Enhanced Hardrock Seismic Imaging Through Multi‐Scale Information‐Guided Unsupervised Learning

Abstract In hardrock or crystalline rock geological settings, due to low impedance contrast, reflected energy is usually weak. In addition, often stronger surface waves and noncoherent noise are observed including high‐frequency scattering noise, which seriously covers the useful reflection signal. Therefore, imaging of hardrock seismic data with a low signal‐to‐noise ratio (S/N) is challenging and requires tailored and cumbersome processing workflows. In this study, we propose an unsupervised learning‐based framework with frequency‐guided constraints for pre‐stack seismic data denoising. The proposed label‐free framework contains two input channels, noisy and time‐frequency‐domain data conditioned through a continuous wavelet transform (CWT) filter. The CWT filtered data provide richer feature representations guiding better the reconstruction of seismic signals. The proposed framework consists of several feature attention blocks with the soft attention mechanism to extract the spatial relationship between noisy and CWT filtered data and assign higher weights to significant features. To improve the denoising performance, we designed a hybrid loss function containing the log‐cosh function, amplitude‐weighted constraint, and frequency‐dynamic weighted constraint. We use one synthetic and two real pre‐stack seismic data sets from two mineral‐endowed regions in Sweden and Canada to test the effectiveness of the proposed network. Compared with the three benchmarks, our proposed framework shows stronger reflection signal recovery and is capable of better attenuating the complex noise. The proposed denoising workflow allows improved delineation of near‐surface structures and the mineral deposits targeted in one of the data sets.