Hyperspectral remote sensing image classification is usually based on the spectral features of objects, but there are plenty of spatial informations in the images. The effective use of spatial information can significantly improve the image classification effect. Because of the special structure of convolution neural network (CNN), CNN has been successfully applied in the field of image classification, and has a good effect on the classification of two-dimensional images. How to improve classification performance through deep learning combined with spatial-spectral information is a key point. Combining the spatial features and spectral information of hyperspectral images, we have developed a three-dimensional convolution neural network model (3D-CNN) for hyperspectral pixel classification, and the multi labels conditional random field is optimized on the basis of the initial classification. Three general open hyperspectral datasets (Indian Pines dataset, Pavia University dataset, Pavia Center dataset) are selected for testing. Experiments show that the accuracy is greatly improved after the classification optimization, the overall accuracy can reach 98%, and the Kappa coefficient reaches 97.2%.