National Center for Farmworker Health

Seismic data interpretation is often challenged by various forms of noise, which can obscure subsurface features critical to hydrocarbon exploration. This study proposes a deep learning approach based on the U-KAN architecture – an enhanced U-Net with Kolmogorov–Arnold Networks (KAN) – to perform seismic denoising on data from Paraná Basin. The preprocessing pipeline includes automatic gain control (AGC), two-stage F-K filtering, trapezoidal bandpass filtering, amplitude clipping, and normalization. The proposed model was trained using a hybrid loss function combining L1 and SSIM, and evaluated using multiple metrics. The generated images achieved an average PSNR of 25.15 dB and SSIM of 0.50, with histogram correlation of 0.9971 and spectral correlation of 0.8933. Compared to the original noisy data (PSNR = 19.09 dB, SSIM = 0.27), the results demonstrate a significant improvement in data quality.