结合半监督聚类的地质图像多条件生成方法

近年来深度学习技术在地质学的应用越来越广泛.地质学科中的一个重要课题是根据稀疏的空间观测数据建立合理的地下模型.最近的工作通过条件生成对抗网络来探索条件化地质建模,产生了逼真且符合空间观测数据的地质图像.然而,多数方法只关注将空间观测数据作为硬条件,忽视了对生成图像中地质属性的调节.本文引入地质属性标签调节地质图像中具体的地质属性表现,将表征地质属性类别的标签数据作为生成条件之一,扩展一个属性分类器与该标签配合,从而实现更可控的图像生成.针对属性标签的人工标注成本大的问题,本文采用半监督聚类利用少量的标注数据为无标签数据自动分配标签.此外,聚类可能产生噪声标签影响建模结果,此方法使用对称交叉熵损失改进分类网络以提高网络对于噪声标签的鲁棒性.本文在黄河地区的河流地质数据集上进行大量实验,结果表明所提出的方法对于不同的属性标签生成了地质模式不同且符合空间观测数据的逼真地质图像,证明了本方法的有效性.

Multi-conditional Generation of Geological Images Combined with Semi-supervised Clustering

Recently, deep learning has become more and more widely used in geology. An important topic in geological modeling is building a subsurface model according to sparse spatial observation data. Deep learning-based geological modeling has been explored through conditional generative adversarial networks, which results in realistic geological images in line with spatial measurements. However, most methods are only conditioned on spatial observations, ignoring the adjustment of geological attributes in images. This study proposes a method to adjust geological images by introducing geological attribute labels on the basis of spatial measurements. The method introduces label data representing a geological attribute category as one of the generation conditions and expands an attribute classifier to cooperate with the label to adjust the generated image, achieving more controllable images. Considering the high cost of manual labeling, this study adopts semi-supervised clustering to automatically assign labels to unlabeled data using a small amount of labeled data. In addition, clustering may produce noise labels that affect the modeling results. In response, the symmetric cross-entropy loss is used to improve the classful network to enhance the robustness of the network against noise labels. Experiments are carried out on a geological dataset in the Yellow River. Results show that the method achieves realistic geological images featuring different geological patterns and conforming to spatial observations for different attribute labels, which proves the effectiveness of the method.