A self-supervised deep learning framework for seismic facies segmentation

Ming Li, Xue-song Yan, Qing-hua Wu,
A self-supervised deep learning framework for seismic facies segmentation,
Expert Systems with Applications,
Volume 288,
2025,
128290,
ISSN 0957-4174,
https://doi.org/10.1016/j.eswa.2025.128290.
(https://www.sciencedirect.com/science/article/pii/S0957417425019098)
Abstract: Seismic facies classification plays a vital role in subsurface exploration and geological interpretation, but the dependence on manually labeled seismic data can hinder scalability and efficiency. In this paper, we propose SSFS (Self-supervised Seismic Facies Segmentation), a novel framework that combines deep clustering with self-supervised learning to achieve seismic facies segmentation without the need for labeled data. Traditional methods rely on costly manual labeling or simplistic clustering, while supervised deep learning struggles with limited labeled data. SSFS addresses these limitations by deep clustering and self-supervised learning. The SSFS framework consists of three key phases: (1) Deep Clustering, where seismic data is sliced into overlapping patches of varying sizes and features are extracted using an attention-based auto-encoder, followed by clustering the features in latent space using K-means; (2) Cluster Merging, where initial clusters are iteratively merged based on cosine similarity to refine the clustering results; and (3) Facies Segmentation, where the merged clusters serve as pseudo-labels for training a segmentation model to refine the facies classification. On the Netherlands F3 dataset, SSFS achieves an average mean Intersection over Union (mIoU) of 83.76% and average pixel accuracy (PA) of 88.43% on four test seismic profiles. Moreover, we show that SSFS can enhance supervised seismic facies analysis by initializing popular segmentation models with the pre-trained weights from the SSFS framework. Quantitative validations across two datasets demonstrate that SSFS-initialized models consistently improve class accuracy (CA) by 2–5% compared to models trained from scratch for minority facies classes. The effectiveness of SSFS is quantitatively validated through several metrics, highlighting its potential for both self-supervised and supervised seismic facies classification tasks. Our results suggest that SSFS provides a promising solution for seismic facies analysis, especially in scenarios with limited labeled data, and has the potential to significantly improve seismic interpretation workflows in practice.
Keywords: Seismic facies classification; Self-supervised learning; Deep clustering; Seismic signal analysis; Attention-mechanism