This project was created over two days as part of the OGA/Agile Hackathon in Aberdeen, UK, November 16-18th 2018.
The aim of the project was to use machine learning to identify whether images of seismic data contained particular feature classes. In this case four features were extracted from OGA lines in the mid-north sea.
- Salt
- Gas Chimney
- Flat Stratigraphy
- Noise
- Create a training dataset of images by capturing screen shots of data all at a common scale. The images are not of equal size.
- Label the images so they can be categorized and put into a Pandas DataFrame
- Noise images were created by subsampling a few large noise screenshots with random patches.
- Randomly split the data within categories into training and test data.
- Concatenate the different categories together.
- Output the data to pickle.
- Import data and create new features using image processing. (Edge detection and Gabor filter).
- Output expanded DF Pickle.
- ML Feature Extraction
- Training
A large set of random patches can be generated from the input images in the process_to_input notebook. This is done using scikit learn image_extract_patches_2d
.
Quentin Corlay; Amaechi Halim; Tony Hallam; Saleem Ramy; Shaji Matthew; Elia Gubbala; Zh Cui