Progressive Self-Supervised Inversion With Multiscale Strategies for Resolving Multiparameter Crosstalk in a Marine Data Case

Parameter crosstalk in multi-parameter full-waveform inversion (FWI) remains a critical challenge, often degrading the accuracy of subsurface models, particularly for density estimation. To address this, we propose CNN-MMI (Convolutional Neural Network-based Multi-scale Multi-parameter Inversion), a self-supervised deep learning framework for simultaneous inversion of P-wave velocity, S-wave velocity, and density from elastic seismic data. Unlike existing self-supervised CNN-based approaches that target single-parameter acoustic inversion, CNN-MMI introduces three fully decoupled independent networks for Vp, Vs, and ρ respectively, severing the interparameter gradient contamination pathways that cause crosstalk at the architectural level. Following the self-supervised paradigm, the framework iteratively generates pseudo-labels from progressively refined prior models, eliminating the need for large labeled datasets. A hierarchical multi-scale workflow sequentially incorporates low-to-high-frequency data to enhance deep-layer resolution and reduce initial model dependency. Furthermore, a parameter-specific scaling coefficient inspired by FWI step-length control is introduced to independently regulate the convergence behavior of each network, improving both inversion stability and accuracy. Numerical experiments on the Sigsbee and Marmousi models, as well as a field marine dataset, demonstrate that CNN-MMI significantly outperforms conventional elastic FWI in suppressing parameter crosstalk and achieving accurate multi-parameter reconstruction, particularly for density inversion.