159 / 2022-02-27 20:22:51

深度学习算法在露头岩性识别中的应用 ——以无人机采集鄂尔多斯盆地石河子组露头影像数据为例

Deep Learning,Artificial Intelligence,image classification,lithology identification,Unmanned aerial vehicle (UAV),Shihezi Formation

摘要：随着人工智能技术的不断发展，许多不同的领域都开始尝试将机器学习、深度学习技术应用到本领域。随着研究的进行，深度学习技术也开始应用到岩性识别方面。无人机影像具有敏捷性、全局性且能够克服对人不利的地理条件的特点，为地质勘探工作提供了高分辨率，细节丰富的图像。在高分辨率无人机露头图像的基础上，将深度学习算法和图形图像处理算法应用于府谷天生桥露头岩性识别工作。本文提出了一种卷积神经网络与超像素分类结合的露头照片岩性识别算法（CNN-SLIC）。首先对露头照片进行切分并构建数据集，训练出神经网络模型，然后用神经网络对图像进行内容识别，确定岩性的分布；然后使用超像素分类算法划分出岩体边界。最后融化岩性识别结果和岩体边界。本文还采用了语义分割的算法对无人机图像进行岩性识别。语义分割是在像素级别上的分类，相比于其他对象识别方法，语义分割的精确性更高。CNN-SLIC算法既能够准确的识别岩体边缘，又能准确的对岩性进行划分。CNN-SLIC的分类正确率达到了88.7%，而使用语义分割的方法对于岩性的识别准确率更高，但是对于物体边缘的识别不够清晰。这一研究不仅对露头沉积学、遥感地质学的发展具有重要的理论意义，而且对促进了人工智能技术与地质学的结合意义重大。



Abstract: With the development of artificial intelligence technology, many different fields have start to apply machine learning and deep learning technology. With the progress of research, deep learning technology has also begun to be applied to lithology identification. UAV images have the characteristics of agility, globality and ability to overcome unfavorable geographical conditions, providing high-resolution and detailed images for geological exploration work. On the basis of high-resolution UAV outcrop images, deep learning algorithms and graphic image processing algorithms are applied to outcrop lithology identification. This paper proposes an outcrop photo lithology recognition algorithm (CNN-SLIC) that combines convolutional neural network and superpixel classification algorithm. First, segment outcropping photos and constructed the dataset, and train a neural network model. Then, the neural network is used to identify the content of the image to determine the distribution of lithology; then the superpixel classification algorithm is used to delineate the boundary of the rock mass. Finally, melt lithology identification results and rock mass boundaries. In this paper, the algorithm of semantic segmentation is also used to identify the lithology of UAV images. Semantic segmentation is classification at the pixel level, which is more accurate than other object recognition methods.The CNN-SLIC algorithm can not only accurately identify the edge of the rock mass, but also accurately divide the lithology. The classification accuracy rate of CNN-SLIC reached 88.7%, while the method using semantic segmentation has higher accuracy for lithology recognition, but the recognition of object edges is not clear enough. This research not only has important theoretical significance for the development of outcrop sedimentology and remote sensing geology, but also has great significance for promoting the combination of artificial intelligence technology and geology.