Abstract: Aiming at the problems of feature information loss and poor generalization ability caused by convolutional neural network (CNN) classification of remote sensing image, an improved capsule neural network classification model based on channel attention and mixed attention is proposed. Firstly, in order for the capsule neural network to adapt to large-size input images, the two maximum pooling layers are used in the feature extraction module. Secondly, in order to improve the classification accuracy, the SENet attention and CBAM attention are added to the last layer of the feature extraction module for improving the feature extraction module. Finally, the sample set is randomly divided into training set, verification set and test set, and the training set and veritication set are further used to train the model, the test set to test the model, and the AID data set is used to verity the generalization ability of model classification. The experimental results show that the accuracy and Kappa coefficient of the improved capsule neural network based on the SENet network are higher than other models, and the generalization ability is also. The overall classification accuracy and generalization ability of the proposed model are significant improved, thus verifying the feasibility and usability of the method.

Key words: remote sensing image, capsule neural network, classification accuracy, generalization ability, attention mechanism