Fault Segmentation on Seismic Data Using CNNs

Mapping fault planes using seismic images is a crucial and time-consuming step in hydrocarbon prospecting. Conventionally, this requires significant manual efforts that normally go through several iterations to optimize how the different fault planes connect with each other.​
Many techniques have been developed to automate this process, such as seismic coherence estimation, edge detection, and ant-tracking, to name a few. However, these techniques do not take advantage of the valuable experience accumulated by the interpreters. ​ Using CNNs, however, is a technique to automatically detect and map fault zones using 3D seismic images in a similar fashion to the way done by interpreters.

Dataset

The training data consisted of 200 synthetic seismic volumes and corresponding volumes containing fault labels. The validation data consisted of another 20 synthetic seismic volumes with corresponding volumes containing fault labels.

The training dataset was further expanded by inclusion of 5 seismic sections from F3 dataset from the Netherlands.

The labels for the seismic section were created manually using adobe illustrator.

Finally, 10,000 tiles(128x128) were created from these 5 seismic sections and included into the dataset.

Model

Our problem required a simplified U-net.

Accuracy

Several models were trained. To further increase the accuracy of the model, 5 inline sections from F3 dataset from the Netherlands were taken at random. Fault labels were created manually and each section was broken into several (128x128) tiles and fed into the model. Resultant training dataset, including the synthetic dataset, consisted of 86800 different seismic sections with their corresponding labels. As a result, accuracy furthur increased. Final Accuracy of the model: 89.21%

Results

The model was tested on an actual seismic dataset.

Process at a glance