GeologicalScienceandTechnologyInformation Vol.36 No.4 Jul doi: /j.cnki.dzkq ,,. [J].,2017,36(4): , 2 2, (1.

Size: px

Start display at page:

Download "GeologicalScienceandTechnologyInformation Vol.36 No.4 Jul doi: /j.cnki.dzkq ,,. [J].,2017,36(4): , 2 2, (1."

税黄
5 years ago
Views:

1 GeologicalScienceandTechnologyInformation Vol.36 No.4 Jul doi: /j.cnki.dzkq [J] (4): ( ; ) : Kappa 92% : ; ; :P627 :A : (2017) Hinton [8] 2010 Hinton [9] (DBN) 2014 Lü [10] DBN DBN [11] 2015 [12] [1] SVM 2015 Li [2-4] SVM [5] [13] 90% [6] 10a SVM DBN [7] DBN : : : (1975 ) @ qq.com

2 4 : 245 n m v iw ij h j (3) i=1 j=1 :θ={w ij a i b j }w ij RBM i j a i i b j j 1 (Restricted Boltzmann Ma- chine RBM) (4) : P(vh θ)= e-e(vh θ ) Z(θ) Z(θ)= e -E(vh θ) vh [14] 1.1 RBM v (5) : P(v θ)= 1 e Z(θ) -E(vh θ) (5) h RBM v RBM ( ) w ij a i b j RBM (6) RBM 1 1 RBM Fig.1 RBM model P(v i =1 hθ)=σ(a j + w ij h j ) (8) j 1.2 RBM (1) : θ={w P(h v)= n ij a i b j } p (h j=1 j v) (1) RBM Hinton [9] (2) : P(v h)= m p (v i=1 i h) (2) (contrastivedivergence CD) RBM RBM ( v ) n (1) X ( h ) m V (0) :v= (v 1 v 2...v n )h= (h 1 h 2...h n ) (9) : v i i h j j P(h (0) j =1 v (0) )=δ(w j v (0) ) (9) RBM (3) : (9) n m E(vh θ)=- a iv i - b j h j - i=1 j=1 :Z(θ) ; (4) (6) : P(h j =1 vθ)=σ(b j + i v iw ij ) (6) :σ(x) x<0 σ(x)=0; x> 0σ(x)=1 (7) : σ(x)= 1 (7) 1+e -x (8) : (2)

3 (10) : h (0) ~ P(h (0) v (0) ) (10) (3) h (0) (11) : P(v (1) i =1 h (0) )=δ(w ih T (0) ) (11) BP (4) v (1) ~P(v (1) h (0) ) DBN (5) BP ( ) (12) : P(h (1) i =1 v (1) )=δ(w iv (1) ) (12) (6) (13) : W W +γ(p(h (0) =1 v (0) )- h i RBM P(h (1) =1 v (1) )v (1) T (13) w ij RBM RBM RBM (pre-training); 3 2 (deep belief network DBN) RBM RBM 2.1 DBN RBM a. RMB(xh (1) );b. RBM(h (1) h (2) );c. 3 DBN Fig.3 TrainingprocessofDBN DBN 2 2 DBN Fig.2 StructureofDBN P(xh 1 h n )= P(x h n )( n-2 P (h k k=1 h k+1 ))P(h n-1 h n ) (14) 2.1 (1) RBM RBM(h (2) h (3) ) h (1) x RBM (xh (1) ) RBM w (1) a (1) b (1) h (1) h (1) h (1) h (2) RBM(xh (1) ) w (1) a (2) b (2) RBM RBM x (2) DBN n (14) :

4 : 247 3 3.1 4 40 07 ~40 24 112 52 ~113 31 53km 31km 1018km 2 3 5 DBN Fig.5 CurveofDBN recognitionaccuracyvaring withthenumberofiterations 92% DBN RBM DBN 6-a 3 ; 4 Fig.

