Research On Remote Sensing Image Classification Based On Sparse Theory

With the development of remote sensing technology,the spectral resolution,temporal resolution and spatial resolution of remote sensing images are gradually improving.People can obtain more useful information from remote sensing image.At present,remote sensing technology has been widely used in military,environmental monitoring,agriculture and other fields.In recent years,sparse theory has been applied in remote sensing images as a research hotspot in image processing.Based on the research of sparse theory,this paper studies two kinds of remote sensing images about multi-spectral and hyperspectral,in order to improve the classification of mangrove remote sensing images.The main contents are as follows:(1)For multi-spectral remote sensing images,a remote sensing image classification algorithm based on weighted neighboring smoothing sparse representation is proposed.Firstly,aiming at the special nature of mangrove remote sensing imagery,it fuses many features such as geography and texture except the spectral features.Then K-singular value decomposition(K-SVD)algorithm is used to train the sample data to form an over-complete dictionary,the Manhattan distance average value represents the similarity among pixels,and different weights are assigned to different neighboring pixels through a weighting function.Finally,through the orthogonal matching pursuit algorithm(OMP)to solve sparse coefficient,and the category of the pixels is determined according to the reconstruction residuals.The effectiveness of the algorithm is demonstrated on the TM multispectral remote sensing image dataset of Zhangjiangkou Mangrove Nature Reserve.(2)For hyperspectral remote sensing images,a hyperspectral dimensionality reduction algorithm based on supervised sparse embedding preserving projection is proposed.Due to the large number of hyperspectral bands and the large amount of computation,dimensionality reduction is an important research hotspot in hyperspectral images.The traditional Sparse Preserving Projection(SPP)is an unsupervised dimension reduction algorithm.Based on the manifold structure information of the sample,the weight information matrix is built by using the tag information to highlight the inherent local manifold effect of the labeled sample.Considering the same sample spacing based on the SPP objective function,that is,the reconstruction residuals of the objective function and the similar sample spacing are both minimized at the same time.The effectiveness of the algorithm is demonstrated in the hyperspectral public data set Indian Pines and Zhangjiangkou Mangrove Nature Reserve HJ1A-HSI.Remote sensing image classification research based on sparse theory has good prospects and development,and this paper applies it to remote sensing images of mangrove wetlands.According to different remote sensing images,different decision-making methods are proposed,which have significant theories and actual value for wetland conservation monitoring and classification.