A Deep Semi-Supervised Learning Approach for Seismic Reflectivity Inversion

In seismic inversion, supervised learning is highly effective when abundant paired data are available, yet such conditions are often not met in this domain due to the scarcity of well data. This scarcity hinders the effectiveness of supervised learning in accurately inverting seismic data. To address this challenge, we explore the potential of semi-supervised learning, which requires a careful balance between supervised and unsupervised tasks. We propose a dual-network structure for semi-supervised learning, enabling predictive mapping and reconstruction paths. Conventional deconvolution methods assume a stationary seismic wavelet across the seismic section, an oversimplification that fails in areas where localized wave propagation effects cause significant deviations, leading to inaccurate estimations of subsurface properties. In response, we propose a practical methodology that leverages on building an initial robust model trained on synthetic reflectivity-seismogram pairs, and subsequently improving the model generalizability through semi-supervised transfer learning. Our methodology, tested on both a synthetic 2-D wedge model and the Marmousi2 dataset, not only outperforms conventional inversion algorithms but also outperforms supervised learning approach. It excels in sparsity recovery of reflectivity estimates, maintains high accuracy in noisy conditions, and ensures spatial continuity between adjacent seismic traces.