[Paper Review] Netherlands Dataset: A New Public Dataset for Machine Learning in Seismic Interpretation

Reinaldo Mozart Silva, Laís Baroni+4 more|arXiv (Cornell University)|Mar 26, 2019

Seismic Imaging and Inversion Techniques39 citations

TL;DR

The paper introduces the Netherlands interpretation dataset for seismic ML, detailing 9 horizons and ~190k labeled seismic images derived from public Netherlands F3 data, and demonstrates two DL tasks (classification and semantic segmentation) using the dataset.

ABSTRACT

Machine learning and, more specifically, deep learning algorithms have seen remarkable growth in their popularity and usefulness in the last years. This is arguably due to three main factors: powerful computers, new techniques to train deeper networks and larger datasets. Although the first two are readily available in modern computers and ML libraries, the last one remains a challenge for many domains. It is a fact that big data is a reality in almost all fields nowadays, and geosciences are not an exception. However, to achieve the success of general-purpose applications such as ImageNet - for which there are +14 million labeled images for 1000 target classes - we not only need more data, we need more high-quality labeled data. When it comes to the Oil&Gas industry, confidentiality issues hamper even more the sharing of datasets. In this work, we present the Netherlands interpretation dataset, a contribution to the development of machine learning in seismic interpretation. The Netherlands F3 dataset acquisition was carried out in the North Sea, Netherlands offshore. The data is publicly available and contains pos-stack data, 8 horizons and well logs of 4 wells. For the purposes of our machine learning tasks, the original dataset was reinterpreted, generating 9 horizons separating different seismic facies intervals. The interpreted horizons were used to generate approximatelly 190,000 labeled images for inlines and crosslines. Finally, we present two deep learning applications in which the proposed dataset was employed and produced compelling results.

Motivation & Objective

• Motivate the need for high-quality labeled datasets in seismic interpretation for DL success.
• Introduce the Netherlands interpretation dataset derived from public Netherlands F3 seismic data.
• Provide dataset characteristics including horizons, images, and how labels were generated.
• Demonstrate two deep learning applications (classification and semantic segmentation) using the dataset.

Proposed method

• Reinterpret 3D Netherlands F3 seismic data to define 9 horizons based on seismic facies.
• Generate ~190,000 labeled seismic images by intersecting horizons with seismic lines to produce labeled intervals.
• Create classification tiles by dividing images into 64x25 tiles and assigning the predominant horizon class; balance the training data.
• Apply deep learning to classify rock-layer textures achieving 81.6% accuracy on a test split.
• Apply transfer-learning/segmentation approaches to produce pixel-wise (semantic) labels and report IoU metrics in related works.

Experimental results

Research questions

• RQ1 Can DL models classify rock-layer textures from seismic tiles with high accuracy using the Netherlands interpretation dataset?
• RQ2 Can semantic segmentation of seismic images achieve high mean IoU on horizon-based labels using this dataset?
• RQ3 How does tiling strategy and class balancing affect DL performance on seismic interpretation tasks?

Key findings

• Classification on the provided test set achieved 81.6% accuracy (confusion matrix shown in the paper).
• The dataset enables high-performance semantic segmentation in related works, with reported mean IoU values exceeding 90–98% under various setups (from referenced works).
• Tiles and labeling approach balance classes and allow standard metrics to be applied for evaluation.
• The dataset comprises 9 horizons, 651 inlines, 951 crosslines, and ~190k labeled images suitable for DL benchmarking.
• Two DL applications (rock-layer classification and semantic segmentation) have been demonstrated using this dataset.