Seismic Resolution Enhancement Using a Transformer by Fusing Local and Global Attention

Abstract

High-resolution seismic data is crucial for accurately characterizing thin reservoirs and small-scale faults, providing the foundation for lithology and fluid identification. However, image super-resolution techniques based on deep-learning methods, particularly convolutional neural networks, are fundamentally constrained to local receptive fields. Such architectures struggle to capture the long-range spatial dependencies required to model continuous seismic events, often leading to oversmoothed results or structural artifacts. To address this problem, we develop a dual-attention transformer termed Seis-DAT. The core of Seis-DAT is a dual-attention transformer block that integrates window-based self-attention to capture local textures with channel attention to provide a global receptive field. This design enables the model to learn local details and global geological trends, ensuring both fidelity and structural consistency in the reconstructed results. To mitigate the scarcity of high-quality labeled data, we construct a large-scale synthetic dataset incorporating complex faults and folds. Furthermore, a composite loss function combining the ?1 loss and multiscale structural similarity loss is employed to dynamically balance pixel-level accuracy and structural integrity through a linear weight scheduling strategy. Synthetic tests demonstrate that the proposed method achieves higher peak signal-to-noise ratio and structural similarity index measure values than conventional comparative methods. Applications to real seismic data show that Seis-DAT generates clearer seismic events with enhanced lateral continuity, thereby improving seismic data resolution.