Progressive Pseudo-label Self-supervised Waveform Inversion and Imaging Based On Multiscale Strategies: A Marine Data Case Application

Full waveform inversion (FWI) and reverse time migration (RTM) have shown great potential in subsurface imaging, yet they remain challenged by migration artifacts and limited deep illumination. A novel PPS-DLI (Progressive Pseudo-label Self-supervised Deep Learning Inversion) framework is proposed that integrates deep learning with multi-scale waveform inversion and imaging. The approach introduces a progressive model-based pseudo-label self-supervised strategy that iteratively refines velocity models by leveraging predictions from previous network stages. Combined with a frequency-based multi-scale inversion scheme, the method enables robust background velocity reconstruction at low frequencies and detailed imaging at higher frequencies. Numerical experiments demonstrate that PPS-DLI significantly reduces the migration artifacts, provides additional structural information with better illumination, improves the deep layers and sub-salt imaging, and preserves amplitude fidelity for enhanced lithological interpretation. These advantages position PPS-DLI as a powerful tool for high-resolution and noise-resistant seismic imaging.