Research On Hyperspectral Image Classification Algorithm Based On Nearest Regularized Subspace

In recent decades,hyperspectral remote sensing has been rapidly developed as an important earth observation technology.Its spectral resolution is up to the nanometer level,which can provide rich image and spectral information.It has become increasingly popular in dealing with environmental monitoring and has shown good prospects,such as agricultural monitoring,forestry,ecosystem monitoring,and mineral identification.Therefore,the research of hyperspectral image classification still has important theoretical significance and practical application value.Hyperspectral image classification is based on the spectral and spatial features of image pixels,determining and labeling category attributes for different categories of features represented by each pixel.The main basis for distinguishing the feature categories is that different species have different spectral curves and geometric features.However,hyperspectral image has a high spectral dimensionality,which is easy to cause dimensionality disasters,and the cost of labeling is also high.In addition,hyperspectral image is often affected by noise and data loss.Traditional hyperspectral classification algorithms have certain defects in the face of the above problems.This thesis takes the nearest neighbor regularized subspace(NRS)as the basic framework,which is more active in recent years.Taking full advantage of the spectral and geometric properties of hyperspectral image,designing corresponding efficient algorithms by techniques of dimensionality reduction,cooperative training and image recovery.The main work and research results of the thesis are as follows:1.It is very difficult to label hyperspectral image.Most of the algorithms cannot obtain satisfactory classification results when the training samples are insufficient.To address the challenge,a hybrid collaborative training algorithm based on support vector machine(SVM)and NRS is proposed.To improve hyperspectral image classification accuracy,SVM classifier and NRS classifier as the base classifier of collaborative training,the training samples are augmented with the fake-labeled samples during the iteration,combining with the super pixel voting method to perform the hyperspectral image classification.The experimental results show that,the proposed algorithm achieves more superior performance in hyperspectral image classification.2.Hyperspectral image has high spectral dimensionality,and it is easy to cause over-fitting of the model using traditional classification algorithms.To address the challenge,a hyperspectral classification method for combining NRS and rotational invariant linear discriminant analysis(RILDA)is proposed.Firstly,the spectral dimensionality of hyperspectral image is reduced by RILDA.Then the processed data is classified by NRS.Finally,combining with the super pixel voting method to perform the hyperspectral image classification.The results show that,the overall accuracy and the required time of classification have been significantly improved.3.Hyperspectral image is typically corrupted by noise,occlusion or data loss.Obtaining a good performance for most regression-based methods is difficult.To address the challenge,we present a novel robust regression-based nearest regularized subspace(R~2NRS)for hyperspectral image classification.In our method,each band of a pixel is assigned with a regularized regression coefficient in the NRS model to reduce the influence of those bands corrupted during classification,combining with the super pixel voting method to perform the hyperspectral image classification.The results show that,the superior performance of our method for hyperspectral image classification for the case when some bands of the image are corrupted by noise or data loss.