Research On Feature Transformation And Classification Algorithms Of Hyperspectral Imagery Based On Superpixel Segmentation

Hyperspectral imagery(HSI)is a three-dimensional imagery containing tens or even hundreds of consecutive spectral bands acquired by an imaging spectrometer.This three-dimensional imagery has wealth information both in spatial and spectral,and has been widely used in many fields such as surface classification,target detection,environmental management,agricultural monitoring and military target identification.Therefore,it has played an important role in China's national economy and national defense construction.However,classification,as an important technology for hyperspectral imagery processing and application,still faces many problems and challenges.On the one hand,the high nonlinearity presented by hyperspectral imagery data makes it difficult to directly classify its raw data to achieve considerable results.On the other hand,extremely limited labeled samples make it difficult to estimate the accuracy of the parameters of the classification model.In order to tackle these problems with much better solution,based on the characteristics of hyperspectral imagery and the analysis of existing researches,this paper focuses on how to use superpixels to achieve effective small-sample classification of surface materials from two aspects.Firstly,in order to extract features containing much richer spectral-spatial information from the hyperspectral imagery and to make full use of these features,a multi-task learning framework based on support vector machine(SVM)is proposed.On the basis of the original Schrodinger feature mapping,this framework replaces raw features with multiple sets of Gabor features and adopts a feature-based dimensionality reduction method based on the superpixel level,which fully exploits the spatial-spectral information of hyperspectral imagery and reduces the time complexity.After obtaining multiple sets of spectral-spatial features,in order to make better use of these features for classification,a multi-task learning method based on support vector machine is adopted for the first time in this paper.Compared with traditional multi-task learning methods based on sparse representation,the method proposed in this paper reduces the time complexity significantly.Experimental results show that the multi-task learning framework proposed in this paper can extract much better features and can obtain much higher classification accuracy than some current advanced feature extraction methods.For another,in the process of decision fusion of hyperspectral imagery classification results,in order to obtain a classification result that is more consistent with the ground truth,this paper proposes a novel method based on local binary pattern(LBP)with superpixel guidance.First and foremost,the core part of this method is to be able to generate the superpixels which are consistent with the distribution of the ground truth.This can be achieved by over segmentation and region merging,which are proposed in this paper.Then,the uniform local binary pattern(ULBP)is used to extract the texture features of hyperspectral imagery and each pixel can be pre-predicted to the probability of each class by using support vector machine.Finally,a probability-oriented classification strategy is applied to classify each pixel based on superpixel-level guidance.Experimental results have demonstrated that the proposed method is more effective and powerful than several state-of-the-art methods in the small-sample situation,especially for hyperspectral images with more evenly distributed surfaces.