Hyperspectral image classification based on residual dense and dilated convolution

Abstract

The approach based on Convolutional Neural Network model has been widely employed in the field of hyperspectral image classification, demonstrating promising classification performance. However, traditional CNN methods only extract deep features from the end of the network, without considering the combination of shallow and deep features from the network. Moreover, most methods extract features using fixed convolutional kernel, which cannot be changed according to the complex spatial structure of hyperspectral image, and ignore the features of data in spatial-spectral distribution. To solve the aforementioned issues, we propose a method for classifying hyperspectral image based on residual dense and dilated convolution (RDDC-3DCNN). First, our method uses a 3D convolutional neural network as the base structure, and the raw hyperspectral data cube is used as the input of the network. Then, a residual dilation dense block incorporating dilated convolution is devised, which enabled the extraction of spatial-spectral features from hyperspectral image using 3D dilated convolution, while simultaneously fusing spatial-spectral features from different levels. After that, shortcut connections is added into the designed module to fully combine the shallow features and deep features of the network. Finally, the fused features obtained are sequentially passed through fully connected layer, dropout layer, and softmax layer to accomplish classification. The experimental results show that the method achieved classification accuracies of 98.87%, 99.33% and 99.32% on the the Indian Pines, the University of Pavia and Salinas datasets, respectively. Experimental results indicate that the proposed RDDC-3DCNN method can make full extract of the spatial and spectral information of the image, and it has better performance than some advanced algorithms in HSI classification.