3-D UXSE-Net for Seismic Channel Detection Based on Satellite Image Enhanced Synthetic Datasets

Abstract

Channels are essential indicators of sedimentary environments and play a vital role in geological applications, such as hydrocarbon exploration, sediment transport, and the study of ancient river geomorphology. Deep learning (DL) techniques have shown great potential in improving channel detection accuracy and efficiency. However, insufficient labeled training data remains a key challenge for refining DL models. To address this issue, we propose a workflow that automatically generates synthetic datasets by integrating channel features extracted from high-resolution satellite images. We first extract river channel features and grayscale values from satellite images. These extracted features are then used to construct reflectivity models, incorporating structural deformations based on seismic reflector dips. The reflectivity models are subsequently convolved with wavelets to generate synthetic datasets. These synthetic datasets are used to train the proposed 3-D UXSE-Net, which integrates the 3-D UX-Net architecture with the squeeze-and-excitation blocks. The model generates improved feature representations that enhance performance by combining convolutional neural networks for local feature extraction and Transformer-based modules for capturing global context. We validate our approach by applying the model to both synthetic and 3-D field seismic datasets. Our results show that 3-D UXSE-Net outperforms baseline methods, including the coherence-based approach and other DL models, and demonstrates strong generalization to field data even when trained solely on synthetic data. Comparisons of different methods highlight the effectiveness of the synthetic data generation approach for DL model training.