Research On Hyperspectral Image Classification Algorithm Based On Deep Learning

As the basis of hyperspectral image processing technology,hyperspectral image classification has always been a hot spot for scholars.However,hyperspectral image data has the characteristics of high dimensionality,information redundancy,different spectra of the same substance and foreign objects of the same spectrum,which brings great challenges to hyperspectral image classification.In recent years,many scholars have used deep learning methods for hyperspectral image classification.Although some good results have been achieved,there are still some problems: the hyperspectral image data presents a 3D structure with rich spectral information,but spatial information relatively few,how to make full use of spatial information to improve classification accuracy has yet to be explored.Hyperspectral data has the phenomenon of unbalanced distribution between classes,and it is difficult to find a dimensionality reduction method suitable for hyperspectral data.Hyperspectral images have high-dimensional characteristics,and some deep learning methods cannot be directly used for hyperspectral image data.By constructing a deep neural network model suitable for hyperspectral images,this paper studies the classification of hyperspectral images from the aspects of data dimensionality reduction,feature extraction,and feature fusion.It mainly includes the following three aspects:A classification method of hyperspectral image based on multi-scale convolutional neural network is proposed for explore the nonlinear characteristics of hyperspectral image data,and to alleviate the problems of gradient disappearance,overfitting,and precision reduction caused by excessively increasing the depth of the network to the traditional convolutional neural network.Firstly,the isometric feature mapping algorithm is used to process hyperspectral data to mine the nonlinear characteristics and maintain the inherent geometric properties of the data.Secondly,a multi-scale convolutional neural network with three 2D convolutional layers of different scales is set up the model extracts the shallow and deep spatial spectrum features of the hyperspectral image,and adds a residual network to the convolution operation to alleviate the problems of overfitting caused by the depth of the network.Finally,the Softmax classifier is used to complete the classification.Experimental results show that the proposed method has better classification performance.In order to extract more discriminative hyperspectral image features and optimize the classification performance of hyperspectral images,a hyperspectral image classification method based on multi-scale local binary mode and residual convolution is proposed.Firstly,use multi-scale local binary mode to preprocess the image,and then use the designed residual convolution network to process the preprocessed image to fully extract the spatial features of the hyperspectral image.Secondly,use the same residual convolutional network to process the original image to extract the spectral information of the image.Finally,the spectral information and spatial information are merged and fed into the Softmax classifier to complete the classification.The experimental results on Indian Pines and other data sets show that the proposed method can effectively improve the classification accuracy of hyperspectral images and solve the problems of easy confusion of similar spectral information.For solving the problem that the 3D convolutional neural network is not accurate in classifying hyperspectral images under the condition of limited training samples,a hyperspectral image classification method combining Gabor filtering and 3D/2D convolution is proposed.Firstly,3D Gabor filter is used to extract the texture information of the image,3D Gabor filter group can filter all the bands of the hyperspectrum at the same time,can extract the texture features while retaining the image spectrum information to ensure that the image spectrum information is not destroyed,and reduce the calculation amount.Secondly,use three different scales of 3D convolution operations to extract the shallow and deep spatial spectrum information of the image,and then use a 2D convolution to further enhance the spatial information.Finally,use the Softmax classifier classification.Experimental results show that compared with classification methods such as SSRN,the proposed method has better classification performance and is an effective hyperspectral image classification method.