Simplifying Horizon Picking Using Single-Class Semantic Segmentation Networks

Abstract

Seismic image processing plays a significant role in geological exploration as it conditions much of the interpretation performance. The interpretation process comprises several tasks, and Horizon Picking is one of the most time-consuming. Thereat, several works proposed methods for picking horizons automatically, mostly focusing on increasing the accuracy of data-driven approaches, by employing, for instance, semantic segmentation networks. However, these works often rely on a training process that requires several annotated samples, which are known to be scarce in the seismic domain, due to the overwhelming effort associated with manually picking several horizons in a seismic cube. This paper aims to evaluate the simplification of the labeling process required for training, by using training samples composed of disconnected horizons tokens, therefore relaxing the requirement of annotating the full set of horizons from each training sample, as commonly observed in previous works employing semantic segmentation networks. We assessed two state-of-art neural networks for general-purpose domains (PSP-Net and Deeplab V3+) using public seismic data (Netherlands F3 Block dataset). Our results report a minor impact in the performance using our proposed incomplete token training scheme compared to the complete one, moreover, we report that these networks outperform the current state-of-art for horizon picking from small training sets. Thus, our approach proves to be advantageous for the interpreter, given that using partial results instead of providing a full annotation can reduce the user effort during the labeling process required for training the models.