Research On Sparse Unmixing And Classification Algorithm For Hyperspectral Image

With the rapid development of remote sensing technology,hyperspectral remote sensing image data has been widely used in environmental monitoring,mineral exploration,military target recognition and other fields.Therefore,the processing of hyperspectral remote sensing image data is very important and has practical application value.The unmixing accuracy of hyperspectral image data and the classification accuracy of hyperspectral remote sensing images affect the follow-up application of hyperspectral remote sensing image data.Therefore,the unmixing and classification of hyperspectral images is a key issue in hyperspectral remote sensing image processing,and an important basis for follow-up research and application.Among them,in the hyperspectral unmixing algorithm,the unmixing algorithm based on sparse constraints has become a hot spot in the unmixing algorithm.The existing multi-objective sparse unmixing algorithm can solve the problem that the non-convexl₀norm and weight parameters cannot be adaptively selected in the traditional sparse unmixing algorithm.However,there are still some shortcomings of the random grouping strategy and the unicity of the knee point selection,resulting in a problem of hyperspectral unmixing accuracy is not high.In the hyperspectral image classification algorithm,the existing hyperspectral image generative adversarial networks?GAN?classification algorithm can solve the problem that the hyperspectral image data has large redundancy and small training set,but the algorithm still has the problem of not being able to extract all spectral feature and spatial-spectral feature,so the accuracy of hyperspectral image classification needs to be improved.In this paper,the following two aspects are studied in view of the deficiency of existing hyperspectral image sparse algorithm and hyperspectral image classification algorithm:Firstly,aiming at the problem that the hyperspectral image unmixing accuracy is not high due to the lack of random grouping strategy and the unicity of the knee point selection in the existing hyperspectral image multi-objective sparse unmixing algorithm,this paper proposes a hyperspectral image sparse unmixing algorithm based on the evolutionary algorithm for large-scale many-objective optimization?LMEA?.For the first time,the decision variable grouping strategy in LMEA is adpoted,and a constrained knee point area selection strategy is proposed to improve the accuracy of hyperspectral sparse unmixing.In this paper,the algorithm is applied to the two-objective sparse unmixing framework based on spectral feature and the three-objective sparse unmixing framework based on spectral-spatial feature.Experiments on common simulated and real hyperspectral image datasets show that the proposed algorithm obtained the best unmixing accuracy.It is also proved that the algorithm is more robust to noise by using the spectral-spatial feature.Secondly,aiming at the problem that extracted spectral features and spatial-spectral features are not comprehensive existing in the existing hyperspectral image GAN classification algorithm,which leads to the low classification accuracy of hyperspectral image,this paper proposes a hyperspectral image classification algorithm based on two-channels GAN.First of all,an improved one-dimensional GAN classification framework and an improved two-dimensional GAN classification framework are designed and constructed to extract more comprehensive spectral features and spatial features,respectively.Based on the above two frameworks,the idea of two-channels is adopted in hyperspectral image GAN classification model for the first time.The two-channel GAN classification framework is designed and built to extract more comprehensive spatial-spectral features and send them into the classifier,thereby improving the classification accuracy of the hyperspectral image.By experimenting with three commomly used hyperspectral datasets,the proposed algorithm obtains the best classification results compared with other algorithms,and verifies the effectiveness and advancement of the proposed algorithm.