Aiming at the problems of traditional shallow machine learning methods applied to high resolution image classification, we propose a high resolution image classification method combining with minimum noise fraction (MNF) rotation and convolution neural networks (CNN). MNF is used to analyze the initial unsupervised pre-training CNN. Linear correction function is adopted as the activation function of the neural network to increase the training speed. In order to reduce the missing of image features in the process of the pool, the sampled image features are put into Softmax classifier under the principle of maximizing sampling probability. Experimental image of typical regions is selected and classified by using the proposed classification method, and the classification results are compared with those of support vector machines classification method and artificial neural network classification method. The results show that the classification accuracy of the proposed method is superior to the shallow machine learning classification methods, and can fully excavate the spatial information of high resolution remote sensing images.