Hyperspectral Image Classification Based On Deep Network And Guided Filtering

Hyperspectral image is an image obtained by simultaneous imaging of a specific region on dozens or hundreds of successive bands by hyperspectral sensor,and the classification of hyperspectral images is an important branch of hyperspectral image technology.Hyperspectral image has the characteristics of a wide range of bands and high spectral resolution,which provide a good help for the classification of ground objects.However,it faces problems such as high data dimensions,multi-band redundancy,and lack of enough training samples.Therefore,how to effectively use these information and extract hyperspectral image features to improve the classification results of hyperspectral images has become an important issue in this field.In order to study how to simplify the complex and redundant hyperspectral image information,and fully extract and use the hyperspectral image features to achieve better classification performance,this paper proposes several hyperspectral image classification methods based on deep network and guided filtering.The main work is as follows:(1)A hyperspectral image classification method based on denoising,combined dimensionality reduction and guided filtering is proposed.In this method,the original hyperspectral image is transformed by minimum noise fraction rotation to achieve initial denoising,which is then processed by dimensionality reduction using principal component analysis and used as the input image of guided filtering.The guided image is obtained by dimensionality reduction of the original hyperspectral image using independent component analysis.After that,the input image and the guide image are input to the guide filter with the increasing of filter radii to perform multi-scale filtering operation,so as to obtain the multiscale spatial features.Finally,the support vector machine is used for classification.This method mainly solves the problems of insufficient utilization of spatial features and difficulty in classification of small samples by the existing techniques,which reduces the computational complexity,and improves the classification effect.(2)A spectral-spatial classification method of hyperspectral images based on deep adaptive feature fusion is proposed.This method implements the deep adaptive fusion of two hyperspectral features,and then it performs spectral-spatial classification on the fused features.A U-shaped deep network model with the principal component features as the model input and the edge features as the model label is designed to adaptively fuse two kinds of different features.One comprises the edge features of hyperspectral images extracted by the guided filter,and the other comprises the principal component features obtained by dimensionality reduction of hyperspectral images using principal component analysis.The fused new features are input into a multi-scale and multi-level feature extraction model for further extraction of deep features,which are then combined with the spectral features extracted by the long short-term memory(LSTM)model for classification.The U-shaped deep network adaptively fuses two different features,which improves the problem of low classification accuracy caused by a single feature.The spectral-spatial classification of fused features further improves the classification results of hyperspectral images.(3)A spectral-spatial classification method for hyperspectral images based on image enhancement and deep network is proposed.The method uses two-dimensional convolutional neural network to enhance the low resolution features of the input image and obtains the high resolution features of the enhanced image.Considering the original hyperspectral image has many bands and scattered features,this method performs guided filtering transformation on hyperspectral image to obtain multi-scale filtered images with low dimension,concentrated features and image edge information,which are then enhanced and input into the classification network to improve the classification accuracy.The enhancement technology effectively uses the automatic learning and updating characteristics of neural networks,which makes the edge information of the enhanced multi-scale filtered images clearer and the difference between classes more obvious.At the same time,the performance of image classification is improved.The model design of this method in image enhancement is universal,which makes this method has good generalization ability.In addition,compared with other methods using three-dimensional convolution neural network,the calculation cost of two-dimensional convolution model of this method is lower.