Seismic Stratigraphic Interpretation Based on Unsupervised Validation and Spectral Clustering Sampling

Abstract

With the continuous development of computer technology and significant improvements in computing power, deep learning has found increasing applications in seismic stratigraphy interpretation, showcasing notable advancements over traditional methods. However, due to the unique characteristics of seismic data, labeling such data has become extremely challenging and time-consuming, necessitating the involvement of professional geologists. Consequently, few-shot learning has garnered considerable attention for seismic image segmentation. Nevertheless, two key challenges remain in few-shot learning: selecting more representative samples and validating the model during training. As the availability of labeled samples decreases, we are left with inadequate data for the validation set. In this article, instead of solely focusing on enhancing the network structure, we propose the utilization of spectral clustering sampling (SCS) methods for training data selection. In addition, we introduce a metric called sum of differences (SD), which can be computed without the need for labeled data, to replace the conventional validation set loss employed in traditional validation approaches. Notably, by employing SCS methods for training data selection and introducing the SD metric to replace traditional validation set loss in F3 dataset, we have achieved remarkable outcomes.

Keywords

• Deep learning
• image segmentation
• sampling method
• seismic interpretation