Interval velocity is a key parameter for obtaining high signal-to-noise ratio, high fidelity, and high-resolution seismic profiles. In recent years, with the help of the nonlinear mapping ability of deep learning, interval velocity modeling method based on seismic reflection waveform data of different types has development. However, current research is mainly focused on supervised learning, i.e., training a large number of "seismic waveform data-velocity model labeled" data pairs into the input network.This approach faces two drawbacks: first, the high cost and large amount of data for obtaining the actual subsurface velocity structure as a label, and second, the problem that the accuracy of velocity modeling depends directly on the nonlinear mapping ability of the neural network after supervised training.Therefore, in this paper, we propose a semi-supervised self-training velocity modeling method, which trains the initial supervised network with a small number of "seismic waveform data-velocity model label" data pairs, and then generates velocity pseudo-labels using unlabeled seismic waveform data. The "seismic waveform data-velocity model label" data pair and the "unlabeled seismic waveform data-velocity model pseudo-label" data pair are mixed to retrain the network and the velocity model pseudo-labels are iteratively updated using the new network model until the network model is updated by the velocity pseudo-label self-training, and then semi-supervised self-training is realized to improve the accuracy and generalization of the velocity modeling, accuracy and generalization.Meanwhile, in order to compress the spatial and channel redundancy in the convolutional neural network and to improve its performance, spatial and channel reconstruction convolution (SCConv) was used to construct the SCConv-Unet network.Finally, in order to verify that the semi-supervised self-training method is suitable for velocity modeling and can improve the accuracy of velocity modeling, numerical experiments are conducted using a fault velocity model, a horizontal laminar velocity model, and a velocity model containing velocity anomalies.The experimental results show that the accuracy of the semi-supervised self-training velocity modeling method can further improve the velocity modeling accuracy of the supervised learning method; making full use of the potential of the velocity-free labeled seismic waveform data can effectively improve the velocity modeling accuracy and reduce the cost of dataset production. In addition, the SCConv-Unet network shows good generalization ability and nonlinear mapping ability, which helps to accelerate the convergence speed of semi-supervised iterative training.