Abstract
Seismic facies segmentation plays a critical role in seismic interpretation and geological analysis, providing essential support for subsurface stratigraphic characterization and hydrocarbon reservoir identification. Although deep learning methods have made significant progress in this field, conventional supervised segmentation models typically require large volumes of high-quality labeled data and can only recognize the fixed categories defined in the training set, limiting their adaptability to variations in seismic data distributions across different survey areas. Moreover, these models cannot incorporate expert feedback to refine predictions, lacking interactive and iterative optimization capabilities. To address these limitations, we propose UmixClick, an interactive seismic facies segmentation network based on a Mix-Transformer encoder and a Multiscale Self-Adaptive decoder. The model leverages user clicks for guidance, enabling open-ended exploration of unknown or complex subsurface structures, while the multi-scale feature extraction mechanism enhances the accuracy of boundary delineation and irregular geological body identification. Experiments on the F3 dataset demonstrate that UmixClick achieves superior interactive segmentation performance and strong generalization ability. By integrating interactive labeling with transfer learning strategies, the model effectively overcomes the cross-domain adaptation challenges faced by conventional approaches, offering a novel solution for seismic facies segmentation and annotation.