Automatic Velocity Picking Based on Improved Mask R-CNN
Automatic Velocity Picking Based on Improved Mask R-CNN
复制标题
DOI：
10.1109/tgrs.2023.3335250
复制
发表时间：
2023
影响因子：
8.2
通讯作者：
Yijun Yuan;Ying Li;Xiaoxuan Fang;Fengfeng Shi
中科院分区：
文献类型：
--
作者：
Yijun Yuan;Ying Li;Xiaoxuan Fang;Fengfeng Shi
速度谱分析是确定动校正速度的主要方法。然而，传统的速度谱分析方法效率低下，并且通常受主观因素的影响，导致所选择的速度与实际速度的偏差。因此，我们开发了一种使用基于掩码区域的卷积神经网络（Mask R-CNN）的速度拾取方法。该方法首先生成大量的训练数据集，包括速度谱和标签，然后将它们馈送到网络进行训练。在用训练数据训练网络之后，它自动地从速度谱中提取时间-速度函数，并将它们输出以用于后续的数据处理。该方法将地震数据处理中速度-能量簇的拾取转化为图像处理中的目标检测。因此，为了使该方法能够准确地检测速度谱中的速度-能量簇，根据速度-能量簇的特点，对传统Mask R-CNN的锚尺度和锚比进行了改进，构建了适合速度-能量簇检测的Mask R-CNN结构。通过改进非最大值抑制算法，解决了传统Mask R-CNN中对象重复检测的问题。对合成数据和实测数据的实验结果表明，该方法能够快速、准确地从速度谱中提取时间-速度函数。在速度拾取、动校正共中心点道集和叠加剖面等方面，均取得了比传统的Mask R-CNN和人机交互速度拾取（MIVP）方法更好的效果。
Velocity spectrum analysis is the main method used to determine the normal moveout (NMO) velocity. However, conventional velocity spectrum analysis methods are inefficient and are typically affected by subjective factors, resulting in a deviation of the selected velocity from the actual velocity. Consequently, we developed a velocity-picking method that uses a mask region-based convolutional neural network (Mask R-CNN). The proposed method first generates a large number of training datasets, including velocity spectrum and labels, and then feeds them to the network for training. After the network is trained with the training data, it automatically extracts time–velocity functions from the velocity spectrum and outputs them for subsequent data processing. The proposed method converts the picking of velocity-energy clusters in seismic data processing into object detection in image processing. Therefore, to enable the method to accurately detect velocity-energy clusters in the velocity spectrum, according to the characteristics of velocity-energy clusters, the anchor scale and anchor ratio of the traditional Mask R-CNN are improved and a Mask R-CNN structure suitable for velocity-energy cluster detection is constructed. We solved the problem of object duplicate detection in the traditional Mask R-CNN by improving the nonmaximum suppression algorithm. The results of experiments conducted on synthetic and field data indicate that the proposed method quickly and accurately extracts time–velocity functions from the velocity spectrum. Furthermore, it obtains better results for velocity picking, NMO-corrected common midpoint gathers, and stacked sections than the traditional Mask R-CNN and the man–computer interaction velocity-picking (MIVP) method.