Self-Supervised Learning for Seismic Data: Enhancing Model Interpretability With Seismic Attributes

Deep learning (DL) has shown great potential in geosciences, such as seismic data processing and interpretation, improving decision-making and reducing analysis time. However, DL faces two main challenges. First, many DL models rely on labeled data, which can be time-consuming to obtain. Second, the predictions from these models often lack interpretability, making it difficult to use them for high-value decisions. To address these limitations, we propose a novel workflow that eliminates the need for labeled data and enables interpretation of the results, highlighting key geological features. The proposed workflow trains a vision transformer (ViT) to produce six attention maps, focusing on diverse and relevant regions, by assigning higher attention values. We first train the ViT using a modified distillation with no labels (DINO) method specifically designed for the seismic domain and monitor for overfitting. Then, to evaluate the focus of each attention head (AH), we use nine seismic attributes as predictor features for the assigned attention using a gradient boosting model. Finally, the method samples the seismic attributes in stationary regions of the attention maps and calculates Shapley additive explanations (SHAP) values to determine the most impactful attributes on the attention prediction. Each AH can concentrate on unique geological features of the input seismic image, as indicated by different relationships between SHAP values and seismic attributes. Additionally, regardless of location, each AH can detect the same geologically significant pattern based on the attributes used. The proposed workflow enables the interpretability of the model’s importance, guided by expert knowledge through seismic attributes.