Why using CNN for seismic interpretation? An investigation
This study first applies two most popular neural network frameworks, the multi-layer perceptron (MLP) network and the convolutional neural network (CNN), to the problem of seismic salt-body delineation and compare their performance. Then, we investigate two factors that contribute to the better performance of the CNN framework in understanding seismic signals and identifying the important seismic structures by feeding the generated CNN features into training a MLP network . Specifically, on one hand, the CNN is capable of automatically generating a suite of features from the original seismic images, which reduces the dependency on interpreters for computing and tuning seismic attributes. On the other hand and more importantly, the CNN classification is patch based, in which local seismic reflection patterns are taken into account for defining and learning the features of the target structures. In this way, the random/coherent seismic noise and processing artifacts of distinct patterns can be effectively identified and excluded.
CNN features vs. Seismic attributes
The 9 seismic attributes selected for the MLP-based classification, including five GLCM-based attributes (a) contrast, (b) dissimilarity, (c) energy, (d) entropy, (e) homogeneity, and (f) standard deviation, and three edge-detection attributes (g) variance, (h) semblance, and (i) similarity. All attributes have been normalized before classification.
The 8 CNN attributes automatically generated from the first convolutional layer of the CNN framework. The corresponding convolution mask is displayed at the bottom-left corner of each attribute map. Note the apparent difference between them and traditional seismic attributes.
The 16 CNN attributes automatically generated from the second convolutional layer of the CNN framework. The corresponding convolution mask is displayed at the bottom-left corner of each attribute map. Note the apparent difference between them and traditional seismic attributes.
Result analysis
To verify the contributions of the two factors, we feed the features automatically extracted from the CNN network into the same MLP network, at sample- and pattern-level, respectively. The corresponding classification results are illustrated below. We notice that: (a) the noise robustness is significantly improved by the pattern-level classification, since the target salt boundaries commonly have strong directionality whereas the seismic noise is often randomly distributed in the cube and has distinct patterns; (b) the use of the second-layer CNN attributes leads to better performance on defining the features of the target salt-bodies, particularly in the zones of weak reflections as labeled by the white rectangle.
Conclusions
The CNN has proven its superiority in learning and identifying important geological structures such as faults and salt domes from the original post-stack amplitude in a more efficient manner, compared to traditional multi-attribute-based classification schemes. This study has performed an investigation of the two factors that contribute to such superiority. First, the CNN automatically generates a suite of features and optimizes them during the training process. Second and more importantly, the CNN classification is patch-based that incorporates local seismic reflection patterns into building the mapping relationship between the seismic signals and the target structures.
Acknowledgments
The deconvolutional neural network algorithm is implemented based on the open-source Python package TensorFlow developed by Google Brain.
Reference
Di, H., and G. AlRegib, 2019, A comparison of seismic saltbody interpretation via neural networks at sample and pattern levels: Geophysical Prospecting. [Link]
Di, H., Z. Wang, and G, AlRegib, 2018, Why using CNN for seismic interpretation? An investigation: 88th Annual SEG Meeting Extended Abstracts, 2216-2220.