SeisDeepNET: An extension of Deeplabv3+for full waveform inversion problem

Determining velocity model in complex geology settings such as environments with a salt body, is one of the significant challenges in seismic data analysis. One highly accepted way to determine the seismic velocity model is full waveform inversion (FWI). FWI suffers from number of problems such as cycle skipping and high computational cost. In this paper we propose a deep network, so called SeisDeepNET, to provide high resolution velocity model. Since FWI is highly nonlinear and dimensions of the seismic field data (as input of the network) differ from the dimensions of velocity model (as output of the network), the proposed deep learning method requires an encoder-decoder network with efficient architecture in order to effectively approximate the inverse operator and estimate model details. The encoder section of the proposed network is in accordance with the encoder of the DeepLabv3+ network, but it has different architecture in the decoder section. In order to enlarge the size of image in the decoder section, we incorporated 2D transposed convolution and bilinear interpolation separately into the decoder of the proposed network and assessed their performances on 1600 simulated 2D velocity models. The results showed that 2D transposed convolution, due to its learnable parameters, preserved the model details better. New architecture enabled the network to generate superior velocity model than counterpart networks like Unet, even with less training data. We also evaluated the effect of MSE and MAE metrics on the estimated velocity models of the network in the presence of salt body.