Reimaging Shallow StructureGreg DePaul ([email protected]) and Jeremy Wood ([email protected])
Stanford Computer Science
Abstract
We graphically highlighted anomalies in seismicreflection data to aid in the detection of signifi-cant formations.
Introduction and Data
Seismic Reflection is the method of sending a pulse(an acoustic pressure wave) into the ground and de-tecting the reflected energy using an array of sen-sors. Using the time it takes for the repeated pulsesto return to the surface, seismic reflection recon-structs the stratigraphic features within the crust.Our goal, after inputting our data set, is to developan algorithm with the ability to distinguish interest-ing features of the shallow structure and appropriatecoloring, based on seismic data. Automating the lo-calization of irregular underground features wouldspeed up the processing and tedious analysis of geo-physical exploration.Our data sets were provided by the National Centersfor Environmental Information under the NationalOceanic and Atmospheric Administration. Depictedon our board are east coast lines 14 and 17 in theGeological Survey of US Atlantic Continential Mar-gin.
Figure 1: Processed Shallow Structure
Before applying machine learning, the data is vi-sually uniform, as depicted in Figure 1. However,upon applying machine learning methods, this be-comes visually revealing of its structure.
Figure 2: Selected patterns from the
Metric
We define the metric as follows. Initially we considerthe horizontal mean:
µxy = 1m2
m∑i=1
m∑j=1||v(x+i,y+j)|| (1)
Where v(x+i,y+j) is the RBG vector of the pixel(x + i, y + j). This makes sense to use since wereare working in black and white data. Then we candefine our metric as a sort of variance:
σxy2 =
m∑i=1
µxy − m∑j=1||v(x+i,y+j)||
2(2)
Clustering
We chose to perform K-means clustering in orderto effectively group similar patterns together. Theprocedure is as follows:1 Perform a motion estimation algorithm over thedata set by progressing a box metric over theentire set.
2 Each box then assigns a metric over thatparticular region.
3 Regions of similar metrics are then clusteredtogether.
4 Repeat until the metric labeling becomes stable.
Results
Figure 3: Selected Results
The first image of Figure 3 represents the overlay performed on Figure 1. The third image of Figure 3represents the overlay performed on the second image of Figure 3.
Discussion
Applying the clustering algorithm requires assum-ing a certain number of features you would want tobe made apparent. In all our examples we typicallyselect 3 or 4 clusters of patterns. Selecting morepatterns results in a "checkerboard" behavior, sim-ply because you have allowed the algorithm to dis-tinguish so many patterns that nearly component isattributed as an interesting feature. This is evidentin Figure 4 which represents an overlay over Figure1 with the assumption of 8 clusters.
Figure 4: Checkerboard Effect
Future Directions
Currently our algorithm relies on the motion estima-tion to lie exactly on the block grid. In the future,it would be better to allow more maneuverability. Itwould also be beneficial to include a technique thatcould handle sloped inclines, as evident in the thirdimage of Figure 3.
References
• Lillie, R. J. (1999). Whole earth geophysics: Anintroductory textbook for geologists and geophysicists.Upper Saddle River, N.J: Prentice Hall. Pages 118 through127.
•Evans, Wecht. Machine Learning for Seismic ReflectionData. Retrieved from Harvard Faculty of Arts and Sciences:http://people.fas.harvard.edu/~wecht/Kevin_J_Wecht/Data_Analytics_files/Final_Report.pdf
•NOAA. (1977). ECT14-17 [Digital SEG-Y]. Retrieved fromhttps://www.ngdc.noaa.gov/mgg/trk/trackline/coral_seal/ect14-17/seismics/data/digital/
