【摘要】 基于UNet++对P波初至拾取进行研究，首先对UNet++进行降维，并从网络结构的深度和单个Block的操作两方面对原始网络进行改进；然后给模型选择损失函数和优化器，让模型能够有优化的目标和方向；接着进行数据的预处理，筛选信噪比小于20 dB的数据出来，对其进行小波阈值去噪、归一化处理；最后是训练和验证，选择在验证集上表现最优的一个模型作为最终模型。经过150条测试集数据的测试，证明所使用方法在均值、方差、命中率3项指标上均优于STA/LTA和AR-AIC,其中P波初至拾取的精确率高达98.00%,为P波初至自动拾取提供了一种新思路。更多还原

【Abstract】 This paper studies the P-wave first-arrival picking based on UNet++. First, UNet++ is dimensionally reduced, and the original network is improved from the depth of the network structure and the operation of a single block. Then, the model selects a loss function and an optimizer to let the model find optimization goals and directions; then data preprocessing is performed, and the data with a signal-to-noise ratio less than 20db are screened out, and wavelet threshold denoising and normalization are performed on them; finally, training and validation are performed, and the performance on the validation set is selected. The optimal model is used as the final model. After the test of 150 test set data, it is proved that the proposed method is superior to STA/LTA and AR-AIC in the three indicators of mean, variance, and hit rate. The precision of the P-wave first arrival picked up by the method is as high as 98.00%. This work provides a new idea for automatic pickup of P-wave first arrivals.更多还原