Automatic event picking in prestack migrated gathers using a probabilistic neural network
|
|
- Reynard Horn
- 5 years ago
- Views:
Transcription
1 Automatic event picking in prestack migrated gathers using a probabilistic neural network Michael Glinsky, Bill Butler, Sandhya Devi and Jim Robinson Shell E&P Technology Company Greg Clark Lawrence Livermore National Laboratory Peter Cheng and Gary Ford University of California, Davis 10/8/97 PNN event picking 1
2 Agenda introduction and problem definition overview of processing scheme detailed example of algorithm application summary and future work 10/8/97 PNN event picking 2
3 There is a velocity analysis bottleneck in prestack migration velocity model v(x,y,z) update the velocity model ray trace form prestack migrated gathers pick events 10/8/97 PNN event picking 3
4 Why does this bottleneck exist? iterative process of converging to velocity model is limited by picking events picking is currently performed manually 2D datasets, >10,000 traces not feasible with larger datasets 3D datasets, >1,000,000 traces done manually for quality control eliminate loop skips goal is to automate the event picking process manual picks only for training and context reduce manual picking to less than 0.1% of data 10/8/97 PNN event picking 4
5 Neural networks automate picking and tracking of events on CRP panels (x,y) shot (x,y) geophone z z θ θ CRP panel 10/8/97 PNN event picking 5
6 Several basic ideas and issues guide this work use the simplest techniques that prove to be effective use approach of: signal / image processing machine vision supervised learning techniques feature analysis is the key GIGO exploit information about: prior knowledge from human experts spatial context 10/8/97 PNN event picking 6
7 Processing flow raw data event picks feature definition feature selection voxel classification valley finding and constraints 10/8/97 PNN event picking 7
8 Algorithms used in the processing flow feature definition 2D Gabor transforms semblance amplitude histogram proximity feature selection sequential forward selection voxel classification probabilistic neural network (PNN) connected components valley finding and constraints size continuity 10/8/97 PNN event picking 8
9 The algorithm used event features PNN threshold and connected components proximity features PNN threshold valley finding and constraints event image 10/8/97 PNN event picking 9
10 Prestack migrated data (raw data) deepwater GOM 2D dataset JIGSAW prospect 10/8/97 PNN event picking 10
11 Useful features of the raw data statistical moment mean, etc. moment over red box semblance Gabor transforms magnitude & phase 2 scales 4 slopes /8/97 PNN event picking 11
12 Event feature images are formed Gabor magnitude (large, 0 ) Gabor phase (large, 0 ) standard deviation Gabor magnitude (large, 50 ) mean semblance 10/8/97 PNN event picking 12
13 Features are ranked via Sequential Forward Selection algorithm )! Gabor phase (rad/ Gabor magnitude (arb. units) large scale, O distance between event and background cluster used!bhattacharyya distance GM small,0 GM large,0 raw GM small,-25 GP small,-25 GP small,0 GP large,0 GM = magnitude of Gabor transform GP = phase of Gabor transform 10/8/97 PNN event picking 13
14 Posterior probability image using event features as input training set 107 events 100 background 20 out of 468 CRPs 0.5% of picks probability of correct classification 89% to 96% 1 probability of event 0 10/8/97 PNN event picking 14
15 Proximity features allow for human contextual input offset time probability of eve probability close to picked event 10/8/97 PNN event picking 15
16 Posterior probability image using proximity features as input 193 picks used 1% of picks mask for constraining search space 10/8/97 PNN event picking 16
17 Binary labeled image 10/8/97 PNN event picking 17
18 Connected components labeled image 10/8/97 PNN event picking 18
19 Event image size one time / offset / cloud continuous max posterior probability 10/8/97 PNN event picking 19
20 Automated picks compared to human picks Human edited FLIRT picks Automated picks 10/8/97 PNN event picking 20
21 PNN prevents loop skips in low signal to noise data FLIRT only 10/8/97 PNN event picking 21
22 Neural network picker applied to 2D GOM dataset (JIGSAW) 0.5 s 0 32 km probability of event 1 neural network picks expert picks 0 5 s 10/8/97 PNN event picking 22
23 Neural network picks compare well with expert picks 2 s 0 8 km probability of event 1 neural network picks expert picks s 10/8/97 PNN event picking 23
24 Neural network sometimes consistently picks on different loop 3.1 s 9 km 26 km probability of event 1 neural network picks expert picks 0 5 s 10/8/97 PNN event picking 24
25 Neural network picker is not as aggressive as expert 0 0 subpoint 9 km time of event pick 2.2 s offset 4 km expert 3 s PNN 10/8/97 PNN event picking 25
26 Computer time needed 15 ms / voxel / feature interpreted MATLAB PowerBook 5300c 7 features used (6 Gabor, raw data) 60 µs / voxel / feature Sparc Ultra 1 compiled C++ 7 days for 4 OCS blocks 10/8/97 PNN event picking 26
27 Summary and future work Gabor transform captures character of event better than semblance and amplitude histogram PNN combines features into best guess prevents loop skips proximity is a way to quantify where to look could enable 3D PSDM cost reduced from $75,000 to $6,000 (4 OCS blocks) cycle time reduced from 12 weeks to 1 week improve robustness of inversion (more picks) further evaluation needed on other datasets tracking of PNN result needs to be improved (aggressiveness) 10/8/97 PNN event picking 27