Research And Application Of Deep Learning Capsule Network In Hyperspectral Remote Sensing Image Classification

Deep learning has been applied in many fields,helping it solve some difficult problems.Capsule network is a new deep learning model.In this model,the input and output of each capsule are vectors.High-level features are obtained by clustering low-level features through dynamic routing algorithm.Hyperspectral remote sensing image can provide spectral and spatial information,which often used in image recognition and classification.The use of capsule network to classify hyperspectral remote sensing images is the focus of this article.In this paper,based on the characteristics of hyperspectral remote sensing images,the capsule network is improved to make the model more suitable for the classification task of this type of image.For the research object,the basic characteristics of hyperspectral remote sensing images are analyzed,and the commonly used image classification and evaluation standards are introduced.For the network model,the basic structure and principle of convolutional neural network and capsule network are studied.Convolutional neural network can fully extract image features,but the pooling operation will lose information of the relative position relationship of the research object;the capsule network can retain the spatial location information of the research object,but its ability to extract the low-level features of the image is limited.On the basis of these two networks,the concept of multi-scale convolution kernel in Inception module is introduced to design a multi-scale 3D convolution-capsule(MS3DCONV-CAPSNET)model.In order to make the model more suitable for the classification of hyperspectral remote sensing images,the MS3DConv-Caps Net model uses a three-dimensional convolution kernel.For the dynamic routing algorithm,the classic dynamic routing algorithm is studied and improved.In the improved dynamic routing algorithm,the consistency measurement is carried out by using the length similarity and direction similarity of vectors respectively.This paper uses two data sets to train and test the MS3DConv-Caps Net model respectively,and compare the classification results with CNN through OA,AA,Kappa and other evaluation criteria.The values of OA,AA,and Kappa of the model on the Indian Pines dataset are 92.82%,92.49%,92.6%,respectively,which are 7.74%,7.84%,and 7.69% higher than the CNN model.The values of OA,AA and Kappa of the model in the data set of Pavia University are 94.77%,94.33% and 94.61% respectively,which are 8.27%,8.31% and 8.24% higher than those of CNN model.