Background The efficient application of depth migration methods has enabled the extensive implementation of depth-domain seismic data interpretation. Previous studies on the impacts of time-to-depth conversion and velocity models merely determine that an increase in depth corresponds to reduced dominant wavenumber of seismic waves and waveform stretching, without considering the changes in amplitude and phase.

Objective This study aims to accurately characterize the waveforms of nonstationary seismic signals in the depth domain and enhance the impedance inversion accuracy.

Methods First, based on energy dissipation and frequency dispersion effects during the propagation of seismic waves, this study investigated the complex mapping relationship between nonstationary seismic reflected waves and formation impedance and introduced the Q model that reflected the absorption attenuation of strata for seismic waves. Nonstationary convolution formula for seismic reflection data were proposed to simultaneously describe the amplitude attenuation, phase distortion, and decrease in the primary wavenumber of seismic waves. Accordingly, the impedance inversion equation was established under spatial constraints. Second, this study estimated the Q model for strata using deep learning technology. The multi-head self-attention mechanism was introduced into the network structure, allowing for the extraction of accurate attenuation characteristics of depth-domain seismic signals. The assumption of a known Q model in the conventional inversion process was abandoned. Instead, some synthetic data were employed for network training and validation, ensuring the convenient implementation of the estimation method. Third, the depth-varying seismic wavelets were calculated using Q values yielded from the network, and the multi-trace impedance inversion method based on the sparsity constraint from the l_p norm was employed. Ultimately, the high-resolution absolute impedance data volume in the depth domain was determined.

Results and Conclusions  Validation using the Pluto model demonstrated that the Q model and nonstationary inversion achieved using deep learning technology yielded a relative error in impedance of 13.7%, suggesting significantly improved inversion accuracy compared to the results of conventional stationary inversion (48.2%). Tests using field seismic data from an exploration area of the Jinzhong coalfield indicate that deep-domain nonstationary seismic inversion technology can capture the physical property parameters of subsurface media more intuitively and accurately. The impedance determined using the inversion technology showed a high similarity of 0.9488 to the impedance curve determined based on log data, thus avoiding instability caused by multiple processing steps such as inverse Q filtering and recursive inversion. The results of this study provide depth-domain stratigraphic information for reference in subsequent seismic interpretation.