Seismic Facies Segmentation Using Atrous Convolutional-LSTM Network
Palavras-chave:
Seismic Classification, Semantic Segmentation, Deep Learning, Convolutional Recurrent Networks, Atrous ConvolutionalResumo
In this paper, we provided new end-to-end approaches to the task of seismic image segmentation, as
human analysis requires a lot of effort and time due to the large pixel dimensions. Given that seismic dataset
contains temporal information along its axis (inline and cross-line), we also proposed the use of recurrent neural
networks (RNN) together with convolutional layers. After several experiments, we found that the application
of crop and rescale (zoom) as a data augmentation technique, as well as the use of focal loss, shows significant
improvements in performance and training speed. Our best LSTM-based model achieved a very close to the best
one using fewer parameters.
