In recent years, convolutional neural network has been widely used in the field of computer vision and achieved good results in image classification. In the field of remote sensing image processing, the demand for hyperspectral image (HSI) classification is also increasing. However, as the training of CNN requires large amounts of labeled hyperspectral images, convolutional neural networks are difficult to apply to this domain. Considering this problem, we proposed a hyperspectral image classification model with multi-scale convolutional neural networks. The input of the model is the original hyperspectral image and the output is the final classification result. The characteristics of this model are as follows. First, the model can extract spatial features of the input data from multi-scales automatically, rather than requiring handcrafted features. Second, multi-scales feature extraction is adopted, and more samples are involved in the classification, solving the problem of obtaining large amounts of labeled hyperspectral data. Third, the model proposed in this paper consists of a multi-scale convolutional spatial feature extraction layer, a feature fusion layer, a normalization layer, a dropout layer and an activation layer, which are more suitable for hyperspectral image classification. Finally, the experimental results for the Indian Pines dataset show that the classification model proposed in this paper is better than other state-of-the-art classification models in terms of the overall accuracy, average accuracy and the Kappa coefficient.