Research On Hyperspectral Image Classification Method Of Graph-based Semi-supervised Learning With Weighted Features

Hyperspectral images have a large amount of ground object information,and related image classification techniques have been applied in various fields.However,there are also various problems in the classification process of hyperspectral images,the most prominent of which is the high cost of manually labeling ground objects,resulting in fewer labeled samples,which affects the accuracy of classification.Therefore,under the condition of small samples,improving the accuracy of object classification has become an important research direction in the field of hyperspectral image classification.This paper takes deep autoencoder and sparse regularization as the core technology,integrates the idea of graph learning,and proposes a hyperspectral sensing image classification method of graph-based semi-supervised learning with weighted features,which is used to improve the accuracy of image classification under the premise of small samples.The specific research content and innovative ideas are as follows:Firstly,a deep autoencoder network is used for feature extraction to solve the problem of information redundancy in hyperspectral remote sensing images.Taking the deep autoencoder network as the main framework,the superpixel segmentation technology is used to extract some spatial information.And through the graph convolution operation,the spatial information and spectral information are fused to obtain the space-spectral joint feature.Compared with the traditional method of using the original data for composition,it can effectively solve the problem that the graph structure is sensitive to noise,and greatly improve the accuracy of composition.Secondly,the feature weighting scheme in the sparse graph regularization classification model is improved,and a similarity attenuation coefficient is introduced to reflect the contribution difference of adjacent pixels with different spatial distances to the central pixel.The similarity attenuation coefficient correction feature weighting scheme is introduced on the basis of Gaussian weighting.Experiments show that this method can describe the spatial similarity between pixels,so as to better integrate more spatial information.Finally,the deep self-encoder network is improved,and a shallow feature extraction mechanism based on convolution is proposed,which can extract more generalized spectral features,thereby solving the problem of lack of labeled samples.This method only extracts spectral features,which can not only extract spectral features with generalization,but also meet the requirements of shallow mechanism for less computation.Experiments on different scene datasets show that the method can achieve high classification accuracy under the premise of small samples.