Seismic fault detection from post-stack amplitude by convolutional neural networks

This study aims at implementing popular CNN tool for the purpose of seismic fault detection, which is superior in two ways compared to the traditional sample-based multi-attribute classification schemes. First, the CNN network defines, learns, and classifies the faults based on local seismic reflection patterns, so that the seismic noises and processing artifacts of distinct patterns can be effectively identified and excluded. Second and more importantly, the CNN network builds the mapping relationship between the seismic signals and the faults using the original seismic amplitude, instead of manually selected seismic attributes, so that the entire process requires less from an interpreters and is applicable to vast datasets without repeated efforts in attribute selection. As the testing dataset, we use a subset (484 inlines x 501 crosslines x 151 samples per trace) of the 3D GSB data that is featured with polygonal faults

Workflow description

[if !supportLists]A. Training image preparation. For the supervised CNN classification in this study, we prepare the training images in three steps. First, the faults in three vertical sections, including crossline #2700, #2800, and #2900, are manually interpreted. Then, the 3 labelled crossline sections are discretized to provide us with a total of 18,248 seeds on the target faults. Next, one image patch of the original post-stack amplitude is retrieved in a size of 31 inlines by 31 samples centered about each of the labelled seeds. Finally, these images are resampled to be 32 by 32 to facilitate the convolution and pooling operations used in the CNN convolutional layers.

B. CNN classifier training. The simplest 1-layer CNN network is used for fault classification in this study. In particularly, it consist of one convolutional layer followed by one fully-connected layers. The input seismic images are 32 by 32. The convolution masks have a size of 9 by 9. The convolutional layer generates 16 features. The 2x2 maximum pooling is used to reduce the dimensions of output features after convolution and hence to control overfitting. The fully-connected layer has 1024 neurons, and the softmax cross entropy is computed for measuring the probability error between the classification and the true labels. The prepared 318,240 images are used for training the CNN network in 300 epochs.

C. Volumetric processing. The trained CNN classifier is applied to the entire seismic survey for generating a fault volume. At each seismic sample, an image patch of 31 crosslines by 31 samples is first retrieved from the original amplitude and then resampled into 32 by 32. Then its label is predicted by the trained CNN classifier and assigned to the central sample.

Result

For demonstrating the accuracy of the proposed CNN approach, we first compare the results with the traditional multi-attribute based classification methods. Figure 3 displays the comparison of the crossline section #2800, among which we notice that the CNN result (d) is clean and closest to the manual interpretation (a). Next, for addressing the concern of overfitting, Figure 4 displays the clipping of the generated fault volume to four randomly-selected vertical sections that were not used in the training process. It is clear that, the CNN classification successfully detects the faults as thin lineaments, indicating that the trained CNN classifier is capable of learning the target seismic features from the original post-stack seismic amplitude and detecting the identical ones accurately.

Figure 3. A comparison of labelling the faults using the traditional multi-attribute based SVM (b) and MLP (c) and the proposed CNN (d) in the crossline sections #2800.The manual fault interpretation is shown in (a). Note that the CNN detection is the closest to the manual interpretation.

Figure 4 The clipping of the fault volume (in black) by the 1-layer CNN network to four randomly-selected vertical sections, including inline #1791, inline #2011,crossline #2600, and crossline #3000 from left to right, overlaying the original seismic amplitude (in blue-white-red). Note the good match between the CNN detection and original images.

Conclusions

Compared to the traditional multi-attribute based techniques, the CNN network is capable of optimally connecting the seismic images with the target faults using the original post-stack amplitude, instead of manually selected seismic attributes, so that the entire process saves interpreters lots of efforts in selecting and generating seismic attributes. Moreover, the CNN classifier utilizes the local seismic reflection patterns, instead of stand-alone attribute values, so that the seismic noises and processing artifacts of distinct patterns can be effectively identified and excluded.

Reference:

Di, H., Z. Wang, and G. AlRegib, 2018, Seismic fault detection from post-stack amplitude by convolutional neural networks: 80th EAGE Conference and Exhibition.