Multiscale dilated dense network for hyperspectral image classification

Abstract

Hyperspectral image has rich spectral information and can accurately reflect the real condition of ground objects, so it is widely used. However, there are huge challenges in hyperspectral image, such as few training samples, the phenomena of “the foreign matter same spectrum” and “the same thingdifferent spectrum” can not be distinguished. The hyperspectral image classification method based on convolutional neural network can solve the above problems and achieve good classification performance. Most classification methods based on convolutional neural network only extract the deep semantic information at the end of the network, ignoring the shallow details. At the same time, Methods of this type increase the receptive field by designing deep network structure, which is prone to gradient vanishing. To solve the above two problems, a multiscale dilated dense network is proposed for hyperspectral image classification. Firstly, hyperspectral data cubes with different neighborhood sizes are selected as network inputs to construct a multi-channel network structure with different scale inputs. Secondly, a multiscale feature mapping module with fusion of dilated convolution is designed. Different numbers of dilated convolution are used to increase the model's receptive field without increasing network parameters, and spatial-spectral features of different scales are fused. Skip connections are added between the designed modules to fully combine the shallow and deep features of the network, without skip connections in the modules, reduce the computation of the overall network. Finally, the resulting features are input into the fully connected layer and softmax to complete the classification. In addition, in order to prevent the network from overfitting, we add dropout regularization method after the fully connected layer to improve the generalization ability of the model. Experiments are conducted on three publicly available datasets, Indian Pines, University of Pavia and Salinas, with accuracy rates of 99.28%, 99.48% and 99.25%, respectively. The experimental results show that compared with the classification methods based on convolutional neural network, the proposed method can effectively extract multiscale features of hyperspectral image and enhance the expression ability of spatial-spectral fusion features, and the classification performance is better.