Hyperspectral Image Classification Based On Low-rank Representation

Recently,hyperspectral image classification has become the main method of object detection in the earth's surface and been applied in many fields.However,there still exist several problems in hyperspectral image classification:(1)Labels in hyperspectral images are limted.(2)There are lots of types of spectral-spatial features.The dimensionality of them is high,which is difficult to process.(3)There are large differences in the area of land covers,which is not easy to be characterized with spatial features.Low-rank priori knowledge is explored after analyzing the characteristics of hyperspectral data.Low-rank representation(LRR)is employed in hyperspectral image spectral-spatial classification and the corresponding efficient algorithms are designed,in order to solve the three problems mentioned above.The contributions of this paper are as follows:(1)Semi-supervised hyperspectral image classification using local low-rank representation is proposed.Considering the difficulty in obtaining labels,semi-supervised learning(SSL)has been applied widely in this paper,in order to make full use of unlabeled pixels.Firstlt,the labels of unlabeled pixels are initialized by label propagation algorithm.Secondly,a local neighborhood is explored for each unlabeled pixel.Low-rank coefficients are obtained by low-rank representation(LRR)and thenlocal affinity matrix is built.Finally,semi-supervised classification is conducted based on the local affinity matrix.The proposed method is evaluated in Indian Pines dataset and Botswana dataset,Experimental results on Indian Pines and Botswana data set show that the proposed method achieves satisfying performance with different scales of training sets and avoids label ambiguity in boundary regions.(2)A robust discriminative extraction of multiple spectral-spatial features for hyperspectral image classification is proposed in this paper.The proposed method explores the property of different spectral-spatial features and the extracted features are robust to noise.Firstly,the proposed method acquires relationship among samples by LRR coefficients,and it constrains that the relationship among samples is retained in lwo-dimensional subspace.Secondly,the projection matrix for each type of feature is given a weight,in order to preserve their specific characteristics.Thirdly,an ideal low-rank O-1 matrix is designed to make full use of the discriminative information,which makes the LRR coefficients optimal.Finally,support vector machine is employed for classification after the low-dimensional features are obtained.The proposed method is evaluated in Indian Pines dataset?Urban dataset and Pavia University Scene dataset.Experimental results show that the proposed method outperforms some of the state-of-art dimensionality reduction and classification methods in hyperspectral image processing,even the data contains noise.(3)Hyperspectral image classification based on multi-scale spectral-spatial features(MSSFE)is proposed.In order to hanle the problem that the area of different land covers differs from each other,multi-scale strategy is used to characterize the distribution of land covers,which enhances the effectiveness of spectral-spatial features.Firstly,spectral-spatial features in different sizes of spatial neighborhood are extracted through multi-scale local weighted filters.Then robust discriminative multiple features extraction is employed for dimensionality reduction after concatenating features filtered with multiple scales.Finally,support vector machine is used for classification.Experimental results on Indian Pines and Botswana data set reveal that the proposed method is able to improve the accuracy of hyperspectral image classification.The classification result is better than that obtained by using single scale of spectral-spatial feaure.