Research Article Improving Hyperspectral Image ... Research Article Improving Hyperspectral Image

download Research Article Improving Hyperspectral Image ... Research Article Improving Hyperspectral Image

of 9

  • date post

  • Category


  • view

  • download


Embed Size (px)

Transcript of Research Article Improving Hyperspectral Image ... Research Article Improving Hyperspectral Image

  • Research ArticleImproving Hyperspectral Image ClassificationMethod for Fine Land Use Assessment Application UsingSemisupervised Machine Learning

    Chunyang Wang,1 Zengzhang Guo,1 Shuangting Wang,1 Liping Wang,2 and Chao Ma1

    1School of Surveying and Land Information Engineering, Henan Polytechnic University, Jiaozuo 454000, China2Sansom Institute for Health Research and School of Pharmacy and Medical Science, University of South Australia,Adelaide, SA 5001, Australia

    Correspondence should be addressed to Zengzhang Guo; and Liping Wang;

    Received 25 August 2014; Accepted 13 September 2014

    Academic Editor: Tifeng Jiao

    Copyright 2015 Chunyang Wang et al. This is an open access article distributed under the Creative Commons AttributionLicense, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properlycited.

    Study on land use/cover can reflect changing rules of population, economy, agricultural structure adjustment, policy, and traffic andprovide better service for the regional economic development and urban evolution. The study on fine land use/cover assessmentusing hyperspectral image classification is a focal growing area in many fields. Semisupervised learning method which takes alarge number of unlabeled samples and minority labeled samples, improving classification and predicting the accuracy effectively,has been a new research direction. In this paper, we proposed improving fine land use/cover assessment based on semisupervisedhyperspectral classification method. The test analysis of study area showed that the advantages of semisupervised classificationmethod could improve the high precision overall classification and objective assessment of land use/cover results.

    1. Introduction

    Remote sensing can quickly obtain surface information,achieve understanding, and study surface characteristics ofthe spatial distribution through transferring, processing, andanalyzing the data. The advantage of high spectral resolutionremote sensing is that it can obtain many continuous bandspectral images therefore, it achieves a fine description ofground targets and reaches the purpose of identifying fea-tures, especially suitable for plant fine classification comparedwith the conventional remote sensing methods [15].

    The land use information has great significance to landresource survey, planning, and dynamic monitoring. Theanalysis of land use/cover assessment using hyperspectralimage remote sensing has attracted more and more attentionin recent years [69]. Currently, supervised classification andunsupervised classification are the two traditional classifica-tionmethods for land use/cover.The supervised classificationis based on class probability density function for samplesin spatial feature. Generally, it takes higher classification

    accuracy and needs a lot of correct training samples. Palreported the usage of the extreme learning machine (ELM)algorithm for land use of hyperspectral image and achievedgood effect [10]. Bao et al. proposed that SVM and randomfeature selection (RFS) are applied to explore the potentialof a synergetic use of the two concepts in order to producehighly accurate [11]. Stankevich et al. proposed a new super-vised hyperspectral imagery classification using the imageryspectral bands as fuzzy data source attributes and cumulativemutual information resulting fuzzy classification as decisiontree inducing criterion [12].

    Unsupervised classification is a clustering method andcan detect unknown classes on images. The advantage ofunsupervised classification is simple and efficient.However, itcannot guarantee the real relationship between the clusteringfeatures classes and surface features classes [13, 14].

    Accessing training data for land cover classification usinghyperspectral data is time consuming and expensive especial-ly for hard-to-reach areas. Semisupervised learning researchfor land use/cover classification using a small amount of

    Hindawi Publishing CorporationJournal of SpectroscopyVolume 2015, Article ID 969185, 8 pages

  • 2 Journal of Spectroscopy

    labeled samples based hyperspectral image becomes a newresearch hotspot. Rajan et al. proposed that an active learningis well suited for learning or adapting classifiers when thereis substantial change in the spectral signatures betweenlabeled and unlabeled data [15]. Jun and Ghosh proposed asemisupervised learning algorithm called Gaussian processexpectation-maximization (GP-EM) for classification of landcover based on hyperspectral data analysis [16]. Munoz-Mariet al. proposed a semiautomatic procedure to generate landcover maps from remote sensing images [17]. Jun and Ghoshproposed a semisupervised spatially adaptive mixture model(SESSAMM) to identify land covers from hyperspectralimages in the presence of previously unknown land-coverclasses and spatial variation of spectral responses [18].

