Research Of Feature Extraction And Classification Methods For Hyperspectral Images

Hyperspectral remote sensing technology can capture image data in hundreds of narrow spectral bands.Hyperspectral image is rich in spectral information and spatial information to explore the physical and chemical properties of the objects.Therefore,the objects,unrecognized in the wide band remote sensing field,can be identified in hyperspectral image.However,there are some challenges in classification applications,such as the properties of high band dimension,strong band correlation,a large dataset and the limited labeled samples.The research of feature extraction and classification methods for hyperspectral image has attracted more attention of domestic and international scholars and become one of the key technologies in hyperspectral image processing and a focus study of many fields.In order to improve the classification performance for hyperspectral image,this paper deeply studied the feature extraction methods and classification algorithms,with the support of National Natural Science Foundations of China(No.61231016,No.61301192,No.61152004).The main research work and innovations are listed as follows:1)An automatic band feature extraction method based on a novel wrapper multiple improved particle swarm cooperative optimization and support vector machine model(MIPSO-SVM)is proposed.In the model,firstly,an improved particle swarm algorithm(IPSO)is proposed by a new update strategy of position and velocity.IPSO uses discrete particle swarm optimization to update the band coding part of particles,and updates the SVM parameters by a strategy of random disturbance.Then,in the process of cooperative evolution,genetic algorithm(GA)is incorporated to improve the premature convergence of PSO.The experimental results on Indian Pines data and Pavia University data demonstrate that the proposed method select best band combination and the optimal SVM parameters,and thus improves classification accuracy.2)The classification performance of sparse-based classification method for hyperspectral image is easily affected by the power of sparse representation dictionary on the sample.Firstly,a new discriminative sparse representation classification approach using spectral data and extended local binary patterns texture is proposed.In this method,an improve LBP coding(HLBP)for hyperspectral image is developed,and a new optimization problem for dictionary learning is constructed by combining the discrimination function with the representation error by sparsity.The learned dictionary is benefit to classification for its within-class cohesion inter-class distinguishability.The experimental results show that with enough samples,the proposed approach outperforms the SRC algorithm by nearly 9 percent in classificatin accuracy.There are not enough samples for classification dictionary learning while the labeled samples are limited.Thus,a sparsity-based classification method based on specific-class dictionary and spatial characteristics is developed.In this method,some specific-class professional dictionaries are generated by k-means clusting algorithm,and are merged to construct the sparse representation dictionary.An adaptive spatial constraint is used to obtain the sparse coefficient for a test pixel by using the information of its neighborhoods.Experiments results on hyperspectral images show that the proposed algorithm achieves good classification performance while using a small number of samples.3)A new classification approach using random forest with spatial cooperative constraints for hyperspectral image is proposed.Firstly,the spectral feature is extracted by minimum noise fraction,which can reduce influence of noise,and the spatial feature is extracted by the extended morphological profiles of the image.Then random forest is constructed based on the extracted spatial and spectral features respectively.Finally,the label constraints based on space continuity is used to constraint the results by using the label information of its neighborhoods on image space.The classification result is decided by voting strategy.Experimental results demonstrate that the proposed approach is insensitive to noise and can obtain outstanding classification accuracy(more than 88%)even under strong additive noise environments(the signal noise ratio is 3.86dB).4)A novel tensor-based spectral-spatial feature extraction and classification method for hyperspectral image is proposed.In this method,a series of tensor operation are defined firstly,and then the principle component analysis is extended to its tensor variant(called TPCA)for extracting the spectral-spatial features.Furthermore,a certain pixel is represented as a tensor not a vector by itself and its neighbors,and the tensor feature is extracted by TPCA.Finally,the tensor feature is rearranged to its vector form for the traditional classifiers.Compared with the non-tensor methods and the recent tensor-based methods,the proposed method improves the classification accuracy significantly.