A Classification Method For Hyperspectral Image Based On Conv-capsNet

Hyperspectral image classification plays an important role in environmental monitoring,agriculture,marine and other fields.The spectral information and spatial position information contained in it are the key factors for its classification.Recently,deep learning methods have shown very good performance in classifying hyperspectral images.In particular,Convolutional Neural Networks(CNN)can extract spatial and spectral features,but to extract these,convolutional neural networks often require deep network layers and a large amount of training data,which slows down the network's training speed.What's more,although CNN can extract such features,it cannot obtain the relative spatial position relationship between entities.Capsule Network(Caps Net)can extract pose and establish a hierarchical relationship between local and whole without requiring large training samples,but its structure has limited its ability to extract simple features.To solve these problems in CNN and Caps Net,this paper proposes a general network architecture which called Convolution-Capsule Network(Conv-Caps Net)that combines them.It increases the number of convolutional layers in the network,so it performs better in extracting information.In order to make it more targeted when dealing with the classification of hyperspectral images,and to obtain better classification results,this paper has made specific designs and improvements based on this general model.It uses three-dimensional convolution to replace the commonly used two-dimensional convolution,so it can overcome the limitation of the number of network layers that the convolutional neural network is subjected to when extracting feature information.It can extract the spatial and spectral features of the target without an excessively deep network.And because of the capsule layer,it can also extract relative spatial position features between targets that are difficult to be extracted by other neural network,and obtain a hierarchical expression of local features and overall features.And in this paper,the loss function proposed by Hinton et al.is optimized,a new loss function that can better measure the stability of the network is designed.This is beneficial to adjust the network in time and obtain better experimental results.In addition,in order to reduce the impact of redundant data on classification accuracy during network operation,this paper also proposes a method to reduce redundant data in hyperspectral images to a certain extent.It performs operations such as separation and edge extraction on different types of data in the data set,which reduces redundant data in training and improves the accuracy and authenticity of experimental results.Experimental results based on the Indian Pines data set demonstrate that the proposed method can achieve high classification accuracy over 91%,and achieves an accuracy of more than 98% on most classes.