The invention provides a hyperspectral image classification method based on combined multi-level spatial spectrum information CNN, and mainly solves the problems of poor classification performance andpoor classification area consistency of hyper-spectral images. The method comprises the following implementation steps: inputting a hyperspectral data set; constructing a convolutional neural networkand a multi-stage spatial spectrum information extraction network; generating a combined multi-level spatial spectrum information convolutional neural network CNN; inputting a training sample set, and training the network by using a loss function; and inputting the test data set, and classifying the hyperspectral image by using the trained combined multi-level spatial spectrum information convolutional neural network CNN. According to the invention, the built combined multi-stage spatial spectrum information convolutional neural network CNN is used; according to the method, the multi-level spatial information and the global inter-spectral information of the hyperspectral image can be extracted and fused, the problems that in the prior art, spatial feature information is not fully utilized, a convolution kernel cannot extract the spectral global information, and consequently the consistency of classification areas is poor and the precision is not high are solved, and the classificationaccuracy of the hyperspectral image is improved.