First break picking with deep learning – evaluation of network architectures

Abstract

In recent years, various convolutional neural network architectures have been proposed for first break picking. In this paper, we compare the standard auto-encoder and U-net architectures as well as versions enhanced with ResNet style skip connections. The U-net appears to have become the standard network for segmentation, judging from the number of published articles. Still, there is some variety in neural network architectural choices. In this paper, we assess the impact of neural network depth, width and input data size, as well as some small modifications for deep networks offered by the ResNet. In general, results improve as the networks get deeper, but with diminishing returns. The more complex the data, the more benefit the deeper networks bring. We use complete shot gathers, albeit rescaled for efficiency, to train the neural networks. For shot gathers with a simple piecewise linear moveout, this approach yields results with good accuracy when gathers are resampled to 128 × 128 samples. For shot gathers with more complex first break moveout, using our approach it is advised to stay close to the original dimension of each gather for best accuracy, at the expense of increased training times. A good trade-off between network depth, image size and training times is to use a nine-stage U-net with 256 sample images. Despite the advantages in other applications, the basic U-net outperforms a U-net with ResNet features. We show that changing the input data dimensions for trained networks does not work, despite the fact the fully convolutional networks are independent of image size. The U-net based first break picking is not sensitive to picking errors, as in many cases the neural network predictions are better than the training data where the training data have random mispicks. This suggests a practical application; namely, to train or re-train a pre-trained network on a single data set after conventional first break picking with the objective of improving conventionally picked first breaks.

A Single Data Extraction Algorithm for Oblique Photographic Data Based on the U-Net

“…The binary cross entropy (BCE) loss function is used to train the U-Net network, which can quickly fit the training data [63][64][65]. However, in the unbalanced data, the BCE results often cannot reflect the actual results of less data.…”

Common‐offset domain reflection–diffraction separation with deep learning

“…Although most geophysicists will be familiar with the concept of convolutions, the overview given by Dumoulin and Visin (2018) will benefit the reader that is interested in the practical details. The original U‐net was used for the segmentation of biomedical images and also has segmentation application in geophysics, such as salt segmentation (Shi et al., 2019) and first break picking (Zwartjes & Yoo, 2021). Since the U‐net architecture is basically an enhanced auto‐encoder, it is also used frequently in regression problems, such as interpolation (Fang et al., 2021) and velocity model building (Yang & Ma, 2019).…”

“…Unfortunately, there is no theoretically optimal design choice for neural network architectures for a particular problem, and therefore this design process mostly comes down to trial and error to empirically decide which architecture and network hyperparameters give the best results. The neural network architecture can be tweaked with various elements that can each have a positive impact on the final outcome, for example shown in Zwartjes and Yoo (2021). However, in practice, the differences will be small and incremental.…”

Building initial model for seismic inversion based on semi‐supervised learning

“…Deep learning is a data‐driven algorithm that can learn implicit nonlinear relations from label data and use it to solve nonlinear problems. Deep learning has been widely used in geophysical problems such as seismic facies analysis (Liu et al., 2021; Nishitsuji & Exley, 2019), first‐break picking (Wang et al., 2019; Yuan et al., 2018; Zwartjes & Yoo, 2022), fault identification (Huang et al., 2017; Wu et al., 2019; Zhou et al., 2021) and model building (Araya‐Polo et al., 2017; Fabien‐Ouellet & Sarkar, 2019; Ovcharenko et al., 2022). Recently, the application of deep neural networks in reservoir characterization has also been investigated (Chen & Saygin, 2021; Dhara et al., 2023; Di & Abubakar, 2021; Sun et al., 2021; Wu et al., 2021).…”