Research Of Hyperspectral Image Classification Based On Convolutional Neural Network

Hyperspectral images contain rich spatial information and spectral information,which can meet the requirements of ground objects feature classification,and have been widely applied to many practical scenes.Deep learning is widely used in hyperspectral image classification because of its powerful feature learning ability and classification ability.This paper focuses on the in-depth study of hyperspectral image classification methods based on convolutional neural networks.Firstly,aiming at the problem of limited labeled samples and high dimension of hyperspectral images,a hyperspectral image classification framework based on improved multiscale 3D residual convolutional neural network is proposed.The three-dimensional convolution is improved and combines with multi-scale filters and residual units to extract the deep abstract features of the image.The method considers the balance between training samples and network depth,reduces the amount of parameters and effectively improves the classification accuracy.Secondly,aiming at the problem of image information loss caused by the continuous convolutional pooling operation in convolutional neural network,a hyperspectral image classification method based on multi-scale dilated convolutional neural network is proposed.The method introduces dilated convolution and constructs a multi-scale aggregate structure,which uses different dilated rates to extract image multi-scale features,and uses shortcut connections in multi-scale channels to accelerate convergence and prevent over-fitting.Thirdly,the performance of the proposed method is evaluated on the Indiana Pine data set and the Pavia University data set,and compares them with other classification methods.The experimental results show that the proposed method can extract better classification discrimination features and achieve higher classification performance.