Hyperspectral Image Feature Extraction Via Spectral And Spatial Kernel Extrem Learn Machine

The classification of hyperspectral images(HSIs)is one of the most important research topics in the field of HSIs processing.However,some challenges,such as data redundancies caused by a number of spectral bands,etc,existed in HSIs classification,thus cause a great challenge for HSIs classification.Hence,extracting efficient and effective features from HSIs become more and more important.Recently,kernel extreme learning machine(KELM)has been proposed and attracted many attentions due to its merits,such as the fast speed and generalization,etc.Spectral and Spatial KELM(SSKELM),as the extension of KELM,produced higher classification accuracies than ELM for HSIs classification due to the use of spatial information of HSIs.The SSKELM,however,has a drawback which lacks the ability for extracting efficient and effective features from corrupted HSIs.This paper,hence,proposed a framework for HSIs feature extraction based on SSKELM.The main contributions of this work can be summarized as follows:(1)In order to address the drawback existed in SSKELM which can't extract the efficient features from HSIs,this work proposed a classification framework which combines the SSKELM and Local Binary Pattern(LBP).The Principal Component Analysis(PCA)used for dimensionality reduction of HSIs followed by LBP for feature extraction,then the SSKELM algorithm classifies the data extracted by LBP.The experimental results show the good performance of the proposed framework in terms of classification accuracies compared with SSKELM and other state-of-the-art algorithms.(2)In addition,this paper also proposed a method which combines the Extended Multi-Attribute Profile(EMAP)and SSKELM in order to address the same drawback existed in SSKELM.Firstly,the PCA has been used for dimensionality reduction,then the EMAP extracted the effective features from HSIs followed by SSKELM for classification.The experimental results show the good performance of the proposed framework in terms of classification accuracies compared with SSKELM and other state-of-the-art algorithms,even at the situation with a small number of training samples.(3)In order to improve the classification accuracies furtherly,this work proposed a feature fusion framework based on EMAP and LBP.The SSKELM classify the combined HSIs data extracted from LBP and EMAP after the PCA has been imposed on HSIs for dimensionality reduction.Experimental results,even with a small number of training samples,show the good performance of the fusion framework compared with the single feature extracted by EMAP or LBP.