Automated Seismic Horizon Picking Using Vision Transformer

Horizon picking is essential in seismic interpretation due to its impact on subsequent processes. Traditional manual and automatic methods are time-consuming and dependent on human expertise. Recently, deep learning techniques have been adopted to overcome these issues. Traditional convolutional neural networks (CNNs) have difficulty effectively capturing global features and complex spatial relationships in seismic data. An architecture based on the self-attention mechanism called Vision transformers has recently been introduced to overcome this challenge. In this study vision transformers were used to investigate its potential for picking horizons on seismic data for effective subsurface characterization. We established a training method for a Vision Transformer (ViT) model grounded in structural geological modelling. This approach allows for the rapid and accurate mapping of horizon labels using structural modelling techniques. The findings indicated that Vision transformer potential to horizons with a high level of accuracy and surpass traditional approaches' limitations.