<p>Seismic exploration faces significant challenges due to inherent limitations in field data acquisition, resulting in low-resolution, noise-contaminated images. These data quality issues are critical bottlenecks in subsurface characterization, as limited frequency bandwidth blurs thin-bed interferences, while low signal-to-noise ratios obscure subtle fault discontinuities, directly hindering precise reservoir delineation and trap identification. This study introduces the Enhanced Parallel Attention Convolutional Network (EPACNet), an advanced deep learning framework designed for simultaneous seismic super-resolution reconstruction and noise attenuation. The proposed architecture incorporates two key innovations: an Overlapped Patch Merging module that effectively preserves high-frequency features during downsampling operations, coupled with an Enhanced Parallel Attention Convolution module that synergistically combines multiscale channel and pixel attention mechanisms to optimize both structural coherence and fine-scale reflectivity preservation.</p><p>Comprehensive evaluation on synthetic datasets demonstrates EPACNet’s superior reconstruction accuracy, achieving an improvement of 3.97 dB in Peak Signal-to-Noise Ratio (PSNR) and an increase of 0.053 in Structural Similarity Index (SSIM) compared to conventional U-Net architectures. The framework exhibits enhanced capabilities in resolution enhancement and noise suppression while maintaining computational efficiency for practical deployment. The application of Field data further validate the method’s generalization capability, with processed results showing remarkable improvements in geological feature representation, including: enhanced fault discontinuity sharpness, refined thin-layer boundary delineation, and extended frequency bandwidth characteristics. These advancements significantly improve seismic interpretation accuracy for hydrocarbon exploration and reservoir characterization applications.</p>