Weakly Supervised Learning for 2D Seismic Impedance Inversion

Seismic impedance is an important attribute reflecting the subsurface structure. Recently, deep learning methods have been widely applied in seismic inversion. However, limited labels from well-log data often obstacle to its practical applications. In addition, 1D impedance inversion based on deep learning usually has poor spatial continuity of the results due to the lack of spatial information, and 2D impedance inversion often requires low-frequency background. Aiming at the above problems, we propose a 2D seismic impedance inversion method based on weak supervision. Using the idea of weak supervision, we input the seismic profile and corresponding time data into deep neural networks, and train the networks in a weakly supervised manner according to the logging labels as constraints to obtain the target impedance profile. In addition, the loss item based on the Robinson model and the total variation regularization item are designed to constrain the spatial continuity of inversion results and the smoothness of each trace result. The proposed method can realize 2D impedance inversion in a weakly supervised manner without knowing the low-frequency background, and the inversion results maintain good spatial continuity. The experiments on synthetic data demonstrate the effectiveness of our method.