Research On Classification Method Of Hyperspectral Image Based On KE-3D-CapsNet

Hyperspectral image originates from remote sensing technology and has a wide range of applications including geology,biology,medicine,and military.Hyperspectral image consists of spectral information and image information as threedimensional data that has the characteristics of redundant and huge data volume,plenty of spectral bands,and high correlation between bands.How to extract features is the key to the classification of hyperspectral images.Once deep learning came out,it has succeeded in processing text,speech,images,and video,and applying deep learning to higher-dimensional images has become a new challenge.The convolutional neural network is a basic and widely used network structure in deep learning,but it has the disadvantages of requiring of too many training samples,poor image ambiguity expression and ignoring information by the pooling layer.In contrast,the capsules in the Caps Net encapsulate the state information of all features,regardless of the lowlevel and high-level,in the form of vectors.Compared with scalars,using vectors as the input of the network can be more completely utilized.The image information retains more features.The low-level capsules transmit information to the high-level capsules through a dynamic routing mechanism.Layers are abstracted and classified to obtain classification results.Based on the above reasons,this paper combines the nonlinear dimensionality reduction method KPCA with the extended morphological contour method(EMP),and uses 3D-Caps Net to classify hyperspectral images named as KE-3D-CapsNet,First,for the characteristics of huge hyperspectral image data volume and redundant band information,the nonlinear dimensionality reduction method Kernel Principal Component Analysis(KPCA)is used to reduce the dimensionality of the spectral information to obtain the principal component that represents the characteristics of the original image information;The morphological theory in the digital image processing method extracts the extended morphological profile(EMP)from the spatial information,and perform edge detection to obtain the three-dimensional spectral spatial information;Finally,the obtained information is iterated using the dynamic routing algorithm in 3D-Caps Net.After compression by the squashing function,the probability that the input image pixels belong to various types is obtained.After comparison,the class with the largest probability value is selected as the classification result.KE-3D-CapsNet conduct hyperspectral image classification experiments on the common data set Indian Pines,Salinas and Pavia University,and conduct comparison experiments with other general classification methods under the same experimental conditions by evaluation criteria OA,AA and Kappa coefficient The analysis of experimental results proves that this method has higher classification accuracy than other methods.