Hyperspectral Image Classification Based On Edge-Preserving Filter And Deep Residual Network

Hyperspectral imaging spectrometer is a new type of optical remote sensor,it can obtain spatial and spectral features of hundreds of continuous narrow spectral bands.Therefore,it can generate nanoscale hyperspectral images.As an important part of hyperspectral image processing and application,the purpose of image classification is to put a unique class label for each pixel in hyperspectral images.However,the strong correlation between bands,the complex structure of ground objects and the insufficient sample size of hyperspectral images bring great challenges to the classification task.Therefore,a hyperspectral image classification based on edge-preserving filter and deep residual network is proposed.Firstly,principal components analysis is used to transform the original hyperspectral image into a group of unrelated data and realize image dimensionality reduction;Secondly,in order to increase sample diversity,the range weight is calculated by the first principal component in the joint bilateral filtering,which can improve hyperspectral image quality and fused with spectral features to obtain the spatial-spectral features of the image.Then,in order to avoid networks devolution and other problems caused by the increase of network layers,the convolutional layer is improved as a residual learning module by introducing hoplayer connection in 2DCNN,a deep residual network model is constructed to extract deep spatial-spectral features.Finally,these features are input into the model and the image classification is completed by Softmax classifier.In this paper,the proposed method is contrasted with other related state-of-art methods on two public datasets,and it achieves the highest classification accuracy,in which OA is98.87% and 99.35%,respectively,AA is 98.64% and 98.84%,respectively,and Kappa coefficient is 98.71% and 99.13%,respectively.The result shows that the method considers the important role of the edge structure to a certain extent,and the overfitting phenomenon in convolutional neural network classification is alleviated,so the method greatly improves the classification accuracy of hyperspectral images.This paper has 26 figures,9 tables and 54 references.