4 4 : ~ ~ km 31km 1018km DBN Fig.5 CurveofDBN recognitionaccuracyvaring withthenumberofiterations 92% DBN RBM DBN 6-a 3 ; 4 Fig.4 Panchromaticimageofremotesensingdatainex- perimentarea 3.2 DBN DBN 6-b 64 DBN DBN 784 ; Softmax Logistic Logistic {01} Softmax DBN Logistic { k} Softmax 3.3 DBN : 15 batch DBN 100DBN DBN 7 DBN 5 15 DBN 1 ( ) ( ) ( DBN )3

248 2017 6 DBN (a) (b) Fig.6DBNrecognitionaccuracyvaringwiththenumberofunits(a)andnodes(b)inhiddenlayers(a)andthenumberofnodesinhiddenlayer(b) 7 DBN Fig.7 ResultofDBNclassification 92.23%Kappa 0.

[13] [J]. 201635(1):205-211. 15 batch 100 3 IEEE2015:685-694. 64 0.01 199. [1]. [J]. 201635(6):194- [2] LiXWangG.Optimalbandselectionforhyperspectraldata withimproveddiferentialevolution[j].

5 DBN (a) (b) Fig.6DBNrecognitionaccuracyvaringwiththenumberofunits(a)andnodes(b)inhiddenlayers(a)andthenumberofnodesinhiddenlayer(b) 7 DBN Fig.7 ResultofDBNclassification 92.23%Kappa DBN Table1 ConfusionmatrixofDBNclassification /% /% /% /% = 92.23% 4 Kappa = % : [D]. : [13] [J] (1): batch IEEE2015: [1]. [J] (6):194- [2] LiXWangG.Optimalbandselectionforhyperspectraldata withimproveddiferentialevolution[j].journalofambientin- teligence & HumanizedComputing20156(5): [3] LiXWangL.Onthestudyoffusiontechniquesforbadgeo- logicalremotesensingimage[j].journalofambientinteli- gence & HumanizedComputing20156(1): [4] LiXZhang H.Identificationofremotesensingimageofad- versegeologicalbodybasedonclassification[c] Anon.Bio- berg:springer2015: [5]. [J] (6): [6]. [J] (3): alinteligencemagazineieee20105(4): [9] HintonG E.ApracticalguidetotrainingrestrictedBoltzmann machines[j].momentum20109(1):926. inspiredcomputing:theoriesandapplications.berlin Heidel- [7] ArelIRoseDCKarnowskiTP.Deepmachinelearning:A newfrontierinartificialinteligenceresearch[j].computation- [8] HintonG EOsinderoSYw T.Afastlearningalgorithmfor deepbeliefnets[j].neuralcomputation200618(7):1527- [10]LüQDou YNiuXetal.Urbanlanduseandlandcover classificationusingremotelysensedsardatathroughdeepbe- liefnetworks[j].journalofsensors201520(15):1-10. [11]. [J]. 2014( 1): [12].. [14]Oquab MBotouLLaptevIetal.Isobjectlocalizationfor free -weakly-supervisedlearning with convolutionalneural networks[c] Computer Vision and Patern Recognition.

