Attention-Based 3D Seismic Fault Segmentation Training by a Few 2D Slice Labels

TL;DR

λ-binary cross-entropy (BCE) and λ-smooth L₁ loss are presented to effectively train 3D-CNN by some slices from 3-D seismic volume label, so that the model can learn the segmentation of 3- D seismic data from a few 2-D slices.

Abstract

Detection faults in seismic data is a crucial step for seismic structural interpretation, reservoir characterization and well placement. Some recent works regard it as an image segmentation task. The task of image segmentation requires huge labels, especially 3D seismic data, which has a complex structure and lots of noise. Therefore, its annotation requires expert experience and a huge workload. In this study, we present lambda-BCE and lambda-smooth L1loss to effectively train 3D-CNN by some slices from 3D seismic data, so that the model can learn the segmentation of 3D seismic data from a few 2D slices. In order to fully extract information from limited data and suppress seismic noise, we propose an attention module that can be used for active supervision training and embedded in the network. The attention heatmap label is generated by the original label, and letting it supervise the attention module using the lambda-smooth L1loss. The experiment demonstrates the effectiveness of our loss function, the method can extract 3D seismic features from a few 2D slice labels. And it also shows the advanced performance of the attention module, which can significantly suppress the noise in the seismic data while increasing the model's sensitivity to the foreground. Finally, on the public test set, we only use the 2D slice labels training that accounts for 3.3% of the 3D volume label, and achieve similar performance to the 3D volume label training.

Figures (18)

• Figure 1: (a) shows the trend of $\theta$. (b) shows the attention map label generated when we replace each pixel in ground truth with the response point in (a).
• Figure 2: The working principle of AAM when forward propagating is that the attention map will generate a weight approaching 1 around the fault, while the non-fault area will generate a weight approaching 0. In the decoding part of U-Net, low-level features will be fused, but it contains a strong noise. Before fusion, multiplying the attention map with low-level features can improve the sensitivity to fault areas while suppressing factors such as noise in non-fault areas.
• Figure 3: The AAM is embedded in the basic U-Net framework, which can suppress the noise in the background region of the low-level features in the forward-propagation feature fusion phase and improve the performance by optimizing the sensitivity of the high-level and low-level features to the foreground region with additional gradients in the back-propagation. Moreover, the model is trained by our proposed $\lambda$ loss function, which is labeled with a few 2D slices, it can fully extract 3D spatial information from the limited number of slice labels, and we will describe this process in detail in the next section.
• Figure 4: The elements on the last feature map share a set of weights, which allows us to obtain effective gradients only by training using the labeled voxels in the label.
• Figure 5: 2D visualization of cuboid weights.