A Benchmark Dataset for Semi-Automatic Seismic Interpretation Based on a New Zealand's Seismic Survey

Open access to curated datasets positively impacts on scientific research of machine learning and deep learning techniques. It is a fact that benchmarks and public datasets prepared for data science assist researchers interested in evaluating, testing, and building new data-driven methodologies for specific domain areas.

In geosciences, there has been a remarkable growth of public datasets arranged to address machine learning challenges related to the oil and gas industry, particularly for reserves exploration and data interpretation. 

For these reasons, we present the Taranaki dataset, which is a collection of seismic horizons interpreted for a seismic stratigraphic interpretation study in the Taranaki Basin, offshore New Zealand. This data comprises fourteen seismic horizons that mark stratigraphic discordances in the Tui-3D seismic dataset. We annotated five seismic horizons on 33 inline sections and nine horizons on 19 crossline sections.

Besides, we present the results of a series of experiments that compare a method of interpolation and a method of deep learning for seismic segmentation. The deep learning experiments evaluated the result of different image tile sizes to train the model, which is presented separately in this dataset. 

Finally, we evaluated both methodologies to interpret the horizons of this dataset in selected seismic sections. Also, we assessed the absolute error of each method with the ground truth interpretations proposed in this dataset.