Image True Mask Predicted Mask Predicted Mask
(epoch 5) (epoch 50)
Abstract:
This project utilized a neural network to do a 3D segmentation of sub-earth geological images. The goal of image segmentation is to partition input digital image to meaningful segments. The geological 3D image shows salt rocks surrounded by non salt rocks often referred as sediments. The image of sediment are characterized by rock formation in layers and the image of salt rocks are characterized by irregular shapes and strong reflection on the exterior boundaries. The neural net is trained to identify salt bodies from surrounding background rocks. Both 2D and 3D convolutional neural networks are trained and tested.
Conclusion:
UNet is highly effective in this synthetic geological image segmentation. One possible explanation is the data set has simple geological structure comparing to distinguish 2D and 3D segmentation.
Hospital rely on radiologists to mannual annotate medical images. With the rapid progress made in image recognition, object detection and segmentation int the field of computer vision, they can now rely more on AI and machine learning models to take some labor intensive work. Similarly in energy sector, trained geologists mannually annotate volume of geological images in 2D and 3D. A good understanding of the sub earth geology is required before drilling to maximize the prospect and minimize the drilling hazards. It requires high accuracy for identifying different geobodies underneath the earth surface.
3D and 2D segmentation have similar performace. 2D model obtains a validation accuracy of 0.9809 and a validation loss of 0.0034 after 50 training epochs. 3D model obtains a validation accuracy of 0.9867 and a validation loss of 0.0054.
- Convert the input SEGY format file into numpy and write as csv files by running
python src/sgy_to_np.py - Train the 2D model by running
python src/2d-segmentation-32x32.py - Train the 3D model by running
python src/3d-segmentation-32x32x32.py
The open source data were computed as part of the Advanced Computational Technology Initiative, in partnership with the United States Department of Energy National Laboratories and Technology Centers[https://wiki.seg.org/wiki/Open_data#SEG.2FEAGE_3D_modeling_Salt_Model_Phase-C_1996]. The synthetic data is clean than the real field data and the accurary of prediction is likely to decrease.
-
Obspy
-
scikit-image
-
Matplotlib
-
Tensorflow/Keras
-
Python
-
Numpy
- Compare the performance of 2D and 3D segmentation on real data
- Test other activation\loss functions
- Modification of neural network structure
I'd like to thank Dan Rupp, Brent Goldberg and Joseph Gartner for their guidance, feedback and technical support for this project.
U-Net: Convolutional Networks for Biomedical Image Segmentation
https://arxiv.org/abs/1505.04597
Olaf Ronneberger, Philipp Fischer, Thomas Brox
SaltSeg: Automatic 3D salt segmentation using a deep convolutional neural network
https://library.seg.org/doi/10.1190/int-2018-0235.1
Yunzhi Shi, Xinming Wu, and Sergey Fomel
Understanding Semantic Segmentation with UNET
https://towardsdatascience.com/understanding-semantic-segmentation-with-unet-6be4f42d4b47
Harshall Lamba
Semi-Supervised Segmentation of Salt Bodies in Seismic Images using an Ensemble of Convolutional Neural Networks
https://arxiv.org/abs/1904.04445
Yauhen Babakhin, Artsiom Sanakoyeu, Hirotoshi Kitamura
ObsPy: A Python Toolbox for Seismology M. Beyreuther, R. Barsch, L. Krischer, T. Megies, Y. Behr and J. Wassermann (2010) http://www.seismosoc.org/Publications/SRL/SRL_81/srl_81-3_es/