Structure-aware Hyperspectral Data Classification

With the diversification of data modes,hyperspectral data,as a unique type of data with the advantage of image-spectrum unification,has attracted extensive attention.This data can obtain the spatial dimension and spectral dimension information of the observed objects at the same time,and describe the observed objects in more multi-scale and multi-dimension.At present,the classification model for hyperspectral data is mostly used in the fields of city,farmland,environmental monitoring,etc.,but how to improve the accuracy of the classification model for the observation area classification interpretation is still a topic that needs to be deeply explored.Among the many classification methods,this paper takes representation-based and least-squares regression-based classification models as the main research contents,respectively explores the limitations of the above two methods in the task of hyperspectral data classification,and proposes their respective improved models.The main research contents of this paper are as follows:1.The uncertainty of hyperspectral data in the spectral seriously affects the actual effect of classification model,which attracted attention due to its simplicity and effectiveness of the classification model based on collaborative said,for not give full consideration to improve the spectral dimension difference of samples,and not the class of the specific use of the training sample label information to guide learning coefficient,a structure-aware collaborative representation was proposed,at the same time to consider using training samples of class label information and improve the sample spectra characteristics of uncertainty,in order to obtain more discriminative representation coefficients.In the proposed classification framework,marginal regression was employed.In addition,an inter-class row-sparsity structure was designed to maintain the compact relationship among intra-class pixels and more separable inter-class pixels,thus enhancing the separability of samples.The experiment and analysis of three open source hyperspectral data sets show that this method was superior to the existing representation-based classification model.2.The classification model based on least square regression can effectively deal with multiple classification tasks and is mainly used in the field of face recognition.When applied to the classification of hyperspectral data,this method often overfitting abnormal samples under the condition of spectral uncertainty of hyperspectral data,and it is easy to lose the discrimination information rich in hyperspectral data in the projection space.In order to solve the above problems,an effective method for pixel-wise hyperspectral classification was proposed,which was to learn more robust projection space by considering class separability and data reconstruction ability.The proposed model used the intra-class compactness graph to enhance class separability to avoid overfitting problems,and imposed data reconstruction constraints to retain discriminant information on a limited projection space.Experimental results based on synthetic manifold data and hyperspectral open source data show that the proposed method was good at processing manifold data and had advantages over other classification models based on least square regression and representation.