Research On Band Selection Of Hyperspectral Image Based On Classification Accuracy Prediction

Hyperspectral image processing is the frontier field of modern remote sensing technology development,and its practical application in various fields is increasingly extensive.However,due to the characteristics of high spectral resolution and large number of bands,there are a lot of redundant information among bands,which will occupy a large amount of physical space in storage,and the processing time is long and the efficiency is low.Because of the high dimensionality of hyperspectral image,the limited sample of landmark is prone to dimension disasters when recognizing,detecting and classifying landmarks.In order to reduce the dimension of hyperspectral image,it is necessary to remove the redundant data.The band selection of hyperspectral image can not only reduce the dimension of the original image,but also retain the physical characteristics of the original band.Therefore,this thesis focuses on the characteristics of hyperspectral images for classification and two new band selection methods for hyperspectral images are proposed as follows:Band selection based on K-AP algorithm for hyperspectral images.K-Affinity Propagation(K-AP)algorithm is a highly efficient clustering algorithm,which has been successfully applied in the fields of face recognition and data analysis,but it has not been successfully applied in the field of hyperspectral image analysis.Through in-depth study of K-AP algorithm,this thesis proposes to apply K-AP algorithm to hyperspectral image band selection and effective data compression for hyperspectral image.Firstly,based on the characteristics of the K-AP algorithm,we define a new similarity matrix based on the Kullback-Leibler divergence to measure the bands,and then use the K-AP algorithm to cluster and select the most representative bands.The correlation and redundancy information among the original hyperspectral image bands are effectively reduced.The experimental results show that the proposed method has better performance than that with other popular band selection method.Sparse subspace clustering algorithm is a useful algorithm in data dimension reduction.It mainly uses sparse optimization model to find the sparse representation and sparse parameters of data,and then uses spectral clustering method to cluster high-dimensional data into low-dimensional subspace.Through in-depth study of sparse subspace clustering algorithm,this thesis proposes a band selection method based on sparse subspace clustering algorithm.However,because of the high correlation among the band of hyperspectral image,as well as the spatial-spectrum integration,we improved the sparse subspace clustering algorithm for the characteristics of hyperspectral image,which combined with the spatialspectrum information of hyperspectral image to select the band,thus decrease the band correlation,effectively reducing the redundancy among the bands of hyperspectral image.What's more,the experimental results show that the correlation between the bands of hyperspectral image is narrowed.The commonly used band selection method has a better performance.Two different band selection methods proposed in this thesis have their own advantages.The band selection method based on K-AP algorithm can quickly select the target band subset.Compared with the commonly used band selection methods,it has better classification performance.The band selection method combined with sparse subspace clustering of spatial spectrum information uses L2 norm optimization model to effectively prevent over-sparse.The emergence of sparseness and band selection based on spatial-spectrum information of hyperspectral images can not only select a subset of the target band,but also have strong robustness and excellent classification performance.