Depthwise separable convolution Unet for 3D seismic data interpolation

In seismic exploration, dense and evenly spatial sampled traces are crucial for successful implementation of most data processing interpretation algorithms. Recently, numerous reconstruction approaches based on deep learning have been presented. High dimension-based methods the benefit making full use signal at different perspectives. However, with transformation dimension from low to high, parameter capacity computation cost training neural network increase significantly. this paper, we introduce depthwise separable convolution instead standard reduce operation Unet 3D missing trace interpolation. The structural similarity (SSIM), L1 hybrid loss function, switchable normalization further improve performance network. comparative experiments synthetic field show that can effectively number parameters intensity interpolation results comparable results.