Enhancing seismic data quality: A machine learning approach to denoising and signal damage reduction Available to Purchase

We propose a denoise workflow comprising a supervised ML (Machine Learning) model applied in the common shot domain and a self-supervised ML signal-add back model in the common channel domain. The supervised ML-based denoise (Brusova et al., 2021) learns from training data containing recorded noise and provides robust, high-quality results that can successfully tackle various noise types in a single pass without requiring complex parameterization. However, some signal leakage can occur, producing primary damage. We demonstrate a self-supervised signal add-back technique based on the blind-trace network (Birnie et al., 2021) that mitigates the primary damage and produces a complete denoise solution. The technique is applied to various vintages of offshore streamer data. The results from our novel workflow show significant improvements in recovery/denoise of low frequency (<3 Hz) on legacy streamer data, which is critical for success in deghosting/broadband processing and FWI.