Research On Dimension Reduction And Classification For Hyperspectral Images

With the maturity of hyperspectral imaging technology,its application fields are more extensive.Improving the processing power and classification performance of hyperspectral data is an inevitable trend of social development.It is of great significance to national security maintenance,rational use of resources,environmental monitoring and urban planning management.Classification is a key step in the application of hyperspectral images(HSIs).Since HSIs have a large number of spectral bands,but only a small number of training samples,classification of HSIs becomes a challenging task.However,the spectral bands of HSIs data are usually related.Therefore,before classification,HSIs data is generally dimension reduced.This paper focuses on the two issues of dimension reduction and classification for HSIs.In the aspect of HSIs dimension reduction,a new global-local discriminant analysis dimension reduction algorithm based on weighted spatial-spectral mean filtering(WW)is proposed.The algorithm makes full use of the spectral information and spatial features of HSIs data,as well as global and local structural features.Firstly,in order to enhance the correlation of adjacent pixels,the spatial-spectral features are used to perform weighted spatial-spectral mean filtering(WSSMF)on the pixels.Then,in order to construct the global-local discriminant matrix,we propose a weighted spatial-spectral local neighborhood preservation(WSSLNP)matrix.It is embedded into the regularized linear discriminant analysis(RLDA)model as a local regression operator to form a powerful dimension reduction algorithm.In the aspect of HSIs classification,a spatial-spectral cooperative classification algorithm(SC-NN)is proposed to make full use of the spectral information and spatial continuity of HSIs data.The algorithm first uses the nearest neighbor(NN)classifier to classify the test samples based on spectral information,and then use the spatial structure information to correct the classification map of the previous step.Furthermore,in order to solve the problem of few marker samples,this paper adopts the incremental learning strategy: iteratively selects the test samples with high classification confidence generated by SC-NN classifier and adds them to the training set,and then uses the new training set for HSIs classification.Combining WW dimension reduction algorithm with SC-NN classification algorithm and incremental learning strategy,a spatial-spectral cooperative and global local discriminant classification algorithm(WWI_SC-NN)is proposed.In order to reduce the influence of a large amount of noise in the data for the dimension reduction and classification algorithms,an efficient anti-noise clustering algorithm based on K-Medians is proposed and applied to the data preprocessing of SVM and WWI_SC-NN classification algorithms.REK-SVM classification algorithm and KWWI_SC-NN classification algorithm with stronger anti-noise ability are obtained.