MSWANet: multiscale dilated convolution and wavelet-guided attention network for seismic impedance inversion

Abstract Based on the convolutional-model assumption, traditional seismic impedance inversion often suffers from smoothing and loss of detail due to limited resolution, sensitivity to noise, and difficulties in capturing multiscale geological features. To overcome these limitations, a novel impedance inversion framework named Multiscale Dilated Convolution and Wavelet-Guided Attention Network (MSWANet) is proposed, which integrates multiscale dilated convolutions and wavelet-guided attention within a semi-supervised paradigm. Within this framework, impedance loss and seismic reconstruction loss are jointly optimized using both labeled and unlabeled data, thereby improving stability even when labeled samples are limited. The network expands the receptive field, enhances multiscale feature extraction, and highlights faults and stratigraphic interfaces while suppressing noise. Experiments on the Marmousi2 model demonstrate that MSWANet outperforms traditional Convolutional Neural Network–Gated Recurrent Unit (CNN–GRU) networks in thin-bed resolution, fault recovery, and low-impedance anomaly detection. Application to the Netherlands Offshore F3 Block further shows that stratigraphic continuity is consistently recovered across profiles of different dimensions, underscoring strong geological plausibility. When applied to field data with sparse well control, MSWANet yields reliable matches to well-log data and realistic impedance distributions, effectively mitigating the smoothing and detail loss inherent in conventional approaches.