Dimensionality Reduction For Hyperspectral Image Based On Graph Embedded Framework

With the rapid development of remote sensing and hyperspectral imaging technology,the resolution of hyperspectral images(spatial)and the number of observable bands(spectral)have increased a lot.Therefore,hyperspectral remote sensing data can contain fine spatial information and abundant inter-spectrum information on the Earth's surface.Because this information can reflect the physical and chemical properties of objects on the surface of the Earth,indivisible features that are originally in the visible spectrum can be identified in hyperspectral remote sensing data.At the same time,it also brings huge challenges to the processing of hyperspectral data,one of which is the impact of higher dimensions of hyperspectral data on classification accuracy.In high-dimensional hyperspectral data,there are problems such as redundancy,noise,and dimensional disasters.However,the dimension reduction,ie,dimensionality reduction,is an effective way to solve such problems.On the one hand,data dimensionality reduction can find low-dimensional structures that are hidden in high-dimensional data.On the other hand,it can reduce processing costs and reduce the burden on applications such as classification.This study is based on the graph embedded dimension reduction framework,deeply studying the construction of the graph and its application in dimension reduction,and then using the reduced dimension data,input support vector machine(SVM)classifier for classification processing.The main work of the thesis is reflected in the following aspects:Firstly,aiming at the assumption that the same object exists in the same spectrum in hyperspectral data,and using the spectral similarity to design the similarity distance to contruct a graph,a graph embedding and dimension reduction algorithm based on the spectral similarity is proposed.The algorithm makes full use of the inter-spectral information of hyperspectral data to make full use of the original data.Secondly,aiming at the existing low-rank graph-based discriminant analysis algorithm,which only uses the global information and does not consider the local information,a local preserving low-rank graph-based discriminant analysis algorithm is proposed.Based on the low rank graph,this algorithm introduces the method of preserving local information in the local preserving projection algorithm,which makes the improved algorithm can fully combine the global and local characteristics of the data.Experiments show that the improved dimension reduction algorithm can greatly improve the classification accuracy.Thirdly,the existing top-view graph embedding dimension reduction algorithms include sparse graph embedding,and cooperative graph embedding and low-rank graph embedding use the l1norm,l2 norm,and kernel norm respectively to construct graphs.However,a single graph can not obtain the global optimal and local optimal information of the data at the same time.Considering the advantage of using three graph embedding algorithms at the same time,a multi-graph fusion algorithm that fuses three graphs is proposed,and the classification accuracy of the algorithm is improved from the perspective of feature fusion and decision fusion,respectively.