Hyperspectral Image Classification Based On Local Holding Broad Learning System

With the strong development of aerospace in various countries,hyperspectral image technology has received great attention.In order to better understand the surface of the Earth and other planets,scientists and engineers are constantly exploring new remote sensing technologies.Hyperspectral image technology,as an emerging remote sensing technology,has great application value.In recent years,hyperspectral image technology is developing rapidly,and hyperspectral images are appearing in more and more fields.At present,hyperspectral image technology has become one of the indispensable and important technical means in the fields of aerospace,military,ecological protection and so on.However,deep learning has a complex neural network structure,which requires frequent adjustment of the neural network settings in practical applications,while However,the complex structure of deep learning networks requires frequent adjustment of network parameters during experiments,while the large number of complex deep learning parameters leads to a large amount of time consumption in hyperspectral image classification tasks.In response to the above emerged problems,two classification methods are proposed in this paper,the main elements of which include:(1)BLS only focuses on the separability of various samples,ignoring the relative relationship between samples and the discriminative information,this is propose a broad learning network based on locally sensitive discrimination is proposed for hyperspectral image classification.The method considers the discriminative information of labeled samples and the local flow structure of data samples by introducing a locally sensitive discriminant analysis,and constructs intra-class and inter-class graphs by labeled samples to characterize the relative relationships between data samples.On this basis,the intra-class and inter-class graphs are introduced into the objective function of the width learning system to enhance the discriminative power of the method for data features by minimizing the intra-class graphs and maximizing the inter-class graphs so that the samples of the same class are clustered as much as possible and the samples of different classes are as far away as possible.In three commonly used hyperspectral datasets,it shows a good classification effect compared with other comparison methods.(2)For hyperspectral images with complex geometric structure and spatial layout,it is difficult to fully characterize hyperspectral data with linear sparse features in width learning,this paper propose a local discriminative graph convolution based broad learning network is proposed for hyperspectral image classification.The method can obtain rich nonlinear spatial spectral features in hyperspectral images by using PCA to downscale hyperspectral images,EMAP to extract multi-attribute profile features,and graph convolution operations to aggregate node information in adjacent graphs to learn the context,and then local intra-class and local inter-class graphs are introduced to reflect the local geometric structure and local discriminative information of the input space,respectively.(3)For the above two hyperspectral image classification methods,a hyperspectral image classification tool is integrated in this paper.The tool can classify images by inputting images,and the classification results and the classification result graph can be visualized by the tool.In practical use,it is only necessary to follow the tips of the tool,so it has good practical value.