1. Automatic interpretation of salt geobodies Adam Halpert ExxonMobil CEES Visit 12 November 2010 Stanford Exploration Project

2. Why automate? Save time Save time Manual salt-picking is tedious, time-consuming Manual salt-picking is tedious, time-consuming Major bottleneck for iterative imaging/model-building Major bottleneck for iterative imaging/model-building Maximize expertise Maximize expertise Allow experienced interpreters to focus on more complex geological problems Allow experienced interpreters to focus on more complex geological problems Improve health? Improve health? Manual picking contributes to ergonomic strain Manual picking contributes to ergonomic strain

3. Automation strategies 1.“Traditional” horizon auto-trackers

4. Example image

5. Auto-tracking SEED POINTS

6. Auto-tracking

7. Auto-tracking

8. Auto-tracking

9. Automation strategies 1.“Traditional” horizon auto-trackers -Still requires significant user input -Can get “lost” at local horizon discontinuities 2.Global image segmentation

10. Graph segmentation Any image (seismic or otherwise) can be thought of as a graph Any image (seismic or otherwise) can be thought of as a graph – Each pixel is a node or vertex of the graph – Vertices are connected by edges Each edge is assigned a weight Each edge is assigned a weight – Usually, a measure of similarity or dissimilarity between pixels A segmentation (or graph partition) groups these edges into subsets of the image A segmentation (or graph partition) groups these edges into subsets of the image – Edges between vertices in the same subset (segment) will have low weights – Edges between vertices in different segments will have higher weights – (or vice versa)

11. Pairwise Region Comparison Felzenszwalb and Huttenlocher (2004): Efficient graph-based image segmentation Felzenszwalb and Huttenlocher (2004): Efficient graph-based image segmentation Two major goals Two major goals Capture global aspects of the image Capture global aspects of the image Be highly efficient (~linear with number of pixels) Be highly efficient (~linear with number of pixels) Construct edges between each pixel and its neighboring pixels Construct edges between each pixel and its neighboring pixels Weight the edges based on the highest-intensity pixel between the two endpoints Weight the edges based on the highest-intensity pixel between the two endpoints

12. The algorithm 1.Create the edges and store their location and weight value 2.Sort the m graph edges by increasing edge weight 3.For initial segmentation S 0, each pixel/vertex is its own segment 4.For each graph edge q in the sorted list from Step 1, if the difference criterion is met, S q is created by merging the two pixels or regions the edge connects Otherwise, do nothing Otherwise, do nothing 5.S m is the segmented image [C++ implementation]

13. Example 1: 2D Field

14. Segmentation result 150 x 500: 1 sec

15. Example 2: 2D Synthetic

16. Segmentation result

17. Pick segments to merge      

18. Merged result 1000 x 2760: 41 sec

19. Example 3: 3D Field

20. Segmentation result 114 x 534 x 51: 39 sec

21. Automation strategies 1.“Traditional” horizon auto-trackers -Still requires significant user input -Can get “lost” at local horizon discontinuities 2.Global image segmentation -PRC method requires little user input, but can offer flexibility -Accurately and efficiently segments 2D and 3D images

22. Planned enhancements Segmentation with multiple seismic attributes Segmentation with multiple seismic attributes Increased opportunity for user input/prior knowledge inclusion Increased opportunity for user input/prior knowledge inclusion Ultimately: link segmentation results with velocity updates and imaging Ultimately: link segmentation results with velocity updates and imaging