    So far, these research results have promoted the devel-opment of the land use/cover assessment, but there weresome problems with these study methods. On the one hand,the misjudgment probability of this strategy is relativelybig. On the other hand, these methods still cannot improveclassification accuracy significantly especially because thetypes classified are many.

    A new fine land use/cover assessment method wasproposed using hyperspectral image classification based oncombiningRenyi entropy andmultinomial logistic regressionsemisupervised learning model. Finally, the paper has landuse/cover assessment experiment by the real hyperspectralremote sensing image data. It shows that it can improve theaccuracy of classification and improve fine land use/coverassessment.

    2. Modelling of Fine LandUse/Cover Assessment

    2.1. Differences between the Spectral Characteristics. The fineland use/cover classification is able to use the hyperspectralimage data because the features of spectral characteristicsare the basis of recognition of feature attributes for remotesensing image.The descriptionmodel of hyperspectral imagedata is shown in Figure 1.The spectral curves of different treespecies are shown in Figure 2.

    Furthermore, the green artificial paints and hyperspec-tral remote sensing can identify vegetation that cannot bedistinguished by the human eyes. Liu et al. proposed thatthe outside laboratory spectrometer using ultraviolet, visible,near-infrared spectrometer can be used to measure all kindsof vegetation reflectance spectra such as Indus, Camphor,Broussonetia papyrifera, and Vine vegetable and comparedwith some reflection curve of green paint. The results showthat the spectral curves feature can distinguish betweenvegetation and green paint [19]. The differences betweengreen paint spectrum and vegetation spectrum are shown inFigure 3.

    2.2. The Method of Semisupervised Classification. The coreidea of this method of semisupervised classification is that,firstly, selected small amount of sample data are performedby multinomial logistic regression algorithm. The fittedregression coefficient can describe the direct relationshipbetween selected sample pixel and its category effectively.

    Then, hyperspectral image is classified by using the fittedregression coefficient. Secondly, the entropy of the experi-mental area is calculated through Renyi entropy calculationmethod that was proposed by Renyi in 1961 [20], and thensome unlabeled samples of maximum Renyi entropy areselected from the calculation data to be added to the sampledata. The classification of multinomial logistic regression isnot iterated repeatedly for many times until the classificationaccuracy tends to be stable.

    2.2.1. Problem Description. For hyperspectral classificationproblems, assuming that the hyperspectral remote sensingimage is = (

    1, . . . ,

    ) , represents the number of

    bands, each pixel represents a vector, there are observations, {1, . . . , }, represents category set, = {

    1, . . . ,


    shows that training set contains sample labels, , represents unlabeled sample set, = , and =(1, . . . ,

    ) represents classification categories of sample

    training datasets.

    2.2.2. The Principle of Multinomial Logistic Regression Algo-rithm. The multinomial logistic regression algorithm canpredict the fitted coefficient quickly and accurately [21, 22];the vector of parameter coefficient = (

    1, . . . ,

    ) is

    gotten by using multinomial logistic regression algorithmfor labeled training sample set. Equation (1) represents theprobability formula that probability of the event occurringthat multinomial logistic model is expressed as outcomevariables:

    (= |

    ) =

    exp ( ())

    1 +

    =1exp (


    . (1)

    When is equal to 1, . . . , , the () can be represented


    () = 0+ 11+ 22 + . (2)

    The coefficient of is estimated through the estimationcriteria of Bayesian maximum a posteriori and the log-likelihood algorithm [23, 24]; the solving method is given asfollows:

    = () =





    () log(


    exp ( ()))]




    Equation (3) calculates fitted coefficient of multinomiallogistic classification; the estimating type CLA(

    +1) brought

    by a new pixel +1

    is calculated as follows:

    CLA (+1) = argmax


    . (4)

    Each predicted category (CLA1, . . . ,CLA

    ) for hyper-

    spectral remote sensing image is calculated through (4); itrepresents a classification process is completed.

    The important issue studied is how to add new samples tothe training sample in the process of semisupervised classifi-cation.This process is conducted automatically by predicting

  • Journal of Spectrosc