Preserving High-Frequency Geological Features in Large-Scale Super-Resolution of Shale-Oil Well Logs via Multistage Knowledge Transfer

Summary Traditional super-resolution methods struggle with large-scale logging curve reconstruction, particularly losing critical high-frequency geological information at magnifications exceeding 50X. In this paper, we present a cascaded deep learning framework employing two-stage knowledge transfer for large-scale super-resolution while preserving geological feature integrity. The proposed method first constructs a random forest (RF)-based nonlinear mapping model for knowledge transfer from high-resolution reference data to conventional resolution data, generating multiscale training samples. A dual-branch cascaded network architecture is then designed where the global branch handles contour reconstruction and the local branch focuses on detail extraction, with a graph interaction module (GIM) enabling dynamic fusion of multiscale features. Additionally, a specialized spike enhancement module preserves critical geological features during progressive 2X reconstruction strategies, achieving cumulative 64X super-resolution reconstruction. Validation experiments on five logging curve types from four Daqing Oilfield wells demonstrate superior performance across all tested curves, achieving peak signal-to-noise ratios of 19.33–23.19 dB and an average high-frequency matching rate of 69%, significantly outperforming existing approaches in large-scale super-resolution reconstruction.