Impedance inversion based on deep learning considering structural constraints

Abstract

Summary Seismic impedance inversion aims to provide high-resolution impedance data for subsurface spatial tectonic analysis and reservoir prediction. Traditional inversion methods tend to impose people’s prior knowledge, such as sparsity, into the modeling of inversion processes. Convolutional neural networks(CNNs) as a data-driven method have achieved excellent performance in the field of seismic impedance inversion. However, CNNs only extract local features of labeled impedance, which cannot learn the global and deep feature information interaction, and usually ignore the condition that the inverted impedance should be consistent with the structure of the label. In this paper, we propose a pure Transformer network (STUnet) with a structure similar to Unet for impedance inversion, and introduced the loss of structural similarity to guide the training process. The experimental results on synthetic data show that the proposed STUnet has better performance than other inversion methods based on deep learning.