Hyperspectral Image Classification Based On Spatial-spectral Feature Extracted By Deep Learning

As the remote sensing technology developed,hyperspectral remote sensing has been widely used in urban surveying and mapping,forest monitoring,meteorological prediction,environmental monitoring and so on.Among them,hyperspectral image classification is a key technology for interpreting and analyzing hyperspectral images.It is a research hotspot in hyperspectral image processing as well.And therefore,it has significant academic research and practical application value.In recent years,in view of the characteristics of hyperspectral images such as non-linearity and "image-spectrum merging",deep neural network has been introduced into the field of remote sensing to better extract the nonlinear and deep features of hyperspectral images,so as to make it more suitable for the classification of hyperspectral images and show good performance.However,the existing deep learning-based classification methods still have the problems of high feature dimension,single convolution scale,and insufficient utilization of spatial context information.Meanwhile,the correlation between feature maps in the extraction of spatial features and the context information of the spectral sequence are not considered in the extraction of spectral features,which leads to the need to improve the classification accuracy.In terms of the above problems,this paper has carried out relevant research work,mainly including:(1)A hyperspectral image classification method based on random multiscale convolutional network is proposed.In this paper,a multiscale dimensionality reduction module is introduced in the feature dimensionality reduction process,which takes the diversity in different homogeneous regions into account.For this reason,it can simultaneously extract low-level spectral information and spatial information,as well as obtain a better dimensionality reduction image.Meanwhile,the method constructs feature extractors of different scales in a random convolutional network,which contributes to overcoming the limitation of a single convolutional scale,using the spatial context effectively,and extracting the multiscale feature information of the image directly.Experimental results indicate that this method can primely enhance the classification accuracy.(2)A hyperspectral image classification method based on attention module and multi-grouping strategy is proposed.In this paper,a channel spatial attention module is added to the multiscale convolutional network for spatial feature extraction,which can distinguish the output features of the convolutional layer,focus on more useful features,and enhance the expressive ability of the model.Meanwhile,spectral multi-band grouping strategy has been added in the extraction of the long short term memory network,which can help to consider local and global spectral information,analyze the context information of spectral sequence and extract spectral features more fully.The experimental results verify the feasibility and accuracy of this method in the hyperspectral images classification and obtain better classification results.