Documents

2012_6.pdf

Description
Expert Systems with Applications 39 (2012) 3800–3809 Contents lists available at SciVerse ScienceDirect Expert Systems with Applications journal homepage: www.elsevier.com/locate/eswa Hyperion hyperspectral imagery analysis combined with machine learning classifiers for land us
Categories
Published
of 10
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Related Documents
Share
Transcript
  Hyperion hyperspectral imagery analysis combined with machine learningclassifiers for land use/cover mapping George P. Petropoulos a,b, ⇑ , Kostas Arvanitis b , Nick Sigrimis b a Foundation for Research and Technology – Hellas (FORTH), Institute of Applied and Computational Mathematics, Regional Analysis Division, N. Plastira 100, VassilikaVouton, GR-70013 Heraklion, Crete, Greece b Department of Natural Resources Development & Agricultural Engineering, Agricultural University of Athens, 71, Iera Odos St., 111 18 Athens, Greece a r t i c l e i n f o Keywords: Land cover/use mappingHyperionSupport Vector Machines (SVMs)Artificial Neural Networks (ANNs)ClassificationRemote sensingMediterraneanGreece a b s t r a c t Describing the pattern and the spatial distribution of land cover is traditionally based on remote sensingdata analysis and one of the most commonly techniques applied has been image classification. The mainobjective of the present study has been to evaluate the combined use of Hyperion hyperspectral imagerywith the Support Vector Machines (SVMs) and Artificial Neural Networks (ANNs) classifiers for discrim-inating land-cover classes in a typical Mediterranean setting. Accuracy assessment of the derived the-matic maps was based on the analysis of the classification confusion matrix statistics computed foreach classification map, using for consistency the same set of validation points.Results indicated a close classification accuracy between the two classifiers, with the SVMs somehowoutperforming the ANNs by 3.31% overall accuracy and by 0.038 kappa coefficient. Although both classi-fiers produced close results, SVMs generally appeared most useful in describing the spatial distributionand the cover density of each land cover category. The higher classification accuracy by SVMs was attrib-uted principally to the ability of this classifier to identify an optimal separating hyperplane for classes’separation which allows a low generalization error, thus producing the best possible classes’ separation.On the other, as a key disadvantage of both techniques was identified that both do not operate on a sub-pixel level, which can significantly reduce their accuracy due to possible mixture problems occurredwhen coarse spatial resolution remote sensing imagery is used.All in all, this study demonstrated that, provided that a Hyperion hyperspectral imagery can be madeavailable at regular time intervals over a given region, when combined with either SVMs or ANNs clas-sifiers, can potentially enable a wider approach in land use/cover mapping. This can be of particularimportance, especially for regions like in the Mediterranean basin, since it can be related to mappingand monitoring of land degradation and desertification phenomena which are evident in such areas.   2011 Elsevier Ltd. All rights reserved. 1. Introduction Remote sensing data has been an is an attractive source in thedetermination of land cover thematic mapping, providing valuableinformation  for   delineatingthe extent of landcover classes, as wellas for performing temporal land cover change analysis and riskanalysis at various scales (Kavzoglu & Colkesen, 2009). Suchinformation is also useful in policy decision making, such as whenconcerning environmentally or ecologically protected areas or na-tivehabitat mappingandrestoration(Council Directive92/43/EEC,1992; Fassnacht, Cohen, & Spies, 2006; Sanchez-Hernandez, Boyd,&Foody, 2007). Thematicmaps of landuse/cover arealso linkedtothe monitoring desertification and land degradation, key environ-mentalparameterspronouncedinareassuchastheMediterraneanbasin (Castillejo-González et al., 2009).Producinglanduse/covermappingthematicmapsusingremotesensingdataiscommonlyperformedbydigitalimageclassification(Chintan, Arora, & Pramod, 