**1. Introduction**

The main goal of the high-resolution Velocity Model Building process is to reconstruct the subsurface structures, especially to capture potential geological bodies, such as shallow gas clouds, salt bodies and fault plains. On the other hand, these complex geological structures result in certain types of anomalies. Refs. [1,2] proposed different methods to calculate velocity models. Particularly, gas clouds correspond to lower velocity and a smaller Quality Factor (*Q*) [3], and some quantifiable techniques were proposed to identify shallow gas pockets. A ray-based *Q*-tomography was developed by [4] and has been applied to field data to estimate the effects of shallow gas by representing the pockets as anomalous *Q* bodies [5]. One of the most crucial problems in the *Q*-tomography method is how to predict the *Q* bodies masks accurately, which is used to indicate the location of strong absorptions. Usually, several iterations were needed to improve the precision of the location information, either by manual editing or introducing some attributes as pilot. Refs. [6,7] show exciting *Q*-tomography results, employing the FWI model to produce the masks for the subsequent anomalous *Q*-tomography process to evaluate the shallow gas clouds. However, in the FWI guided *Q*-tomography flow, the accuracy of the masks is strongly dependent on the FWI process, human intervention, and the initial Q-factor model, which are extremely time-consuming and tedious. The initial Q-factor model can be directly obtained from field data by applying sophisticated methods [8]. Furthermore, the initial velocity model usually starts from a smooth one, which is based on manual velocity analysis or vintage experience from that area only; the local detailed information is excluded. Introducing accurate location information into velocity models automatically

**Citation:** Li, Z.; Jia, J.; Lu, Z.; Jiao, J.; Yu, P. Seismic Velocity Anomalies Detection Based on a Modified U-Net Framework. *Appl. Sci.* **2022**, *12*, 7225. https://doi.org/10.3390/ app12147225

Academic Editors: Guofeng Liu, Xiaohong Meng and Zhifu Zhang

Received: 9 June 2022 Accepted: 13 July 2022 Published: 18 July 2022

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2022 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

in the early stages will help to improve the quality of the VMB results and accelerate the model building process significantly.

Since deep learning was proposed by [9], it has received widespread attention and is widely used in computer vision, voice recognition, natural language processing, etc. Recently, machine-learning based techniques were introduced into the seismic data processing and interpretation community. Ref. [10] presented a supervised-learning-based salt body detection algorithm. Three features—amplitude, second derivative and curve length—were selected to characterize voxels of 3D seismic volume, and the algorithm uses small fraction of the characterized voxels for training to predict the whole volume. Ref. [11] developed a texture classification workflow using seismic attributes, clustering techniques and segmentation by thresholds, followed by second step mathematical morphological and basic operations between volumes to improve the detection. A major strategy of this type of method is to apply data mining algorithms [12] on the post-migration volumes. Ref. [13] developed a novel method based on machine learning techniques to automatically identify and localize faults. The method was introduced in the initial stages of the VMB process, when no seismic data had been migrated, which is different from other types of post-migration methods that use processed seismic data or migrated images [14,15].

In the field of computer vision, it is well known that Fully Convolutional Networks (FCNs) and U-Net perform well on image segmentation tasks. These two frameworks were proposed by [16,17], respectively. U-Net is similar to FCN and has been widely used in medical image segmentation. Compared with FCN, the first feature of U-Net is that it is completely symmetrical, which means the left and right hand side are very similar. However, the decoder of FCN is relatively simple, using only a deconvolution operation. The second difference is skip connection: FCN uses summation, while U-Net uses concatenation. The U-Net model modified and expanded the network on the basis of FCN, so that it can use very few training images to obtain very accurate segmentation results. In addition, an upsampling stage is added, which adds lots of feature channels, allowing more texture information of the original image to spread in high-resolution layers. U-net does not have a fully convolutional layer and uses valid for convolution throughout, which ensures that the results of the segmentation are based on no missing context features.

In this paper, a two-step deep-learning based velocity anomalies detection workflow is established. The workflow starts from the pre-migration shot gathers directly, justifiying the presence of anomalies firstly and then predicting accurate location information of the velocity anomalies prior to VMB process by employing a modified U-Net neural network. A set of two-dimensional synthetic model tests are presented to evaluate the effectiveness of the proposed workflow.
