An Optical Spectrum Combined Hyperspectral Image Classification Model And Method Based On Low Rank Representation

Hyperspectral remote sensing is one of the important earth observation technologies which have been developed rapidly in recent 30 years.It has the nanometer level spectral resolution,and can provide rich image and spectral information.Hyperspectral classification is a hot research topics in the applications of hyperspectral remote sensing.Most current classification methods mainly take advantage of spatial information,spectral information and radiate information to improve the classification accuracy.But because of its own characteristics:easily affected by the weather,high dimension data,large scale,high spectral resolution,low spatial resolution,the accuracy and stability of the hyperspectral classification are very difficult to be guaranteed.In this paper,using Low Rank Representation(LRR)as the basic framework,we fully extract spectral features and spatial correlation of hyperspectral images,take advantage of kernel tricks and multiple kernel learning to exploit the abundant spatial and spectral information and design a fast algorithm based on the alternating direction iterative algorithm(ADMM).The validity of the method and the software is verified by the experiment of the actual hyperspectral image.1.we propose a method called kernel low rank representation(KLRR).In this method,the low rank can acquire global correlation constraints on the hyperspectral image;At the same time,we map the traditional linear classifier to a nonlinear classifier using kernel function,which can obtain the characteristic of the high order structure of the data.Also,we use kernel tricks to avoid high dimensional operation.By analysing the four groups of kernels,the classical Gauss kernel,composite kernel,the mean filter kernel and the neighborhood filter kernel,we give the hyperspectral classification algorithm called the low rank representation with the neighborhood filter kernel.The experimental results show that this method is of high accuracy,especially in the case of small training samples.2.we propose a low rank representation of hyperspectral classification based on spectral-spatial MKL.This method aims at the problem that the single feature scale is limited,and the parameters are difficult to be determined in KLRR.In our method,we utilize MKL method in spectral and spatial dimensions,hence not only the spectral feature is extracted from the spectral dimension,but also the local neighborhood information of the image is extracted by the MKL.At the same time,it keeps to describe the global features of hyperspectral images by using the LRR framework.Besides,according to the characteristics of hyperspectral classification,the paper uses the RMKL to choose the kernel functions and the weights of the kernel matrix to avoid from the high time complexity of traditional multi kernel learning method.Experiments on real hyperspectral data show that this method is able to achieve robust,fast and high accuracy in hyperspectral image classification.3.Based on the above algorithms,the hyperspectral remote sensing image's classification processing software is developed based on MATLAB/GUI.It consists of four modules,including document management,hyperspectral image classification,accuracy assessment,analysis and so on.We also give the system framework,the main process design,and the software core module development and testing.