| OpendTect dGB Plugins User Documentation version 4.2 |
|---|
| Prev |
|
Next
|
This window pops up as soon as OpendTect has created or retrieved the train and test sets. In order to create a training set, the software must compute all selected attributes at all picked locations. This may take some time. The user will be notified when the data is collected. Acknowledging the notification will automatically start the training phase.
Training can be stopped and restarted with the Pause (Resume) button. Clear randomizes the weights of the current network. This implies that all training results are lost. The Clear option can be used when over fitting (see below) has occurred and the user wishes to restart training from scratch. The randomized network is then re-trained and the network is stopped before over fitting occurs. This can be done manually (pressing Pause) or by specifying a number in the number of training vectors field.
In Unsupervised training, the network performance is tracked in a graph that shows the average match (confidence) of clustered input. Typically, the average match increases in a step-function. Each step indicates that the network has found a new cluster. Training can be stopped as soon as the average match has reached a stable situation. Usually this will be around 90%.
The colors of the input nodes in an unsupervised network will also change during training. In unsupervised mode, these colors do not indicate that one attribute is more important than another. All attributes in a clustering experiment are equally important.
Optionally, the neural network can be stored immediately on pressing OK. To do this, enter a neural network name in the appropriate field at the bottom of the NN training window.
The Quick UVQ horizon option can be used to quickly segment the seismic waveform along the interpreted horizon. The attributes and data selection are made automatically, contrarily to the standard NN training mode that request the user to define attribute and training locations (picks) before going into the NN training module.
The sole window requests the user to select the input seismic data (2D or 3D according to the type of intepretation). Around 2000 input traces are randomly chosen to train the neural network. The waveform in a given time window is the main parameter to input, together with the output number of classes. It is highly recommended to save the output network.
The time window should match the length of the target of interest, and be extended on either sides by a few samples.
Example: You have a top reservoir horizon and the target is 50m = 32ms thick. An appropriate time window would be [-12, 44] ms along the horizon.
The number of class is hard to determine in advance. Too many classes will give redundant class centres, but too few classes will not match the seismic.
The 'OK' button will start the unsupervised Neural Network training. The training that results above 90% match could then be accepted to create output maps. The 'OK' button in the Neural Network Training window will automatically start processing the two outputs (Segmentation and match grids) on the 3D horizon. The processed grids are only saved to the database using the right mouse click on the processed attributes (Class, Match) of the horizon.
After training an additional Neural Network Info window will pop-up.(see above). This dialog gives a detailed report about the Neural Network training. The Display button in the Neural Network info dialog is used to pop-up a class center display in a 2D viewer (see below). The classes can be color-coded using the Color Table, and the color table can be displayed in the 2D viewer by clicking on the 'show classification' button.
2D QUICK UVQs
The similar workflow is also available for a mapped 2D seismic horizon. Similar settings are required and the processing steps are also the same (as explained previously). After the Neural Network training is finished, the system will prompt to grid and save the classification and the match grids to a 3D horizon. Select or create a 3D horizon to process the classification and match maps.
It is possible to use the output of an unsupervised segmentation for the stacking of seismic data. This functionality allows stacking of multiple cubes based on class number. The input nodes must be a measure of quality for each of these cubes.
This function is only available if the environment variable OD_DGB_QUALITY_STACKING is set to "Yes". The quality stack button appears in the "NN info" window of the loaded neural network. It will not appear if a single volume is used as input to all attributes of the neural network, like for a UVQ classification.
The segment cube must have been previously processed and saved on disk. It represents the first input. The other inputs are the volumes to be stacked, on for each attribute (so the input of the attributes used in the neural network). Once again only stored volumes can be selected.
The output will be a weighted stacking of the input volumes. The weights are represented by the class centre values (one for each attribute), and vary according to the best fitting class centre as represented in the segment volume. Optionaly low weight (relative to the maximum weight in a given class) may discard the corresponding volumes locally. Use a low value for the smoothest output, and a high value for the best discrimination.
In Supervised mode, the network's performance is tracked during training in two graphs: Normalized RMS and % Misclassification:
The Normalized RMS error curves (see network training picture below) indicate the overall error on the train and test sets, in red and blue respectively on a scale from 0 (no error) to 1 (maximum error). Both curves should go down during training. When the test curve goes up again the network is over fitting. Training should be stopped when (preferably before) this happens. Typically a RMS value in the 0.8 range is considered reasonable, between 0.8 and 0.6 is good, between 0.6 and 0.4 is excellent and below 0.4 is perfect. The normalized error is calculated as follows:
The percentage misclassification shown in the lower left corner is a much easier quality control parameter to interpret. It simply shows how the percentage of the training and test set that is classified in the wrong class.
On the right-hand side of the window a graphical representation of the input attributes is shown. The circle in front of the attribute name changes color during training. The colors reflect the weights attached to each input node and are therefore indicative for the relevant importance of each attribute for the classification task at hand. Colors range from red (high weight means high importance) via yellow to red (relative small weights, less important). This feature is very useful when you wish to design small networks to increase processing speed.
Optionally, the neural network can be stored immediately by pressing the OK button. First, enter a neural network name in the appropriate field at the bottom of the NN training window.
The Save misclassified toggle allows saving the misclassified picks in a new Pickset. This Pickset is automatically loaded in OpendTect again. The Pickset can be indicative of picking errors. It is not recommended to bluntly remove the misclassified picks from a Pickset, since good picks, although misclassified during training, still help neural network training.
The Supervised training window from well data is very similar to the training window from a Pickset (see below). The only difference is the display of a scatter plot instead of a % Misclassification plot.
A scatter plot shows the actual target data on the horizontal axis and the predicted target data by the neural network, as it is at that moment, on the vertical axis. Not all nodes are plotted. Only a random selection of the used train and test data is shown. Ideally, after sufficient training, all data points should be on the diagonal. That would mean that the trained neural network predicted all examples correctly. However, this will rarely be the case. In most cases, the data will cluster along the diagonal. The narrower this cloud, the better the neural network is trained.
Overtraining occurs when the Normalized RMS of the test data increases, while the Normalized RMS of the train set decreases. This usually also means that the cloud of train nodes becomes narrower, while the cloud of test nodes becomes wider again.
