Machine/deep learning-based seismic interpretation: An example of salt-body delineation
With the growing complexity of seismic data in both size and resolution, more efficient, accurate, and effective seismic interpretation increasingly relies on the development of powerful computational interpretation tools that are capable of mimicking an experienced interpreter’s intelligence. In recent years, the machine/deep learning has been a popular topic in various disciplines and is attracting more and more attentions from the petroleum industry. Geoscientists desire to explore the massive seismic data in more intelligent ways to extract more useful information about the subsurface reservoirs. This blog will share my findings while playing with the most popular machine/deep learning tools for salt-body detection from the synthetic SEG-SEAM dataset, which is featured with a complex salt intrusive in a folded Tertiary basin and challenges the existing interpretation tools (Figure 1). Both the traditional support vector machine and the emerging convolutional neural network are implemented under the same circumstance to provide a fair comparison of their performances on analyzing seismic data and solving interpretation problems.
Multi-attribute based SVM analysis
The support vector machine (SVM) is one of the most popular algorithms used in the traditional workflow of multi-attribute based seismic feature classification (Figure 2). Particularly,
9 attributes are selected and generated from the SEAM data, including GLCM contrast, dissimilarity, energy, entropy, homogeneity, standard deviation (std), variance, Semblance, and Similarity (Figure 3), all of which well highlights the salt-body boundaries;
210,168 training samples are prepared by manual interpretation on three vertical sections, with half on the boundaries (Figure 4). Note that the boundary in the deep area is recognized and interpreted, even though the signal is weak.
Results:
Classification accuracy: 80.9%, when applying the built SVM model to the 210,168 samples in three vertical sections with manual interpretations (Table 1). Two major observations from the corresponding results (Figure 5):
The boundaries with strong signal are well detected, but thicker than manually interpreted. Also the strong reflection events are detected as boundaries. It may be corresponding to the 38,789 (19.4%) false positives of non-salt samples as salt ones;
No boundary is detected in the deep areas, where the signal is too weak and no attribute supports such feature (denoted by circles). It may be corresponding to the 1,324 (13.2%) false negatives of salt samples as non-salt ones.
Multi-attribute based MLP analysis
The multi-layer perceptron (MLP) also has wide applications in the field of seismic feature delineation, such as gas-chimney detection, since the 1990s. In general, such approach follows the same workflow as shown in Figure 2, and for a fair comparison to the SVM analysis, the same 9 attributes (Figure 3) and 210,168 training samples (Figure 4) are used for building a fully-connected neural network, which contains 3 layers with 16 neurons in each.
Results:
Classification accuracy: 94.3%, when applying the built fully-connected network to the 210,168 samples in the three vertical sections with manual interpretations (Table 2). In particular, compared to the SVM analysis (Figure 5), such multi-attribute based ANN analysis (Figure 6),
successfully decreases the false positives of non-salt samples as salt ones to 9,036 (4.5%), which makes the detection cleaner, thinner, and closer to the manual interpretation (Figure 4).
undesirably increases the false negatives of salt samples as non-salt ones to 2,877 (28.3%), which makes the detection even less complete compared to the SVM detection (Figure 5), particularly the inner boundaries where multiple salt bodies merge (denoted by circles).
Figure 7 displays the normalized weights of the 9 attributes for salt-boundary classification. The GLCM contrast (~22%), energy (~19%), and homogeneity (~20%) are given the largest weights in the trained ANN model, whereas the contributions from the GLCM standard deviation, variance, and similarity are subtle, about 4%, 3%, and 0.3%, respectively. Such difference is consistent with our observations in Figure 3, where the salt boundaries are clearly highlighted by the GLCM attributes, but weakly by the similarity.
Figure 8 plots the accuracy of the learning process of 500 epochs, indicating that the model becomes optimal after 50 epochs with the maximum accuracy being ~84%. To avoid overfitting, cross-validation is used in each epoch: first, the 210,168 training samples are randomly split into the training set and the validation set by a ratio of 0.9; then, the training set is deliberately balanced, so that it contains the same amount of non-salt labels and salt labels; next, the model is adjusted by feeding the training set; finally, the trained model is applied to both the training set and the validation set, providing us with the training accuracy (in blue) and the validation accuracy (in red), respectively.
Amplitude-based CNN analysis
Deep convolutional neural network (CNN), as a class of deep, feed-forward artificial neural network, has been successfully applied to image analysis. By treating seismic data as images, such approach is then implemented for salt-body detection, with original post-stack amplitude as input, instead of computing a series of seismic attributes. 192,480 patches generated from the three vertical sections of manual interpretation are used as the training data.
Results
Classification accuracy: 99.98%, when applying the built CNN network to the 192,480 samples in the three vertical sections with manual interpretations (Table 3). In particular, compared to the multi-attribute based analysis, such CNN analysis (Figure 9)
successfully differentiates the non-salt features and the salt features, with 37 (0.1%) false positive and 0 (0.0%) false negative.
leads to saltbody boundaries closely similar to manual interpretation.
By treating the convolutional layers in a CNN classifier as an attribute generators, the CNN-based classification then follows the same multi-attribute based workflow (Figure 2). Here for investigating how a CNN distinguishes the salt boundaries from the other features, we output the eight "attributes" (in the yellow borders) from a given vertical section (in the blue border), with the corresponding 9x9 masks in the top-right corner. Apparently, the CNN "attributes" are by simple math operators, thereby of no geophysical implications and distinct from those shown in Figure 3. Such attributes may be meaningless to our human eyes, but a computer is capable of combining them in an efficient way by assigning them different weights.
Below is a 3D view of the detected salt-body with clipping to 6 vertical sections, none of which is used in the process of training the CNN classifier. Good match is observed between the detection (in black) and the original seismic images.
In summary, the CNN network has great potential for analyzing seismic data and assisting seismic interpretation. More work is in need in the future for investigating the impacts of its key parameters, addressing the common issues, and more importantly quantifying its performance in seismic object detection, facies analysis, and more applications in this domain.
References:
Di, H., Z. Wang, and G. AlRegib, 2018, Deep convolutional neural networks for seismic salt-body delineation: AAPG Annual Convention and Exhibition.
Last updated: March, 2018