Seismic first arrivals automatic pickup technology is of great significance in seismic exploration. The traditional manual pickup method has problems such as low efficiency and weak noise immunity, so the industry has been looking for more efficient and automated methods. In recent years, deep learning techniques, especially convolutional neural networks, have attracted much attention in seismic wave first arrivals pickup, and by training a large amount of data, the deep learning methods can automatically identify the features and classify them efficiently, which improves the limitations of the traditional methods. In this study, a deep fully convolutional Segnet semantic segmentation network is used to directly output the probabilistic map of the first arrivals location through semantic segmentation of effective waves and noise based on the advantages of fully convolutional neural networks. The method has the advantages of high efficiency and strong noise resistance, and it has achieved good results in the orthogonal data and real dataset tests.