Research On Classification Of Hyperspectral Remote Sensing Image Based On Spatial-Spectral Feature Fusion

Hyperspectral remote sensing image has ultra-high spectral resolution,and can obtain information of hundreds of continuous spectral bands of ground objects,thereby greatly improving the ability to distinguish ground objects.Hyperspectral remote sensing plays an important role in national defense construction and national economy,etc.It has been widely used in precision agriculture,geological prospecting,marine fields,grassland forest fields and environmental monitoring fields.Many hyperspectral remote sensing applications are based on image classification,but the characteristics of hyperspectral remote sensing images,such as high dimensional,high correlation between bands and small number of samples,bring challenges to the classification task,such as dimensional disaster and spatial homogeneity and heterogeneity.In view of the above problems,this paper combines new theories and methods in computer vision,machine learning and deep learning to study the spatial-spectral feature classification methods of hyperspectral remote sensing images.The main research contents are as follows:(1)A spatial-spectral joint classification method of hyperspectral remote sensing image with locality and edge preserving is proposed.In this method,firstly,the first principal component image extracted by principal component analysis is used as the guidance image,and the spatial characteristics of the image are extracted by guided filtering.Then,the spatial feature is embedded in the low dimension through local Fisher discriminant analysis,and the local information of the data is taken into account while the dimension is reduced.Finally,the obtained low-dimensional embedded features are combined with random forest for classification.The experimental results using two hyperspectral images of Indian Pines and Pavia University show that this method can obtain higher classification accuracy compared with other related methods.In the case of randomly selecting 10% and 1% samples from various ground objects as training samples,the overall classification accuracy is increased to 99.57% and 97.79%,respectively.The proposed method takes into account the edge structure information of the image when extracting spatial features,and introduces the local relationship between pixels when performing low-dimensional embedding,extracts more effective spatial-spectral features,and improves the classification accuracy.(2)A hyperspectral image classification method based on guided filtering and autoencoder network is proposed.This method first uses the first principal component image extracted by principal component analysis as a guidance image,and uses guided filtering to extract the spatial features of the image and fuse it with the spectral features to obtain the spatial-spectral features of the image.Then,the spatial-spectral features are input into the deep autoencoder network to extract deep features.Finally,the image is classified by support vector machine.Experiments with two hyperspectral images from Pavia University and Salinas show that this method can obtain higher classification accuracy compared with other related methods.When 3% and 1%samples are randomly selected from various ground objects as training samples,the overall classification accuracy is improved to 98.37% and 99.43%,respectively.The method considers the edge information of the image in spatial feature extraction.The stacked autoencoder network is used to perform deep-level feature extraction on the spatial-spectral features,and low-dimensional embedded features that are beneficial to the classification of ground objects are extracted,which improves the classification accuracy.(3)A hyperspectral image classification method combining joint bilateral filtering and convolutional neural network is proposed.First,the original image is compressed into three bands through the principal component analysis.Then,the first principal component image is used as the guidance image,and the first three principal component images are enhanced with joint bilateral filtering.Finally,the neighborhood information of the pixels in the enhanced image is extracted and input into the convolutional neural network for deep feature extraction and classification.The experimental results using two hyperspectral images of Indian Pines and Pavia University show that this method can obtain higher classification accuracy compared with other related methods.When 10% and 3% samples are randomly selected as training samples,the overall classification accuracy is improved to 98.47% and 99.30%,respectively.The proposed method can enhance the edge information of the input image,and the constructed network model can extract deep features that are beneficial to classification.The experimental comparison results with related deep learning methods show that the proposed method can improve the accuracy of image classification.To sum up,the proposed classification methods have certain theoretical basis,which can make full use of the spatial information of images to improve the classification accuracy of images and lay a foundation for many applications of hyperspectral images.There are 69 figures,13 tables and 142 references in this paper.