Identification of horizons in seismics using convolutional neural network

Resumo

Seismic structural interpretation is an essential step in exploring and producing hydrocarbon reserves.
This interpretation requires identifying geological features such as facies, horizons, and faults in the region of
interest. The manual identification of these features is a time-consuming task. Convolutional neural networks
(CNN) are widely used in computer vision problems, yielding excellent results in many situations, including the
seismic interpretation process. This work studies supervised convolutional neural networks to segment horizon
lines delimiting seismic facies based on seismic amplitude. We evaluate our proposal using the F3 block with the
seismic facie annotations. In the original dataset annotations, the labels were annotated areas for each seismic
facies, so this set of annotations was changed from a multiclass problem to a binary, considering only the boundary
between one seismic facie to its neighbor. The horizon prediction uses the ResUnet network, a combination of
Unet with residual blocks, designed to obtain high performance with fewer parameters. Some loss functions are
analyzed to optimize the segmentation result. Generalized dice loss and Focal Tversky loss functions yield best
results in our experiments. The Dice metric reached an index above 50% with the Focal Tversky loss function,
showing promising results.