seismiQB is a framework for research and deployment of deep learning models on post-stack seismic data. It allows to

• convert SEG-Y to a compressed and quantized data formats, that take 5x less disk space and load slices up to 40 times faster
• work with a number of labels: horizons, faults, facies (both 2d and 3d)
• load crops of data and create segmentation masks to train neural networks, following defined distribution of location generation
• augment seismic images and masks with both usual CV methods like flip, rotate and elastic_transform, as well as to compute geological attributes: phases, frequencies, instantaneous_amplitudes
• define complex neural networks with simple and intuitive configurations: just a few lines of code are enough to define models ranging from vanilla UNet to most sophisticated versions of modern EfficientNets
• apply ready-to-use models and pipelines to detect horizons, faults, alluvial fans and fluvial channels
• export predicted entities to a convenient formats like CHARISMA and FAULT_STICKS for an easy validation by geophysicists

With pipenv:

pipenv install git+https://github.com/gazprom-neft/seismiqb.git#egg=seismiqb

With pip:

pip3 install git+https://github.com/gazprom-neft/seismiqb.git

After that just import seismiqb:

import seismiqb

To get the developer version, run

git clone --recursive https://github.com/gazprom-neft/seismiqb.git

Working with SEG-Y cubes with various indexing headers (e.g. pre-stack and post-stack).

Our dedicated Horizon class is capable of loading data from multiple geological formats, computing a wealth of statistics of it, and a lot more. We also provide interfaces for other types of geological bodies like faults, facies and labels in pre-stack cubes.

A wrapper aroung geometries and labels, that can generate data from random labeled locations from multiple cubes and apply both geological and computer vision augmentations.

In order to evaluate our results (particularly predicted horizons), we developed a few seismic attributes to assess quality of seismic cubes, sparse carcasses and labeled surfaces.

This model spreads a very sparse hand-labeled carcass of a horizon to the whole cube spatial area by solving a task of binary segmentation.

Enlarge picked (possibly by other models) horizons to cover more area.

Applying the multi-class segmentation model to the task of horizon detection. Note that the model was developed with older seismiQB versions and does not work anymore.

Application of a model, trained on a set of cubes, to a completely unseen data.

Please cite seismicqb in your publications if it helps your research.

Khudorozhkov R., Koryagin A., Tsimfer S., Mylzenova D. SeismiQB library for seismic interpretation with deep learning. 2019.

@misc{seismiQB_2019,
  author       = {R. Khudorozhkov and A. Koryagin and S. Tsimfer and D. Mylzenova},
  title        = {SeismiQB library for seismic interpretation with deep learning},
  year         = 2019
}