資源描述：
《基于bp神經(jīng)網(wǎng)絡(luò)的高分辨率遙感影像分類》由會員上傳分享，免費(fèi)在線閱讀，更多相關(guān)內(nèi)容在教育資源-天天文庫。
1、摘要本文以2014年9月獲取的北京市部分地區(qū)的高分二號Pms影像作為實(shí)驗(yàn)數(shù)據(jù),采用神經(jīng)網(wǎng)絡(luò)方法對研究區(qū)域進(jìn)行分類處理及精度評定等研究。利用MATLAB軟件進(jìn)行神經(jīng)網(wǎng)絡(luò)程序開發(fā)，完成對影響進(jìn)行分類，并使用誤差矩陣和Kappa系數(shù)對分類結(jié)果進(jìn)行評定。結(jié)果顯示BP神經(jīng)網(wǎng)絡(luò)分類算法總體精度為81.6%，Kappa系數(shù)為0.765，比最大似然發(fā)監(jiān)督分類總體精度提高了6.5%。主要完成了如下的任務(wù)和相應(yīng)的結(jié)論：1針對高分辨率影像特征，通過地物多樣性和紋理分析，選取具有代表性樣區(qū)，確定分類類別數(shù)，并選取多光譜4個波段信息和紋理特征（通過計(jì)算全色波段灰度共生矩陣選取對比度紋
2、理圖像）作為分類特征數(shù)據(jù)。2選取訓(xùn)練樣本，設(shè)定BP網(wǎng)絡(luò)結(jié)構(gòu)。包括訓(xùn)練樣本歸一化處理，中間層神經(jīng)元數(shù)的設(shè)定和學(xué)習(xí)率的確定。利用MATLAB軟件完成BP神經(jīng)網(wǎng)絡(luò)的設(shè)計(jì)、分類后處理和精度評定。3將BP神經(jīng)網(wǎng)絡(luò)法和最大似然法監(jiān)督分類結(jié)果進(jìn)行對比。該網(wǎng)絡(luò)實(shí)現(xiàn)總體分類精度為81.6%，Kappa系數(shù)為0.7656。最大似然法監(jiān)督分類總體分類精度為75.1%，Kappa系數(shù)0.6886。關(guān)鍵詞：BP神經(jīng)網(wǎng)絡(luò)、高分辨率影像分類、監(jiān)督分類、精度評定5959AbstractThispapertaketheGF2highresolutionimageofBeijingcityas
3、basedexperimentaldata,anduseMATLABsoftwaretodeveloptheBPnetworkprogram.Weusethisprogramtoclassifytheremotesensingimages,thenweuseerrormatrixtoevaluatetheaccuracyofclassificationresult.Comparingdifferentclassificationmethods,wefoundthattheoverallclassificationresultofBPneuralnetwork
4、classificationis81.6%,whichismorethanmaximumlikelihoodmethod6.5%.Theinvestigationperformedandrelevantconclusionsareoutlinedasfollows:BasedonthecharacteristicsofGF2image,thispresentsdataprocessingandimageanalyzingofstudyarea,selectingthetypicalsampleareasandeterminingtheamountofcate
5、gory.Andthen,thefeaturedatatobeclassifiedarepreparedbycombiningthemulti-spectralbandswiththecontrast(CON)textureimagefrompanchromaticimage.ThetrainingsampleareaselectedfromthedataofthedataofstudyareaandtheBPnetworkstructureisconstructed.Toensurethestabilityanimprovetheconvergencera
6、te,someapproachesareexploredtoimprovetheBPalgorithm,suchasnormalizedpretreatmentoftrainingsamples,settingpropertrainingrateandadjustingnetworkstructureincludingthenumberofneuroninhiddenlayer.ThecomputerprogramofBPalgorithmisdesignedanddevelopedintheenvironmentofMATLABtoimplementthe
7、GF2imageclassificationandaccuracyevaluatingofstudyarea.Theexperimentsshowthatthismethodcanattainhigh-accuracy5959classificationresult.Comparedwiththoselikemaximumlikelihoodclassificationapproach,theBPhashigheroverallqualityandKappacoefficientthantheMLCwithimprovedby6.5%whichis75.1%
8、.AndtheBPKappais0.7656high
