High-Resolution Multiparameter Seismic Inversion via Large-Field-of-View Reduced-Dimension Global Attention

Abstract

Conventional deep-learning seismic inversion methods typically rely on local receptive fields, failing to exploit large-scale lateral geological continuity. To address this, a large-field-of-view, high-resolution multiparameter inversion framework driven by a lateral reduced-dimension global attention transformer is proposed. A dual-branch complementary supervision strategy integrates full-coverage low-frequency initial models with sparse high-fidelity well logs, effectively mitigating the ill-posedness caused by extreme well sparsity. In addition, the lateral transformer employs spatiotemporal bilinear tokenization and adaptive nonuniform sampling to achieve global spatial attention with minimal computational overhead. By fusing this module with a multiparameter extraction backbone, the network simultaneously models fine vertical details and broad lateral regularities. Validations on a modified Marmousi2 model and a thin-interbedded field dataset demonstrate that the proposed method outperforms state-of-the-art baselines. It delivers superior blind-well accuracy, enhanced lateral consistency, and precise high-frequency recovery, offering a robust paradigm for practical reservoir characterization in sparsely drilled environments.