遥感图像分形特征提取与分割

遥感图像分形特征提取与分割郑桂香,蔺启忠(,100101):分形理论由B.B.Mandelbrot于20世纪70年代中期创立,现已被广泛地应用于自然科学和社会科学的几乎所有领域本文在前人研究的基础上,利用双毯法(DoubleBlanketMethod)提取出图像的分形特征并用于图像分割,进一步证实了分形在此领域的可行性和有效性首先,通过比较局部分形维数偏移全局分形维数的标准差来确定适合该方法的最优滑动窗口其次,考虑到单尺度分维特征的局限性,提取出多尺度的特征值并建立分形维数谱然后,以模拟图像为例,分析图像中各区域的分维谱,选择适当尺度的分形特征,利用最大似然法对图像进行分割最后,将分形理论应用于遥感图像中,与传统的基于灰度值特征的图像分割方法相比,加入图像的空间分形纹理特征后分割精度明显提高研究结果表明:分维值的大小和变化趋势可以表示不同地物的空间复杂度,结合地物的光谱以及灰度信息能有效地识别目标地物:遥感;多尺度分形;双毯法;图像分割:TP751:A:1000-3177(2008)95-0009-07:2007-06-13:(40371085):(1983~),,,,Email:[e-mail],Pentland,[1],Peleg,,,[2~3],,[4][5]TM,[6];[7],,,,,,,,2Mandebrot:[8];,[9]HausdorffPeleg,,,r2rr,[2]:f(i,j),ur,br92008.1遥感信息u0(i,j)=b0(i,j)=f(i,j)(1):ur(i,j)=max{ur1(i,j)+1,maxd(i,j,m,n)1ur1(m,n)},r=1,2,3!(2)br(i,j)=max{br1(i,j)+1,maxd(i,j,m,n)1br1(m,n)},r=1,2,3!(3),d(i,j,m,n)(i,j)(m,n),:A(r)=i,j(ur(i,j)br(i,j))2r(4)A(r)∀k*r2D,,logA(r)=C1logr+C0(5),C1D=2C1A(k,l,r)=k+wi=kwl+wj=lw(ur(i,j,r)br(i,j,r))2r(6)(LFD)[3],(k,j),(2w+1)#(2w+1),(5)(6),(GFD),,,A(r)∀k*r2D,logrlogA(r),,,,[3,10],,,[11],,3,T.Parrinello[12],,(1(a))256#256,,0~255matlabrandn(0,1),:Iinner(x,y)=100cos(0.03x3)+127(b)(2,10)(a)1(b)213.1:,1(a)3,,21(b),3,[13],2032#32r=10,3#3,5#5,7#7,9#9(LFD)(GFD),2SELFDGFD,SE:0.039381,0.025816,0.020892,0.021352LFDGFD,,5#5,LFDGFD5#5,r=10LFD110遥感信息2008.12r=10,LFDGFD,3(a),(2,2.35),,(2.8,3.0),,,3(b),2:,,(2.8,3),,(a)1(b)23,y=140,,4(1),(2.8,3),,(2,2.35)(50,140),(200,140),,,1,,,2,3(b)4(b),(a)1(b)24(y=140):112008.1遥感信息3.2,2101~100,5:(1);(2),,100:∃,:,(2,10),(0,1),%11185(1)5(2),r=3,10,100,,,2(a)(b)523.3,,,,[11],,6(a)1,6(b)10,99.0404%,Kappa0.9807(a)(b)612,61.0956%,Kappa0.2156,,,1,98.0008%;,12Scale:OverallAccuracyKappaCoefficient376.9292%0.54261076.4494%0.534210076.8892%0.54303,1078.8085%0.583410,10096.2015%0.92343,10088.2047%0.76493,10,10098.0008%0.9597,,r=10r=100,(r10r100)/(r10+r100),r=100,r=10,r=100,12遥感信息2008.193.7625%,Kappa0.8745,(a)10(b)3,10,1007242001821ASTER123,512#512,4.1,,,8,:(1),,;(),,,(2),1230.5560m,0.6610m,0.8070mDN,FDband3FDband1FDband2,,VFDI(VegetationFractalDimensionIndex),FDband3,r=30ASTERr=30,FDband2,r=100ASTERr=100VFDI=FDband3,r=30FDband2,r=100(2),3DN(3)ASTER12315m,,,,8(4),,1,30~50,235~703,45~854.25#5,r=30,40,70,,:∃:82.6986%,Kappa0.7819;,84.7116%(r=70);,r=4085.8426%%,9,,9(a),,9(b),,,&9(d),,5,132008.1遥感信息,,85.8426%,3%,:,,;(),,,,,:,,1A.P.Pentland.Fractalbaseddescriptionofnaturalscenes[J].IEEETrans.,1984,PAMI-6,(6):661~674.2S.Peleg,J.NaoR,R.HartleyandD.Avnir.Multipleresolutiontextureanalysisandclassification[J],IEEETrans.,1984,PAMI(6):518~523.14遥感信息2008.13T.Peli.Multiscalefractaltheoryandobjectcharacterization[J].J.Opt.Soc.AmA,1990,7(6):1101~1112.4,.[J].,2000,12(1):53~57.5.[J].,1993,8(12):23~27.6,,.[J].,1998,2(1):47~50.7.[J].,1998,23(4):370~373.8B.B.Mandelbrot.,.∋[M].:,1999.9B.B.Mandelbrot.TheFractalGeometryofNature.Freeman[M].SanFrancisco,Cali,1982.10JayFeng.Fractionalfractalgeometryforimageprocessing[D].Illinois:NorthwesternUniversity,2000.11,.[M].:,2003.12T.ParrinelloandR.A.Vaughan.Multifractalanalysisandfeatureextractioninsatelliteimagery[J].INT.J.RemoteSensing,2002,23(9):1799~1825.13SonnyNovianto,etal.Nearoptimumestimationoflocalfractaldimensionforimagesegmentation[J].PatternRecognitionLetters,2003(24):365~374.FractalFeatureExtractionandSegmentationofRemoteSensingImageryZHENGGuixiang,LINQizhong(InstituteofRemoteSensingApplication,ChineseAcademyofSciences,Beijing100101,China)Abstract:FractalmethodisanewsubjectwhichwasfoundedbyAmericanscientistB.B.Mandelbrotinthemiddleof1970s,whichiswidelyappliedtoalmostallthefieldsofphysicalandsocialsciences.Basedonpreviousstudies,thispaperextractedthefractalfeaturesofimagesbyusingtheDoubleBlanketMethodandappliedthemtoimagesegmentationwhichshowedthevalidityandfeasibilityoffractalinthisfieldfurther.Firstly,anoptimumwindowwasselectedbycomparingthestandarderrorbetweenlocalandglobalfractaldimensions.Secondly,multiscalefractalfeatureswereextractedandfractaldimensionspectrumswereestablishedwithregardtosinglescale(slimitation.Then,byanalyzingobjectfractaldimensionspectrum,appropriatefeatureswereutilizedtosimulativeimagesegmentationbasedonthemaximumlikelihoodmethod.Atlast,fractaltheorywasbestowedtoremotesensingimage.Relativetothetraditionalmethodconsistedofonlygraylevelfeatures,theoverallsegmentationaccuracywasobviouslyimprovedwhenconsideredthespatialfractaltexturefeatures.Theresultsshowedthatfractaldimensionanditschangetrendcoulddisplayspatialcomplexityofdifferentobjects.Combinedwiththespectrumandgraylevelinformation,theobjectscanbediscriminatedeasily.Keywords:remotesensing;multiscalefractal;DoubleBlanketMethod;imagesegmentation(8)ANewMethodofAcquiringThreeComponentsofDeformationDisplacementBasedonDInSARTechniqueZHAXianjie,FURongshan,LIUBin,DAIZhiyang,SHAOZhigang,HANLibo(SchoolofEarth&SpatialScience,ChineseScienceTechnologyUniversity,Hefei230026)Abstract:AnewmethodofacquiringthreecomponentsofsurfacedeformationusingthreeSARinterferogramswithdifferentsatellite(slineofsightispresented.Toverifythismethod,interferogramsoftwotypeidealsurfacedeformationsaremodeledwiththr