Abstract Seismic velocity is a critical parameter in seismic exploration, and its accuracy significantly affects the reliability of data processing and interpretation results. However, manual velocity picking methods are not only inefficient but also time consuming, making them increasingly inadequate for meeting the demands of practical production work. This paper introduces the multi-information combination convolutional neural network (MCCN) velocity auto-picking method. Building on the foundation of convolutional neural networks, we have designed the network structure of the MCCN method specifically tailored to the characteristics of stacked velocity picking tasks. Given that velocity spectrum energy clusters exhibit both morphological and trend features, we use a regression convolutional neural network to enhance the accuracy of velocity picking. Furthermore, as the velocity spectrum contains interference from multiple waves and other noise, we employ a coordinate attention mechanism to mitigate the influence of interfering information. Our approach involves the simultaneous incorporation of velocity spectrum and CMP information through a dual-combination network, thereby further enhancing velocity picking accuracy. Finally, we compare our method with fully connected convolutional neural networks and manual velocity picking methods, demonstrating the practicality and precision of our proposed approach.