Efficient Sedimentary Facies Recognition Using Vision Transformer and Weakly Supervised Deep Multi-View Clustering

Sedimentary facies recognition plays a crucial role in geological exploration and oil-gas resource evaluation. However, traditional recognition methods are limited by their ability to extract local features and efficiently utilize unlabeled data, making it difficult to effectively handle complex sedimentary facies characteristics. To address this, this paper proposes an efficient sedimentary facies recognition method based on Vision Transformer (ViT) and weakly supervised deep multi-view clustering. First, we use ViT for feature extraction from sedimentary facies images, capturing global information through the self-attention mechanism, thereby enhancing the ability to recognize complex sedimentary facies patterns. Secondly, we introduce deep multi-view clustering, which integrates multi-angle features such as color, texture, and shape, improving the model’s robustness and classification accuracy. Additionally, we apply weak supervision, combining a small amount of labeled data with a large amount of unlabeled data, and use strategies like pseudo-labeling to enhance the model’s generalization ability and reduce the cost of data labeling. Experimental results show that this method achieves higher accuracy and robustness across multiple sedimentary facies datasets, significantly outperforming traditional Convolutional Neural Networks methods in recognition performance. This research provides a novel solution for sedimentary facies recognition, especially suited for scenarios with limited labeled data and complex, diverse features, and offers new research directions for future deep learning-based geological data analysis.