费力。需要对大量薄片进行分析的时间紧迫的大型项目可能需要多名岩相学家，从而有在分析中引入不一致的风险。为了应对这一挑战，我们探索使用深度卷积神经网络 (CNN) 作为加速和自动化微相分类的工具。我们利用基于强大且可靠的 CNN 模型的迁移学习，这些模型经过大量非地质图像训练。通过在“微调”训练中使用的标记薄片数量相对较少，我们能够调整 CNN 模型，该模型在对来自同一数据集的微相进行分类时实现低误差水平 (<5%)，以及中等结果 (<40 %) 用于对来自不同数据集的薄片的微相进行分类。这些备用数据集在两个独立因素上不同于训练数据：薄片来自不同的地层，由不同的实验室准备。虽然 CNN 在生物和制造学科中被广泛接受为有用的工具，但目前在地球科学界并未得到充分利用。我们预计此类技术的使用会增加，以帮助加速和量化各种地质任务。