基于U-net网络的频散曲线自动拾取方法研究

Research on Automatic Picking Method of Dispersion Curve Based on U-net Network

频散曲线拾取是面波勘探的关键环节,旨在通过频散曲线反演出地下横波速度结构。然而目前频散曲线拾取工作主要依靠人工拾取,耗时耗力。为此,本文通过将频散曲线拾取问题看成是图像分割问题,引入U-net网络,发展出一种频散曲线的自动拾取方法。该方法使用频散能量图并使其作为数据集,使用人工手动拾取的频散曲线作为标签集;通过卷积神经网络经由上采样、下采样和跳层链接等步骤学习图片特征,实现频散曲线的自动拾取。模型测试结果验证了利用U-net网络提取频散曲线的准确性。最后本文将训练好的网络模型应用于冰岛南部Ölfusá河岸的Arnarbæli周边试验场地的实际数据频散曲线提取,并将提取结果与手动拾取的频散曲线进行对比。结果表明,利用U-net网络提取频散曲线预测速度快,预测512×512×3大小的图片耗时为96 ms,预测准确度高。

The picking up of dispersion curves is a crucial step in surface wave exploration,aiming to invert the velocity structure of underground shear waves through dispersion curves.However,currently the work of picking up dispersion curves mainly relies on manual picking,which is time-consuming and labor-intensive.Therefore,this article regards the problem of picking up dispersion curves as an image segmentation problem,introduces U-net networks,and develops an automatic method for picking up dispersion curves.This method uses a dispersion energy map as the dataset and manually picked dispersion curves as the label set.By using convolutional neural networks to learn image features through upsampling,downsampling,and skip layer linking,automatic picking of dispersion curves is achieved.The model test results verified the accuracy of using U-net network to extract dispersion curves.Finally,this article applies the trained network model to the actual data dispersion curve extraction of the Arnabæli surrounding experimental site on the banks of the Ölfusá River in southern Iceland,and compares the extraction results with manually picked dispersion curves.The results show that using U-net network to extract dispersion curves has a fast prediction speed,a prediction time of 96ms for images of size 512×512×3,and high prediction accuracy.