Research On Hyperspectral Image Classification Method Based On Extreme Learning Machine

The classification application of hyperspectral images plays a vital role in geological exploration,urban expansion,agricultural and forestry monitoring,military and other industries.Hyperspectral images have excellent spectral information and rich spatial information,and its feature quality is one of the key factors affecting classification performance.Due to the intra-class differences of features and extensive changes in lighting and scale,the classification problem is still challenging.Therefore,how to extract essential features from hyperspectral data is the main research focus of this article.The main work is as follows:(1)It is of great difficulty for hyperspectral images in feature extraction due to their high spectral dimension,strong correlation,and large amount of data.In allusion to the issue that the classic extreme learning machine algorithm is difficult to extract spectral features well,this thesis introduces feature learning technology and comes up with a composite kernel extreme learning machine method based on discriminative information(CKELM-L).CKELM-L makes full use of Gaussian distribution information,by maximizing the inter-class matrix and minimizing the intra-class matrix,the projected low-dimensional data can be closer to the same category and away from different categories to extract spectral features.Considering the correlation between spectral pixels and spatial pixels,the kernel method is used to obtain spatial features.Multi-core learning is adopted for feature fusion operation,and then classified.It can be seen from the classified data that the proposed method retains better spectral features,has low computational complexity and achieves excellent separability.(2)Hyperspectral images have spectral features and spatial features,and the features also have spatial context information.Extreme learning machine is characterized by less training parameters,fast training speed and good generalization ability.In data clustering or classification tasks,it has been applied to obtain feature representations from complex data.Based on this,this thesis proposes a compound kernel extreme learning machine(GCKELM)with graph embedding structure to capture the input data characteristics.Specifically,the hyperspectral data contains spatial features and spectral features.In GCKELM,the spectral features and spatial features extracted by the graph embedding framework containing the intrinsic feature graph and the penalty graph are integrated into the extreme learning machine by the composite kernel method.Therefore,the both local structure and global structure information can be used with extreme learning machine.Experimental results on several standard data sets show that GCKELM can obtain effective and powerful feature representation of original data.(3)Hyperspectral images can provide the spectral characteristics and spatial structure of different substances.In the spectral features,because the correlation coefficient can effectively measure the spectral similarity between different pixels,this thesis proposes a kernel extreme learning machine classification algorithm(CCGCKELM),which combines the correlation coefficient and graph embedding.In this algorithm,the correlation coefficient between each band and its adjacent bands is used as the weight of the kernel method,and different bands are weighted according to the useful information contained in the kernel function,so that the bands with more useful information play a more important role in classification.The experimental results show that the correlation coefficient can affect the feature extraction of hyperspectral images.