6 4 : 249 RecognitionandClassificationforRemoteSensingImage BasedonDepthBeliefNetwork XuLikun 1 LiuXiaodong 2 XiangXiaocui 2 (1.MapInstituteofHubeiProvinceWuhan430074China; (2.WuhanOpticsValeyBeiDouGeo-spatialInformationIndustryCo.Ltd.Wuhan430206China) tion. Keywords:remotesensingimage;depthbeliefnetwork;limitedBoltzmannmachine; classificationmethod 檪檪檪檪檪檪檪檪檪檪檪檪檪檪檪檪檪檪檪檪檪檪檪檪檪檪檪檪檪檪檪檪檪檪檪檪檪檪檪檪檪檪檪檪檪檪檪檪檪 ( 227 ) ApplicationofGeomagneticLoad-UnloadResponse RatioMethodinNorth-SouthSeismicBelt DaiMiao 12 FengZhisheng 3 LiuJian 1 LiDeqian 12 WeiGuichun 12 ShenXuelin 12 (1.KeyLaboratoryofEarthquakeGeodesyInstituteofSeismology ChinaEarthquakeAdministrationWuhan430071China; 2.EarthquakeAdministrationofHubeiProvinceWuhan430071China; 3.EarthquakeAdministrationofJiangsuProvinceNanjing210014China) Abstract:High-resolutionremotesensingimageshavehigh-dimensionalmulti-scalenon-stationaryinter- nalcharacteristicsandmassivemulti-sourceandheterogeneousexternalfeaturesandproviderichspatial information.thispaperexplorestheinteligentextractionandclassificationofhigh-resolutionremote sensingimagesusingtheemergingdepthbeliefnetwork.withalargenumberofexperimentalcomparisons ofboththeclassificationaccuracyandkappacoeficientandthefurtherparametersensitivityanalysisthe paperobtainstheoptimalsetingschemeforparameterssuchasthenumberofnetworklayersthenum- berofhiddenlayerneuronsandthenumberofiterations.comparedwiththetraditionalshalownetwork classifiertheimproveddepthofthebeliefnetworkcanbeterfittheintrinsicstructureofthesampleand theaccuracyoftheremotesensingimageclassificationisabout92%whichisveryefectiveinclassifica- Abstract:Electromagneticpropertieschangeindielectricconductivity duetostresschangesbeforethe earthquakemaycausethecorrespondingchangesinphaseandamplitudeofregionalverticalmagneticfield diurnalvariationcurveandhencegeomagneticload-unloadresponseratiosanefectivemethodforidentifi- cationofabnormalamplitude.thispapercalculatesgeomagneticload-unloadresponseratiosbefore18sig- nificantearthquakesinthenorth-southseismicbeltanalyzesthegeomagneticload-unloadresponseration spatialdistributioncharacteristicsandtheformation mechanism beforeearthquakes.theresultsindicate that:1intheexclusionofspaceweathersurroundingenvironmentinstrumentandmonitoringsystemand otherfactorsgeomagneticload-unloadresponseratioabnormalappearrbeforesomestrongearthquakesac- countingfor44.4%.2geomagneticload-unloadresponseratiothresholdisafectedbylatitudeandinthe conditionofthesamelongitudethegreaterthelatitudethelowerthethreshold.afterrepeateddebug- gingandcontrastthecalculationresultsshowthattheresponseratiothresholdofthenorth-southseismic beltinthenorthernsectionis3andthatofsouthandsouthwestregionsis3.1.3ingeneralabnormal within6 monthsaftertheoccurrenceofmoderateearthquakesthebiggerthenumberofabnormalsta- tionsthegreaterthemagnitudeofstrongearthquakes.theepicenterismorelikelytobelocatednearthe thresholdline. Keywords:geomagneticload-unloadresponseratio;dailyvariationamplitude;abnormityindex;earthquake prediction;north-southseismicbelt

2 209 Softmax Python TensorFlow drop- out 2 Fig.2 Remotesensingimageclassificationflowchart 1 1 ;2 ;3 ; ( ) 3 (3) 1 (4) Fig.1 Remotesensing

2 209 Softmax Python TensorFlow drop- out 2 Fig.2 Remotesensingimageclassificationflowchart 1 1 ;2 ;3 ; ( ) 3 (3) 1 (4) Fig.1 Remotesensing 37 2 2018 3 GeologicalScienceandTechnologyInformation Vol.37 No.2 Mar. 2018 doi10.19509/.cnki.dzkq.2018.0228. [J]. 201837(2)208-214. 1 2 3 (1. 710065;2. 430223; 3. ( ) 430074) Python TensorFlow Dropout