结合U-net与FPN的地震初至波拾取算法

Seismic First Break Picking Algorithm Combining U-net and FPN

精确识别初至波以及拾取地震初至波初至时间对于地震的精确定位、地震孕震机制的解释、地震预警等都具有很大的意义。单纯地依靠地震学专家进行分析会耗费大量的人力和时间。本文将U-net模型编码、解码语义分割的思想与特征图金字塔网络(feature pyramid networks,FPN)的思想结合,提出了一种新的初至波初至时间拾取算法,满足高精度识别与拾取初至波的需求。算法模型分为编码、解码两部分,编码部分分为四层,用于捕获特征,通过卷积核大小、步长等参数实现特征由低层级到高层级的抽象和提取。解码部分分为四层,通过对前一层进行反卷积上采样实现。最后,将编码、解码部分对应层级特征图组合,通过两层全连接神经网络实现最终结果输出。与传统算法相比,U-net FPN模型在对初至波初至时间的分类与定位上都有着显著提升。

Accurately identifying the first break and picking up the first break time are of great significance for the accurate positioning of earthquakes,the interpretation of earthquake preparation mechanism,earthquake early warning and so on.Relying solely on seismological experts for analysis will con⁃sume a lot of manpower and time.This paper combines the idea of encoding and decoding semantic segmentation of u-net model with the idea of feature pyramid extraction of FPN(fea⁃ture pyramid networks)model,and proposes a new first break time pickup algorithm to meet the needs of high-preci⁃sion recognition and picking up first breaks.The algorithm model is divided into two parts:encoding and decoding.The encoding part is divided into four layers to capture features.Features are abstracted and extracted from low level to high level through convolution kernel size,step size and other pa⁃rameters.The decoding part is divided into four layers,which is realized by deconvolution and up sampling of the previous layer.Finally,the corresponding hierarchical feature maps of coding and decoding parts are combined,and the final result output is realized through two-layer fully connected neural net⁃work.Compared with traditional algorithms,U-net FPN model has a significant improvement in the classification and location of first break time.