A Multiclass Training Dataset and Hybrid Neural Network for Simultaneous Karst and Channel Detection

Seismic karsts and channels interpretation is vital for reservoir characterization. These geobodies often coexist in seismic volumes and exhibit similar reflection features, making them challenging to distinguish using existing methods that typically interpret them separately. We propose a multi-class hybrid neural network for simultaneously detecting karsts and channels, offering more efficient and accurate results than independent detection. To address the lack of multi-class labeled data, we developed a workflow to generate a diverse synthetic training dataset. This involves constructing 3D impedance models characterized by stratigraphic sequences with karsts and channels embedded according to geological and empirical data. To simulate diverse models, key parameters such as velocities, dimensions, and orientations of geobodies are randomly assigned within reasonable ranges based on geological context. Reflectivity models are computed from impedance models and convolved with Ricker wavelets to simulate seismic data, with noise added to enhance realism, yielding diverse seismic volumes with multi-class labels. This dataset is used to train a U-shaped hybrid neural network combining a Swin-Transformer encoder with a residual module decoder. The Swin-transformer provides global context-awareness and captures long-range dependencies, while the decoder ensures detailed feature restoration. The network performs multi-class segmentation, simultaneously detecting and distinguishing karsts and channels. Field applications demonstrate that the trained model can detect these geobodies with high accuracy and integrity. From the detection, 3D meshes of the geobodies are constructed to model and analyze their structural geometries. We have made our multi-class training dataset publicly available (https://zenodo.org/records/10781510) for further research on karst and channel interpretation.