Seismic fault detection based on multi-attribute support vector machine analysis

Transcription

1 INT 5: Fault and SEG 2017 Seismic fault detection based on multi-attribute support vector machine analysis Haibin Di, Muhammad Amir Shafiq, and Ghassan AlRegib Center for Energy & Geo Processing (CeGP) Georgia Institute of Technology {hdi7, amirshafiq, alregib}@gatech.edu September 27, 2017

2 Outline Motivation Workflow description SVM and MLP analysis CNN analysis Conclusions 2

3 Motivation Fault interpretation is important for subsurface interpretation and reservoir characterization from 3D seismic data Lots of methods/algorithms available: Attributes: coherence, curvature, flexure, likelihood, and so on Techniques: ant-tracking, Hough transform, time wrapping, motion vectors, and so on Seismic data size is significantly increasing, which requires more interpretation efforts; Machine learning is efficient in big data analysis How is its performance on seismic fault detection? 3

4 Problem definition from machine learning Image Classes Classification Natural 1. Person 2. Vehicle 3. Others Seismic 1. Fault 2. Non-fault 4

5 Multi-attribute SVM/MLP analysis Workflow 5

6 Step 1: attribute selection Group Attribute Measurement Fault Dip Steeply dipping Juxtaposition of positive and Curvature Geometric Lateral variation of the negative curvatures attribute geometry of seismic reflectors Peak flexure with two subtle Flexure sidelobes Geometric fault High values Coherence Low coherence Sobel edge High values Lateral changes in seismic Edge-detection Semblance Low semblance waveform and/or amplitude attribute Canny edge High values using various operators Similarity Low similarity Variance High variance GLCM contrast Statistical analysis of local High contrast GLCM homogeneity distribution of seismic amplitude Low homogeneity Texture attribute Gradient of texture (GoT) Variation of seismic texture High GoT Saliency Attraction to interpreter eyes High saliency 6

7 14 attributes 7

8 Step 2: Training sample picking Training samples: ~20,000 from 3 vertical sections ~ 10 minutes picking 8

9 Step 3: Optimal ML model training Support vector machine (SVM) Manual SVM 9

10 Multi-layer perceptron (MLP) Manual MLP 10

11 Confusion matrix of SVM SVM Confusion matrix of MLP MLP 11

12 Step 4: Volumetric processing SVM MLP 12

13 Result SVM MLP Time 1132 ms Inline 1791 Crossline

14 Interpretation Fault volume imaging Seeded fault picking Automatic fault extraction 14

15 Multi-attribute SVM/MLP analysis Attribute selection is important (e.g., Barnes and Laughlin, 2002; Zhao et al., 2015) a) ideal: perfect classification b) In most cases: overlapping between two features in the attribute domain Variance GLCM Contrast 15

16 Contribution of 14 attributes - Attribute weight matrix 16

17 CNN analysis With original amplitude as input, seismic attribute selection is not necessary The architecture of 1-layer convolutional neural network 17

18 CNN analysis Manual vs. CNN of the training section Four randomly-selected vertical section of the generated fault volume Total precision: 0.88 True positive rate: 0.99 Actual Prediction Non-fault Fault Non-fault Fault

19 CNN analysis 16 attributes 19

20 Conclusions Supervised ML classification is applied to seismic data for fault detection Both SVM and MLP are used for multi-attribute based classification MLP has better performance with enough training samples Attribute selection is important and relies on interpreters experience CNN is capable of generating attributes automatically to complete fault classification, which requires less from an interpreter. More work is in need to Promote sample-level to Image/volume-level detection and segmentation Achieve cross-dataset interpretation Build open and comprehensive training datasets with all features of interest 20

21 in the early morning More information (e.g., recent research, publications, tools, and codes) is available by: Visit our booth: #2109 Visit our center: Visit my webpage: 21