2004). A recent, comprehensive reviewof the variety of classification approaches applied to remote sens-ing data was made available by Lu and Weng (2007). Generally, awidely used categorization of classification techniques includesthree main groups of approaches, namely: pixel-based, sub-pixeland object-based classification techniques. Pixel-based techniquesperformclassification by assigning pixels to land cover classes andthis be achieved by either supervised or unsupervised classifiers.Unsupervised classifiers group pixels with similar spectral valuesinto unique clusters according to some statistically predefined 0957-4174/$ - see front matter    2011 Elsevier Ltd. All rights reserved.doi:10.1016/j.eswa.2011.09.083 ⇑ Corresponding author at: Foundation for Research and Technology – Hellas(FORTH), Institute of Applied and Computational Mathematics, Regional AnalysisDivision, N. Plastira 100, Vassilika Vouton, GR-70013 Heraklion, Crete, Greece. Tel.:+30 2810 391774; fax. +30 2810 391761. E-mail addresses:  petropoulos.george@gmail.com, gpetropoulos@iacm.forth.gr(G.P. Petropoulos).Expert Systems with Applications 39 (2012) 3800–3809 Contents lists available at SciVerse ScienceDirect Expert Systems with Applications journal homepage: www.elsevier.com/locate/eswa  criteria that the classifier combines and re-assigns the spectralclusters into information classes. On the other, supervised classifi-ersusesamplesof knownidentityforeachlandcoverclass, knownas ‘‘training sites’’, to classify image pixels of unknown identity(Campbell,1996).Supervisedclassifiersarealsocommonlydividedinto parametric and non-parametric. In comparison to non-para-metric(suchasArtificialNeuralNetworks–ANNs),parametricpix-el-based classifiers (e.g. the Maximum Likelihood – ML) requireprior knowledge/assumptions regarding the statistical distributionof the data to be classified for the different classes used, informa-tion often difficult to attainin practice. Spectral unmixing is a verydifferent classification approach, which is based in defining differ-ent surface material fractions within an image pixel. Sub-pixelclassification approaches are generally divided into linear andnon-linear unmixing, depending on whether it is assumed thatthe reflectance at eachpixel of the image is a linear or a non-linearrespectively combination of the reflectance of each material pres-ent within the pixel (Plaza, Plaza, Perez, & Martinez, 2005, 2009;Small, 2001). In object-based classification each classification taskaddresses a certain scale, and image information can be repre-sented in different scales based on the average size of image ob- jects, whereas the same imagery can be segmented into smalleror larger objects.Remote sensing data in land cover classification mappingstarted to be routinely applied from the late 1960’s. Since then, awiderangeof space-bornemultispectral systemshavebeenplacedinorbit.Therapidadvanceofairborneandsatellitesensortechnol-ogy has also resulted to the development of hyperspectral sensors.Thesesensorsareabletoprovidespectraldataforeachimagepixelin numerous narrow continuous spectral bands, unlike multispec-tral systems that produce images in a few relatively broad spectralbands. During the last decades, a number of airborne and satellitehyperspectral sensing systemssensors havebeenlaunched. Ontheother, the recent availability of such data has also encouraged thedevelopment of several techniques for analyzing the rich informa-tion content provided by hyperspectral imagery. Hyperion is a sa-tellite hyperspectral sensor onboard the Earth Observer-1 (EO-1)platform, launched under NASA’s New Millennium Program endof 2000. Hyperion acquires hyperspectral images at a 30m spatialresolution and at about 10nm spectral resolution in 242 spectralbandsintotal,70ofwhichfoundinthevisible/nearinfrared(VNIR)and172intheshort-waveinfrared(SWIR)(Han,Goodenough,Dyk,& Love, 2002). The availability of data from Hyperion has createdunique opportunities for remote sensing studies to be conductedexploringthesensorpotentialuseinlanduse/coverthematicmap-ping extraction.Various studies have examined the combined use of Hyperionwith different pixel-based classification techniques