Research On Band Selection Method For Spectral Unmixin

Hyperspectral images contain hundreds of bands,providing rich spectral information for ground object identification and analysis,and achieving great success in precision agriculture,ecological science and other fields.However,the high dimensional data of hyperspectral images also presents a number of challenges for image processing,such as data redundancy,low signal-to-noise ratio,and high spatial and temporal complexity.Therefore,dimensionality reduction is a very important task in the field of hyperspectral image processing.Band selection is a straightforward and effective method for dimensionality reduction,yet most current band selection methods are applied to image classification or target detection,ignoring the important application of unmixing,and they ignore the spatial structure of hyperspectral images and the role of correlation between bands.Therefore,based on the spatial structure of hyperspectral images,this paper focuses on the spectral unmixing band selection method.The main research contents are as follows:(1)Simplex-based band selection method.In order to further improve the performance and efficiency of unmixing,this paper proposes a simplex-based band selection method for spectral unmixing of hyperspectral images,including a band selection method based on the maximum volume of a simplex and a band selection method based on the maximum distance of a simplex,respectively.Both methods are based on the structure of hyperspectral data appearing as convex simplex in space,and take into account the interactions between the bands to find the combination of bands that make up the largest simplex volume and the largest simplex distance in space.(2)A band selection method based on simplex maximum volume optimization and dynamic programming.The simplex-based band selection method suffers from computational overload and long running times when selecting a subset of bands.Therefore,in order to improve the computational efficiency of the algorithm,this paper makes an in-depth study of the first two methods,and proposes an optimization algorithm based on the maximum volume of the simplex and a band selection method based on dynamic programming.The former proposes an optimisation algorithm to find the subset of bands that form the largest simplex volume in space;The latter combines dynamic programming with band selection,introducing an orthogonal subspace projection onto the simplex formed by the image in space,thus extracting the relatively pure feature information in the band.Finally,based on the spatial structure of hyperspectral images,a dynamic programming model is developed to find the optimal subset of bands under multiple constraints.(3)In order to verify the effectiveness of the band selection method proposed in this paper,unmixing experiments are conducted on real hyperspectral images and compared with some representative methods.Experimental results show that the proposed method is helpful to improve the efficiency and accuracy of unmixing,and is an effective dimension reduction method for unmixing.