Microseismic monitoring technique is an important tool for hydraulic fracturing process monitoring and fracturing effect evaluation. For surface monitoring,the P-wave polarity can directly and quickly invert the focal mechanism,while the polarity correction can also improve the imaging accuracy of the diffraction-based location method. Therefore,accurate and rapid determination of P-wave polarity is of great significance for real-time monitoring at surface. Convolutional neural network(CNN)is a deep learning algorithm with powerful feature learning and classification capabilities. It can also be used to determine the P-wave polarity of microseismic events. Since microseismic monitoring at surface mostly uses star,grid,or other regular acquisitions,in this paper,we use the target trace and its neighboring seismograms as input sample to build a multi-trace P-wave polarity classification network model based on convolutional neural network. The results from field data application show that,in comparison to the single-trace based CNN,the multi-trace based CNN model can combine the target trace with the neighboring traces to predict the polarity of the target trace,and improve the accuracy of polarity classification for surface microseismic data from a regular observation system.