Research On High-precision Classification Method Of Hyperspectral Remote Sensing Images

Hyperspectral remote sensing images have the strong practical significance,and contain a large amount of information about spectral dimension features and spatial information description,so it is often used to classify ground objects.Hyperspectral remote sensing satellites are widely used for monitoring and warning,and for resource exploration and crop management in the civilian field.Traditional hyperspectral remote sensing image classification technology can no longer adapt to the development of hyperspectral remote sensing images.At present,the important research is how to extract deep-seated and anti-jamming features in hyperspectral images for classification.The accuracy of the classification of hyperspectral remote sensing images is the cornerstone of hyperspectral remote sensing image processing technology and will directly affect the effectiveness of the image processing algorithm for subsequent identification,tracking and investigation.With the development of hyperspectral remote sensing technology,satellite resolution is getting higher and wider.The amount of hyper-spectral image data is increasing rapidly,but high-resolution spectral image transmission is extremely difficult due to limited by the transmission bandwidth of the star-earth link.These lead to the number of hyperspectral image data samples becomes small.For the above the research of hyperspectral remote sensing images in the field,this paper explores the high-precision classification technology combining spectral dimension information with spatial dimension information through the extraction of deep features.And the optimization settings of algorithm parameters are analyzed,and experimental verification analysis compared with relevant mainstream algorithms are carried out.In this paper,three kinds of hyperspectral classification methods are proposed for different typical application scenarios.Aimming at the lack of spatial information representation in common classification methods and low classification accuracy.A space-spectral classification method with enhanced confidence is proposed.The spectral enhancement algorithm is proposed to improve the anti-jamming ability and classification identification of the algorithm.On this basis,aiming at the problems of low classification accuracy and serious noise interference in the case of limited training samples,a space-spectral combination classification method based on adaptive guided filtering is proposed to further exploit the advantages of space-spectral combination and improve the classification accuracy of the algorithm.The first two methods are developed on the basis of machine learning in the post-classification stage,and require manual extraction of classification features.In order to improve the universality of the algorithm,a three-stage deep learning training method based on automatic weighted fusion of spatial and spectral features was proposed.Specifically,the main achievements of this paper are as follows:1.The traditional hyperspectral image classification techniques mainly focus on the processing of spectral information.However,the classification results produce impulse noise.In this paper,PCA is used to reduce the dimensionality of data to play the role of feature enhancement.Moreover,patch processing is carried out on the pixel points.Spectral information and spatial information are combined together to get super-pixel blocks.In order to improve the characteristic anti-jamming ability,Gabor filter is also used to enhance the spatial information and extract the texture features.Finally,the confidence iteration correction is adopted.By finding the maximum confidence,and SVM method is used to achieve hyperspectral classification based on space-spectral fusion.2.The deep convolutional neural network can achieve better classification effect,which uses residual network to extract features and combines with attention mechanism to enhance features,but the residual network is not good at extracting more features.In order to improve the classification accuracy,this paper introduces a dual-path network to replace the residual network for basic feature extraction,and proposes a three-stage method for model training.The dual-path network has the advantages of both low-level feature reuse of residual network and deeper and more feature exploration of dense connection network.On the basis of retaining the attention mechanism realized by the squeeze and excitation network,a Spatial-Spectral Squeeze-and-Excitation Dual-Path Network(SSSEDPN)is proposed.The three-stage training method focuses on the weighted fusion process of spatial and spectral features to highlight the part of the two features which is more conducive to target classification.It can further improve the classification accuracy of hyperspectral images.3.For the case of limited training samples,the traditional spectral classification methods have the problems of high computational complexity and low classification accuracy due to noise interference.A spatial and spectral combining classification method based on adaptive regularization factor guided filtering is proposed.The improved contrast adaptive LBP description extraction to improve the image gray value contrast description accuracy.And Gabor transform were used to extract spatial texture features.A coarse-to-fine classification framework was adopted,and a guided filtering algorithm based on adaptive regularization factor is proposed to improve the classification accuracy of ground features.Based on the initial classification binary image,an improved guided filter is used to improve classification accuracy and reduce noise interference.