Semisupervised Classification For Hyperspectral Remote Sensing Image

With the development of imaging spectrometer,the remote sensing technology provides spectral characteristic information and also provides the space characteristics of the image at the same time.On this basis,hyperspectral remote sensing images with high spectral resolution,has been used widely in practice.Hyperspectral remote sensing image has much many bands and large amount of information,so we can obtain more fine spectral characteristics of target object,which brings advantages and challenges in classification and detection of hyperspectral remote sensing image.According to the characteristics of hyperspectral remote sensing images,the following questions should be considered:(1)The high number of spectral channels.Hyperspectral remote sensing images have dozens or even hundreds of wavelengths,in result of huge data and high dimension;(2)The high cost of true sample labeling.Labeling samples is a matter that cost human and resources.Hughes proved that when the number of training samples is 6~10 times more than the dimension of data,the classifier works better.As a result,it is difficult to obtain enough and representative true samples.(3)The spatial variability of the spectral signature.The spectral signature is affected easily by its types and changing surrounding environment.The traditional classification methods are generally divided into unsupervised methods and supervised methods.Unsupervised classifications are lack of prior knowledge,so the classification result is not good.Under the conditions of insufficient or poor quality of the training samples,The generalization of supervised classifier that dependent on the training samples is decreased.Considering the above problems,this papers mainly study semi-supervised classification algorithm of hyperspectral remote sensing image,which make full use of the huge amounts of the unlabeled samples.At the same time,kernel-based classifiers are able to handle large input spaces efficiently and deal with noisy samples in a robust way,and we incorporate spatial information to ease the spatial variability of the spectral signature.On this basis,this paper put forward two semi-supervised classifications methods as follows:1 This paper proposes a semi-supervised classification algorithm which combining Learning with local and global consistency(LLGC)and Least squares support vector machine(LS-SVM).LLGC algorithm is a kind of graph-based method,which pass the label information according to the degree of similarity between samples.On the basic idea of thealgorithm,this paper proposes that using LLGC algorithm instead of KNN method labeling neighboring samples of the labeled samples,then expanding these samples to the LS-SVM training sample set and training the LS-SVM classifier.The proposed algorithm not only avoids well the defect of graph-based method,such as the high complexity of time and the transductive learning,but also expands the training sample set of LS-SVM and improves the classification accuracy and generalization ability.Experimental results show that the proposed algorithm improves the classification accuracy and Kappa coefficient.2 It is proposed to introduce dissimilarity in Laplacian support vector machine(Diss-LapSVM).Laplacian support vector machine(LapSVM)is a semi-supervised method that combines the traditional regularization framework and manifold assuming,as the same as the traditional support vector machine(SVM),LapSVM has much advantages,such as simple structure,fast calculation speed,and strong robustness.What's more,LapSVM make full use of the huge amount of unlabeled samples' geometry information by adding the manifold regularization item,which can adjust the classification hyperplane together with labeled samples.Meanwhile,in order to introduce appropriately distribution of unlabeled samples,this paper provides linear neighborhood propagation(LNP)to construct graph Laplacian matrix.Considering the spatial variability of the spectral signature,this paper puts forward to add dissimilarity information to machine' manifold regularization term,which restrains the influence of the spatial variability effectively.The results illustrated that the method can improve the classification accuracy,especially for samples that have similar spectral features.In this paper,the two proposed algorithms have different features: The first is a combination of two kinds of supervised classifier,first of all,we train a kind of classifer using original labeled sample,then choose some samples with high confidence level to extend another classifier' training samples;The second algorithm add unlabeled samples to the objective function directly,we can use the labeled and unlabeled samples together to get a more reasonable classification hyperplane.The two algorithms are kernel method,which have strong robustness in dealing with high-dimensional hyperspectral remote sensing image,and can solve the problem of insufficient training samples.