In order to adapt to efficient seismic acquisition after 5G（fifth-generation mobile communication technology）intelligent node being put into production, the scheme of automatically extracting obstacles has been proposed by using a little self-made training data and innovatively combining traditional algorithms with deep learning. A SegNet networkhas built, obtaining a rough segmentation result.Optimization processing including boundary smoothing and noise elimination, which achieved by utilizing probabilistic graphical model - conditional random fields and feature - space probability fusion respectively.Finally, the segmentation has attained higher accuracy rates on each class of obstacles.Compared with the deep network single using which needs tens of thousands of levels of data to improve the generalization ability,a semantic segmentation method has proposed with strong generalization ability under the condition of a little training data, which can below-costly, flexibly and efficiently used in various specific semantic segmentation scenarios.