for land-use/cover mapping (Eckert & Kneubühler, 2002; Galvao, Formaggio, &Tisot, 2005; Galvao et al., 2009; Goodenough et al., 2003; Pignattiet al., 2009; Walsh et al., 2008). Furthermore, both linear andnon-linear unmixing classification combined with Hyperion forland classification has also been investigated in a few studies(e.g. Pignatti et al., 2009; Ramsey, Rangoonwala, Nelson, Ehrich,& Martella, 2005; Walsh et al., 2008). Others have also exploredthe use of object-based classification with Hyperion imagery anal-ysis for performing land use/cover mapping (Walsh et al., 2008;Wang, Chen, He, Lv, & Liu, 2010). Yet, studies performing compar-ative analysis of the performance of different classification ap-proaches with Hyperion hyperspectral imagery for land use/covermapping in particular, are evidently scarce in the literature (Du,Tan, & Xing, 2010; Pignatti et al., 2009; Walsh et al., 2008; Wanget al., 2010). Furthermore, to our knowledge, an investigation of the potential use of Hyperion hyperspectral imagery combinedwith very powerful machine learning, non-parametric classifierssuch as Support Vector Machines (SVMs, Vapnik, 1995) andArtificialNeuralNetworks(ANNs)forderivinglanduse/covermap-ping has so far been very limited, if not existent. This despite thefact that several studies applied previously employing either mul-tispectral(Canty, inpress;Dixon&Candade,2008;Huang,Davis, &Townshend, 2002; Nemmour & Chibani, 2006) or hyperspectraldata(Karimi,Prasher,Patel,&Kim,2006;Pal&Mather,2005),havedemonstrated their strong potential in thematic informationextraction for land use/cover mapping applications. Such an inves-tigation would also be undoubtedly of great interest if imple-mented in Mediterranean conditions, due to the high relevanceoflanduse/coverinformationextractiontodesertificationandlanddegradation, phenomena often pronounced in such regions(Castillejo-González et al., 2009).In this context, the objective of the present study hasbeen to appraise, to our knowledge for first time, the combineduse of the SVMs and ANNs machine learning non-parametricclassifiers combined with Hyperion hyperspectral imagery forland use/cover mapping in a typical Mediterranean setting. Forthis purpose, as a test site was selected an area located in themainland of Greece for which a Hyperion imagery was madeavailable. 2. Study site The study area, is located approximately 50km north of thecenter of Athens, andis coveringparts of AtticaandVoiotiaprefec-tures (Fig. 1). The surface area covered is approximately 200km 2 ,extending from 23  44 0 to 23  45 0 East, and from 38  5 0 to 38  20 0 North. The area is representative of typical Mediterranean condi-tions in terms of landscape structure and land surface cover varia-tion. The terrain varies from sea level to approximately 1400m,having steep slopes especially in the northern and southern partsof the area around a mountain lying there, Mt. Parnitha. The vege-tation of the area also varies due to the high altitude difference. Atlower elevations, is found mainly scherophyllous vegetation, spar-sely vegetated areas and some agricultural areas. At higher Fig. 1.  Location of our study area. G.P. Petropoulos et al./Expert Systems with Applications 39 (2012) 3800–3809  3801  altitudes, areas are covered mainly by conifers and broadleavedforests as well as transitional woodland/scrubland areas. The partof the studied region in which Mt. Parnitha lies is a National Park,which also due to its biodiversity richness has been included inthe European network of protected areas NATURA 2000(http://www.natura.org/). During the summer of 2007, a wildlandfire that occurred in this area resulted to a serious ecological dam-age to this National Park, destroying a large part of it. 3. Datasets The Hyperion imagery used in the present study was acquiredover out test site on August 27th, 2009, and was obtained at nocost from the United States Geological Survey (USGS) archive.The imagery was received as a full long scene (185-km strip) andat level 1 (L1GST) processing level in GeoTIFF format, written asband-interleaved-by-line (BIL) files stored in 16-bit signed integerradiancevalues. The L1Gproduct is radiometricallycorrected, geo-metrically