High-Resolution Seismic Data Reconstruction Driven by Multi-Source Data and Physical Model

This study proposes a high-resolution seismic data reconstruction method that integrates multi-source data and a physical model within a semi-supervised learning framework. Traditional model-driven methods, such as least-squares deconvolution and inverse Q filtering, offer good generalizability but rely on physical approximations that are challenging to apply in practice. Data-driven methods, particularly those using deep learning networks, provide strong nonlinear mapping capabilities and can incorporate broadband well-logging data. However, the scarcity of well-logging data can hinder network training, leading to poor generalization. To address this, we invert the band-limited reflection coefficient series from post-stack seismic data, convolve them with broadband wavelets, and integrate well-logging data and convolutional models into the network. This approach improves the model’s generalizability, physical interpretability, and high-resolution processing capabilities. The inclusion of broadband well-log data significantly enhances the resolution of the inversion results and broadens the seismic data’s frequency spectrum, thus facilitating better identification of thin layers and providing a more accurate representation of subsurface structures.