Hyperspectral Images Classification Based On Improved Capsule Neural Network

Resently,hyperspectral images classification methods based on convolutional neural networks have achieved outstanding performance.However,due to the complexity of hyperspectral images,the classification methods based on convolutional neural networks is complicated,resulting in the problems of large parameters and low efficiency.Moreover,the cost of labeling hyperspectral images increase with the spatial resolution continues to improve,resulting in small sample problems.In order to solve the problem of complex network structure,this paper proposes two feature extaction methods of hyperspectral image,i.e.,separable 3D-convolution method and 1D-convolutional capsule method,and constructs a simple and efficient hyperspectral image classification method.The separable 3D-convolution method replaces the standard 3D-convolution to extract spectral-spatial information from hyperspectral images.It decomposes the standard 3D-convolution filter into the uncross-channel 2D-convolution filter and the cross-channel 1D-convolution filter,which not only reduces network parameters but also simplifies the network structure.The 1D-convolution capsule method provides the capsule-wise local feature extraction capability of capsule neural networks.It can be understood as the extension of the 1D-convolution on capsules,and have characteristics of local connection,parameters sharing and distributed representation.The 1D-convolution capsule method can avoid the influence of noise on the classification accuracy,thereby further improving the accuracy of capsule neural networks.This paper constructs three classification network to verify the effectiveness of proposed methods.Experimental results on three representative hyperspectral datasets show that the classification method proposed in this paper is superior to 9 well-known hyperspectral image classification methods in terms of accuracy and efficiency.