resampled, and registered to a geographic map projec-tion with elevation correction applied.In addition, it was also collected an imagery fromthe AdvancedLand Imager (ALI) sensor which is also onboard the EO-1 platform.The ALI imagery was of the same overpass day as that of Hyperionimagery and was also obtained from USGS at the same processinglevel astheacquiredHyperionimagery.ALI sensoracquiresdatain10 spectral bands, one panchromatic with a spatial resolution of 10mand 9 other bands covering a wavelength range fromthe vis-ibletoshortwaveinfraredandaspatial resolutionof30m. Inaddi-tion, the CORINE2000 Land Cover (CLC) map ( JRC-EEA, 2005) at aspatial resolution of 100m for the test region was obtained at nocost (from http://reports.eea.europa.eu/COR0-landcover/en). 4. Methods Landuse/coverclassificationwasconductedinourstudysitebyimplementingtheSVMsandtheANNsnon-parametricpixel-basedclassifiers to the Hyperion imagery acquired for the study site.Fig. 2 presents an overview of the methodology followed for thispurpose. Next is provided an overview of the main pre-processingsteps as well as the classifiers implementation to the Hyperionimagery for deriving the thematic maps of land use/cover. 4.1. Data pre-processing  The main pre-processing applied to the collected Hyperionimageryis brieflydescribedinthe subsequent sections. First, usingtheLevel1G/1THDFandGeoTIFFHyperionimagerywasconvertedinto ENVI format files that contain wavelength, full width half maximum and bad band information. This was performed in ENVIimage processing environment (ITT Visual Information Solutions)using the Hyperion_tools.sav toolkit. Subsequently, were removedthe non-calibrated bands of the Hyperion imagery (namely bands1–7; 58–76; 77–78; 225–242). Hyperion VNIR spectrometer has70 bands of which only 50 are calibrated, while the SWIR spec-trometer has 172 bands of which only 148 are calibrated. The198 calibrated bands cover the entire spectrum from 426 to2395nm (USGS, 2008). Also the Hyperion imagery water absorp-tion bands (namely bands 120–132, 165–182, 185–187,221–224) were eliminated in order to reduce to the data the influ-enceofatmosphericscatter,watervaporabsorptioncausedbywellmixed gasses. Bands 77 and 78 were also eliminated because theyhad a low SNR value, and overlapped with band 56 and band 57respectively. In the next step, the Hyperion imagery bands withvertical stripping were indentified based on visual inspection andthose were manually removed (namely bands 8–9, 56–57, 79–82,97–99, 133–134, 152–153, 188, 213–216, 219–220). Verticalstripes are caused by differences in gain and offset of differentdetectors in push broom-based sensors such as Hyperion and ver-tical stripping are usually identified by visual inspection of the im-agedataoratmosphericmodeling(Beck,2003).Then,theat-sensorradiance was computed from the raw Digital Number (DN) values,for all remained spectral bands. This was derived by dividing thepixel’s DN by a constant value, which was 40 for the visible andnear-infrared (bands 8–57) and 80 for the short-wave infrared(bands 79–224) (USGS, 2008). Atmospheric correction was not ap-plied, as according to Datt, McVicar, Van Niel, Jupp, and Pearlman(2003) ‘‘it is not necessary to atmospherically correct image datafor a single observation’’ (in Pengra, Johnston, & Loveland, 2007).Also, taking into account that the Hyperion imagery was alreadyterrain-corrected, no further correction for topographic effectsdeemed necessary.Subsequently, a minimum noise fraction (MNF; Boardman &Kruse,1994;Lee,Woodyatt,&Berman,1990)wasappliedtoHype-rion data set in order to separate noise from data and to minimizethe influence of systematic sensor noise during image analysis, ashas been done previously by other investigators (Galvao et al.,2005; Pengra et al., 2007; Pignatti et al., 2009). The MNF transform identifiessystematicnoise, whichispresumedtoarisefromsensorand processing anomalies, and can then segregate it from themeaningful signal. MNF transformation was performed on all theHyperion bands that had not been masked out (136 in total) inENVI as a linear transformation and is essentially two principalcomponent transformations. The first transformation, is