The Automatic First-break Picking Method Based On U-Net

First-break picking is a very basic and important problem in seismic exploration,because the data acquired near the surface always have strong static correction characteristics.And there will be problems such as low energy and low SNR,which will lead to the difficulty of first-break picking,especially the traditional automatic first-break picking method.In the past,first-break picking basically rely on manpower,so the experience of the people is very important and it requires a lot of manpower and large workload.Then people produced many traditional methods,such as STA/LTA,fractal dimension method,correlation method and so on,but these methods are also need people take part in,and in picking work,people often get into trouble like high noise,lack of the trace,and so on and the traditional methods can't handle these problems well.Now deep learning has a good performance in image processing,so a lot of work appeared to try to apply deep learning methods to the problem of first-break picking,such as using convolutional neural network and circulatory neural network for model training.In image processing,first-break picking can be regarded as the problem of image classification,the records before the first break are one class and the records after the first break are the another class.The records of the first break take shape a line,so it also can be regarded as a problem of image segmentation.Main work of this article is to deal with the first-break picking as image segmentation,and using the convolution neural network called U-Net which is performed well in medical images as the training framework.The training data of the seismic data with labels are synthetic,in addition,the loss function is modified,the weight of the first break part is increased in combination with the image gradient,so that the training of the model is more focused on the first arrival part.From the experimental results,deep learning can provide an intelligent and reliable first-break picking method.