Research On Hyperspectral Target Extraction Algorithm Based On Density Peak Clustering

Hyperspectral sensors have limited spatial resolution and complex features.Each pixel contains more material information,resulting in a large number of mixed pixels in hyperspectral remote sensing images.In order to accurately identify and classify hyperspectral data,the premise is that the representative target pixels can be accurately extracted in the image,so the research of target extraction methods has become a hot spot in the hyperspectral field.This paper is based on a density peak clustering algorithm to statistically distribute the distance and density of each pixel of hyperspectral remote sensing image,and analyze the different attribute pixels to effectively and quickly divide the noise point,category center and background pixels.Based on this clustering idea,this paper studies and improves the hyperspectral image target extraction algorithm.The main work of the thesis is divided into the following three aspects:(1)The density peak clustering method is extended to hyperspectral image,the spectral distance is used instead of the pixel value difference to analyze,and different types of noise are added to the simulated image,and the distribution characteristics of different types of object objects in the density peak decision graph are analyzed.And the effect of different noise conditions on the density peaks in order to obtain better candidate regions in the target extraction.(2)Combined with the density peak clustering algorithm,the automatic target generation process algorithm and the simplex growing algorithm are C-ATGP and C-SGA algorithms to reduce the impact of burst noise on target extraction during imaging.ATGP and SGA are specific target extraction methods and are sensitive to burst noise.In order to accurately extract the representative target pixels from the image,before using the ATGP and SGA algorithms to extract the target pixels,the image is preprocessed according to the results of the density peak decision map,in order to remove the uninteresting pixels.Interference,thereby improving the effectiveness of target extraction.Experimental results on simulated and real hyperspectral images show that the effectiveness of using ATGP and SGA to extract targets in preprocessed hyperspectral images is better than that of direct extraction in original hyperspectral images.(3)Based on the density peak clustering algorithm,a CFDP-ATGP algorithm for target extraction by category is proposed.The core idea of the algorithm is to select candidate target pixels in each cluster,mainly to extract the more pure target pixels in the category.The experimental results on the simulated and real hyperspectral images show that the target extracted by CFDP-ATGP algorithm is better than the traditional SGA and ATGP algorithms,and the target extracted from C-SGA and C-ATGP algorithm is closer to pure pixel(endmember).(4)In order to reduce the influence of complex background on the extraction of target categories of interest,three weighting methods are proposed based on the characteristics of density peak clustering algorithm,and the ATGP,NMF and SGA algorithms are improved respectively.For the non-interesting pixels and complex background pixels,the weights are assigned a small weight,and the pixels with high density are assigned a larger weight,thereby improving the effectiveness of the target extraction.In the simulated image and the real image,the validity of the three weighting methods of the W-ATGP,W-SGA,and W-NMF weighting algorithms is verified.The experimental results show that the effectiveness of several weighted algorithm target extractions is improved in the original algorithm.