based onan estimated noise covariance matrix, decorrelates and rescalesthe noise in the data. The second step is a standard principal com-ponent transformation which creates several new bands contain-ing majority of the information (ENVI User’s Guide, 2005).Herein, the MNF transformation handled the VNIR and SWIR dataseparately because it was superior managing the noise due to itsdifferent structure in the two data sets (Datt et al., 2003). Theresulting MNF bands were analyzed for their spectral informationcontent using eigenvalue plots and individual MNF gray-scalebands. For the inverse MNF transformation nine components inthe VNIR and seven in the SWIR were used. Hyperion final dataset after the implementation of an inverse MNF consisted of 136bands, 46 in the VNIR and 90 in the SWIR. After this step, theresulting image was reduced to a subset of the studied region(Fig. 3). These final 136 bands after this last pre-processing stepwere used in the present study. 4.2. Hyperion classification Apixel-basedsupervisedclassificationusingboththeSVMsandtheANNsclassifierwascarriedoutontheHyperionimage,throughthree main steps. Firstly, the classification key was formulated,which consisted of the classes shown in Table 1. The decision touse this classification scheme was based primarily on photo-inter-pretation of the examination of the ALI panchromatic imagery(10m spatial resolution), assisted by an inspection of the classespresentattheCORINE2000landnomenclaturemap(100mspatialresolution)andourfamiliaritywiththestudyregionfrompreviousworks. Secondly, training sites representative of each of the aboveclasses were collected from the Hyperion imagery following astratified random sampling strategy. Selection of the training siteswasguidedbytheaerialimageryphotointerpretationandselectedfield visits conducted, supported also by the familiarity with thestudy area from previous works conducted in the same region.The training sites were carefully determined and restricted to re-gions where land-cover changes is consistent, and to land-coverwith slight phonological changes. A total of 716 Hyperion pixelswere identified as training data representing the classes defined 3802  G.P. Petropoulos et al./Expert Systems with Applications 39 (2012) 3800–3809  in the adopted classification scheme. Thirdly, the SVMs and ANNsclassifiers were developed and implemented in ENVI image pro-cessing environment, using the training sites collected in the pre-vious step. 4.2.1. SVMs classification This section provides the details concerning the SVMs imple-mentation to the Hyperion hyperspectral imagery for producing aland use/cover thematic map over our studied region. A detaileddescription of SVMs workings was considered unnecessary to beprovided herein, as that can be found elsewhere, for example inBurges (1998) and Foody and Mather (2004). SVMs is a supervised machine learning method that performsclassification based on statistical learning theory (Vapnik, 1995).It is a binary classification method that provides a separation of classesbyfittinganoptimalseparatinghyperplanetoasetoftrain-ingdatathatmaximizestheseparationbetweentheclasses.Essen-tially, the hyperplane is the decision surface on which the optimalclass separation takes place. Intuitively, a good separation isachieved by the hyperplane that has the largest distance to theneighboring data points of both classes. Each training example isrepresented by a feature vector. From a given set of training data,the SVMs classifier calculates an optimal hyperplane characterizedbyavectorthatprovidesthebestseparationbetweenthetwoclas-ses.Theoptimalhyperplaneistheonethatmaximizesthedistancebetween the hyperplane and the nearest positive and negativetraining example, called the margin. To avoid computational over-load, this is not done by evaluating all training points, but only asubset, called the ‘‘support vectors’’ of the algorithm. SVMs canbeextendedtomorethantwoclassesbysplittingtheprobleminto Fig. 2.  A flow diagram summarizing the processing steps adopted in the present research study. G.P. Petropoulos et al./Expert Systems with Applications 39 (2012) 3800–3809  3803
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks