| OpendTect Workflows Documentation version 4.2 |
|---|
| Prev |
|
Next
|
Purpose: Visualize patterns pertaining to a seismic window that is cut out along a horizon (horizon slice).
Theory: In UVQ waveform segmentation, an unsupervised vector quantizer type of neural network clusters seismic trace segments (= waveforms) into a user-specified number of clusters (typically between 4 and 10 clusters). Two output grids are generated: the segment grid reveals patterns pertaining to the studied interval and the match grid shows the confidence in the clustering result ranging from 0 (no confidence) to 1 (waveform is identical to the winning cluster center). A third useful output is a display of the actual cluster centers. As of v3.2 UVQ waveform segmentation can be done in two ways: a fast track approach called "Quick UVQ" and in the Conventional way. The conventional way is a two-step approach: Step 1--the user-defined number of cluster centers are found by training the network on a representative sub-set of the data (typically 1000 trace segments extracted at random locations along the horizon). Step 2--the trained network is applied to the horizon. Each trace segment is compared to the cluster centers to generate two outputs: segment (index of the winning cluster) and match (how close is the waveform to the winning cluster center).
Software: OpendTect + Neural Networks
Workflow - Quick UVQ:
- Load a horizon and select "Quick UVQ" from the horizon menu (right-click).
- Specify the seismic data set, the number of classes and the time-gate and press OK.
- The software selects waveforms at random positions and starts training the UVQ network. The Average match (%) should reach approx. 90. If it does not reach 90 reduce the number of clusters and/or the time-window. Press OK to continue.
- The trained UVQ network is automatically applied to the horizon. Class (=segment) and match grids are computed and added as "attributes" to the horizon.
- A new window with neural network info pops up. Press Display to display the class centers. Press Dismiss to close the window.
- Class and Match are not automatically saved! To store these grids as "Surface data" with the horizon use the Save option in the horizon menu (right click).
Workflow - Conventional:
- Create a New Picksets and fill this with 1000 random locations along the mapped horizon.
- Open the default attribute set called "Unsupervised Segmentation 2D".
- Open the Neural Networks window and Select Pattern Recognition. In the new window specify Unsupervised, the input attributes, the Pick set generated in above and the Number of classes. Note that the window (the input attributes) should be chosen such that it captures the seismic response of the geologic interval to study. To visualize patterns pertaining to a reservoir interval of approx. 30ms 2WT thickness on a mapped top reservoir horizon select a window length of approx. -12 to 42ms. This captures the interval plus the bulk of convolutional side effects on zero-phase data.
- Train the UVQ network. The Average match (%) should reach approx. 90. If it does not reach 90 reduce the number of clusters and/or the time-window.
- Store the Neural Network and Display the cluster centers by pressing Info ... followed by Display ...
- Apply the Neural Network Segment (or Match) to the horizon in batch: Processing - Create output using Horizon, or on-the-fly: right-click on the horizon in the tree.
Tips:
- If you apply the network on-the-fly you probably want to save the result as Surface data with the horizon for later retrieval.