A new method for three-dimensional stochastic reconstruction of spatial data is proposed. This method introduces deep learning into the feature extraction and reconstruction process of discrete spatial data. In the training process, the spatial data features are learned by constructing a deep neural network, and the global correlation between data is obtained; then the reconstruction results are obtained by feature replication. In the training process, this method doesn’t need to scan the training image repeatedly, which is different from the traditional multiple-point simulation. The experimental results show that the structural features of reconstructed spatial data using this method are consistent with the training images.