Seismic Data Resolution Enhancement Using an Improved Residual Channel Attention Network

Accurate interpretation of subsurface structures in geologically complex regions is often hindered by the inherently limited resolution of seismic data. To overcome this limitation, we propose an Improved Residual Channel Attention Network (IRCAN) for seismic super-resolution. We designed a novel IRCAN architecture that integrates Canny edge detection, a multi-scale residual module, and a Bottleneck Attention Module (BAM), where these individual modules are adapted from computer vision. In addition, we introduce two physics-inspired mechanisms, depth-dependent amplitude decay and horizon-guided coherence attenuation, which are novel to seismic processing, and employ a hybrid loss function combining L1 and gradient losses. The architectural modifications enhance sensitivity to subtle structural features, enrich multi-scale contextual understanding, and strengthen spatial and channel-wise feature representations. The physics-inspired mechanisms realistically simulate the degradation effects present during actual data acquisition. Moreover, the hybrid loss function composed of L1 and gradient losses enables the network to balance amplitude fidelity and structural continuity. Experiments conducted on synthetic Marmousi models and 3D field seismic datasets demonstrate that IRCAN achieves substantial improvements in resolution, noise robustness, and geological fidelity. Frequency-domain analyses further verify that the enhanced high-frequency components correspond to authentic subsurface features, rather than artifacts, underscoring the model’s practical utility for seismic interpretation.