Research On Band Selection And Classification Of Hyperspectral Remote Sensing Images

For hyperspectral remote sensing images,it is a challenging task to select discriminative bands due to insufficient training samples and complex noise,and conflicts between insufficiently labeled samples and high-dimensional bands.The traditional deep neural network also faces the problem of overfitting in hyperspectral image classification.Aiming at the above problems,this paper proposes a band selection method and a classification method.The main work is as follows:(1)First,a band selection method combining hypergraph automatic learning and optimal clustering algorithm is proposed.By randomly dividing the entire band space into several subspaces(LVs)of different dimensions,where each subspace consists of a set of low-dimensional representations of training samples consisting of its associated bands.After projecting the samples of all subspaces into the label space,the hypergraph automatic model is used to preserve the local manifold structure of these projections,ensuring that the samples of the same class have a small distance,and the consistency matrix is used to integrate the bands corresponding to different subspaces.Finally,the ranking order is obtained through the clustering formula,and the optimal clustering result is finally selected.On three public datasets,the experiments verify that the proposed combination of hypergraph automatic learning and optimal clustering algorithm performs better than the comparison methods.(2)A multi-scale integrated deep hybrid kernel extreme learning machine based hyperspectral remote sensing image classification algorithm is proposed.First,an adaptive superpixel segmentation technique is used to perform multi-scale oversegmentation of the scene in the HSI dataset;secondly,the superpixel pattern(SP)and attention-adjacent superpixel pattern(ANSP).On this basis,an Integrated Deep Hybrid Kernel Extreme Learning Machine(EDHKELM)is proposed to study the deep features of SP and ANSP.Finally,the class of each pixel is precisely determined by a decision fusion and weighted output layer fusion strategy.Experiments on three different datasets verify that the classification results of the multi-scale deep hybrid kernel extreme learning machine network are better than other methods.