Hyperspectral image classification based on multi-scale hybrid convolutional network

To solve the problems of uneven distribution of hyperspectral image data, insufficient spatial -spectral feature extraction, and network degradation caused by the increase of network layers, a hyperspectral image classification algorithm based on multi-scale hybrid convolutional network is proposed. Firstly, principal component analysis is applied to reduce the dimension of hyperspectral data. Then, the neighborhood extraction is applied to take all pixels in the neighborhood as a sample to supplement the corresponding spatial information. Next, an improved multi-scale hybrid convolutional network is applied to extract features from the preprocessed sample data, and the mixed domain attention mechanism is added to enhance the useful information in the spatial and spectral dimensions. Finally, the Softmax classifier is used to classify each pixel sample. The proposed model is tested on hyperspectral datasets of Indian Pines and Pavia University. Experiments show that the overall classification accuracy, average classification accuracy and Kappa coefficient can reach 0. 987 9, 0. 983 3, 0. 986 2 and 0. 999 0, 0. 996 9, 0. 998 6, respectively. Compared with other classification methods, this algorithm can extract the feature information of hyperspectral images more fully, and achieves better classification results.