Objectives The deep learning model such as convolutional neural network(CNN) has achieved satisfactory results in hyperspectral images classification. However, because traditional neurons can only perform scalar computation, the existing deep learning models cannot model the instantiation parameters of hyperspectral image features, so they cannot achieve satisfactory classification results under the condition of restricted neighborhood scope. Aiming at the problem, we design a new network model by introducing capsule network structure, to further improve the classification accuracy.
Methods The model is composed of traditional convolution layers, capsule layers and fully connected layers, which has a stronger feature representation ability. This model can further improve the classification accuracy of hyperspectral images using the vector calculation of capsule neurons and the spatial relationship between features encoded by the weight matrix. Specifically, the hyperspectral image patches are firstly processed by a convolution layer to extract local features. Next, the primary capsule layer and the digital capsule layer are used to extract the deeper abstract features at higher levels and classify the input data. In addition, the fully connected layers are used for reconstruction to further enhance the abstract modeling ability and generalization ability of the capsule network.
Results Three public hyperspectral images data sets including Pavia University, Indian Pines and Salinas are selected for experiments. The results show that the proposed method outperform the support vector machine(SVM)-based and the traditional CNN-based classification methods.Specifically, compared with the SVM-based methods, the proposed method improves the overall classification accuracy by about 4.7%—7.2%, 8.2%—10.9% and 2.5%—6.9% on three different datasets. Compared with the traditional CNN-based methods, the proposed method improves the overall classification accuracy by about 0.8—3.7%, 2.7—5.5% and 1.3—2.5% on three different datasets.
Conclusions In conclusion, the proposed network model has better classification performance than that of traditional algorithms. Under the condition of sufficient training samples, the overall classification accuracy of the proposed model is higher than the traditional SVM-based and CNN-based classification models, and it has a lower time cost. In additional, the proposed model has better adaptability under the condition of further reducing training samples and pixel neighborhoods.
DownLoad: