Here we present a method for using the spatial x–y coordinate of an image cropped from the cylindrical surface of digital 3D drill core images and demonstrate how this spatial metadata can be used to improve unsupervised machine learning performance. This approach is applicable to any data set with known spatial context, however, here it is used to classify 400 m of drillcore imagery into 12 distinct classes reflecting the dominant rock types and alteration features in the core. We modified two unsupervised learning models to incorporate spatial metadata and an average improvement of 25% was achieved over equivalent models that did not utilize metadata. Our semi-supervised workflow involves unsupervised network training followed by semi-supervised clustering where a support vector machine uses a subset of M expert labeled images to assign a pseudolabel to the entire data set. Fine-tuning of the best performing model showed an f₁ (macro average) of 90%, and its classifications were used to estimate bulk fresh and altered rock abundance downhole. Validation against the same information gathered manually by experts when the core was recovered during the Oman Drilling Project revealed that our automatically generated data sets have a significant positive correlation (Pearson's r of 0.65–0.72) to the expert generated equivalent, demonstrating that valuable geological information can be generated automatically for 400 m of core with only ∼24 hr of domain expert effort.

中文翻译：

---

利用机器学习中的空间元数据改进地质钻芯的客观量化

在这里，我们提出了一种使用从数字 3D 钻芯图像的圆柱表面裁剪的图像的空间x – y坐标的方法，并演示了如何使用该空间元数据来提高无监督机器学习性能。该方法适用于具有已知空间背景的任何数据集，但是，这里它用于将 400 m 的钻岩图像分类为 12 个不同的类别，反映岩心中的主要岩石类型和蚀变特征。我们修改了两个无监督学习模型以纳入空间元数据，与不使用元数据的同等模型相比，平均提高了 25%。我们的半监督工作流程涉及无监督网络训练，然后是半监督聚类，其中支持向量机使用M 个专家标记图像的子集为整个数据集分配伪标签。最佳性能模型的微调显示f ₁（宏观平均值）为 90%，其分类用于估计井下大量新鲜岩石和蚀变岩石的丰度。在阿曼钻探项目期间回收岩心时，对专家手动收集的相同信息进行验证表明，我们自动生成的数据集与专家生成的等效数据集具有显着的正相关性（皮尔逊r为 0.65-0.72），这表明有价值的地质信息只需领域专家 24 小时的努力即可自动生成 400 m 的核心。