The Research Of Sparse Subspace Clustering Algorithm And Its Application In Hyperspectral Image

With the rapid development of hyperspectral remote sensing technology in recent years,it is possible to extract the feature details of hyperspectral lands.And the hyperspectral image(Abbreviated as HSI)effectively combines the spatial information,spectral information and radiation energy of materials.Since HSI lands contains the intrinsic property and different lands have their own electromagnetic wave reflection characteristics,the different reflection characteristics from different materials are the theoretical basis for HSI clustering.The methods of HSI clustering based on statistical learning theory and machine learning are the main development trend in recent years.The HSI clustering techniques are able to determine the clustering of unknown lands according to electromagnetic spectrum reflection characteristics of different lands,which can obtain spatial size and location distribution information of different lands and provide people with determination of some unknown areas of interest.Therefore,the HSI clustering techniques have a strong practical significance in the field of atmospheric monitoring and military reconnaissance.The methods based on sparse subspace clustering are very popular for HSI clustering in recent years.Sparse subspace clustering(SSC)algorithm is a clustering framework based on spectral clustering(SC),which can obtain low dimensional structure from high dimensional data via the Laplace feature mapping and further obtain clustering results via SC.To some extent,SSC algorithm can perform dimension reduction for high dimensional HSI samples,which is beneficial to clustering.Directly applying SSC algorithm into HSI clustering usually causes that the clustering accuracy is not high and the geometric structure information including spatial information in HSI is not fully extracted.Moreover,traditional unsupervised clustering has limited the improvement of HSI clustering precision due to lacking of the guidance of supervised information.Based on the above analysis of applying SSC to HSI,this thesis mainly makes the following improvements based on SSC algorithm.Firstly,considering abundant spectral information,spatial information and high dimensional properties of HSI data,traditional SSC algorithm only utilizes the single sparse representation coefficients of HSI samples to construct similarity matrix and further obtains clustering results via performing SC.In essence,the single sparse representation coefficient lacks the utilization of overall information and do not capture the structural information of HSI data including rich spectral information and spatial information,which leads to the reduction of clustering precision.Consequently,we adopt the obtained sparse representation vectors of HSI samples to construct cosine similarity matrix.Moreover,we further build Cosine-Euclidean similarity matrix(CE)based on cosine similarity matrix,which merges unique spatial information of HSI.Theoretically,there are some factors,e.g.the atmospheric effects and angle of the sun which can influence the valid utility of HSI information in clustering.So we propose the dynamic weighting adjustment to construct similarity matrix based on sparse representation vectors and cosine measure,which is called Cosine-Euclidean dynamic weighting similarity matrix(CEDW).The CE and CEDW methods can combine rich spectral information and unique spatial information of HSI,which better improves the precision of HSI clustering.Secondly,considering the limitation of clustering precision via traditional SSC,we use a little supervised information of HSI samples to deliver known label information to a great number of unknown HSI samples via sparse representation classification(SRC),which makes up the class probability of HSI samples and predicts the class information of HSI samples.Moreover,with prior knowledge,the class probability model can specify the membership of the unknown sample corresponding category,improve the compactness of block diagonalization of similar samples and expand the consistency of similar samples.Finally,the merger of sparse representation and class probability well promotes the improvement of clustering precision.Based on above theoretical analysis,we conduct a variety of experiments via our proposed model on public HSI data sets including Pavia University,Indian Pines and Pavia Centre.The experimental results show that our proposed methods based on SSC are favor of HSI clustering,make block diagonalization of sparse representation coefficients clearly and demonstrate the effectiveness of the proposed methods.