摘要
Background:Treespeciesrecognitionisthemainbottleneckinremotesensingbasedinventoriesaimingtoproduceaninputforspecies-specificgrowthandyieldmodels.Wehypothesizedthatastratificationofthetargetdataaccordingtothedominantspeciescouldimprovethesubsequentpredictionsofspecies-specificattributesinparticularinstudyareasstronglydominatedbycertainspecies.Methods:Wetestedthishypothesisandanoperationalpotentialtoimprovethepredictionsoftimbervolumes,stratifiedtoScotspine,Norwayspruceanddeciduoustrees,inaconiferforestdominatedbythepinespecies.Wederivedpredictorfeaturesfromairbornelaserscanning(ALS)dataandusedMostSimilarNeighbor(MSN)andSeeminglyUnrelatedRegression(SUR)asexamplesofnon-parametricandparametricpredictionmethods,respectively.Results:TherelationshipsbetweentheALSfeaturesandthevolumesoftheaforementionedspecieswereconsiderablydifferentdependingonthedominantspecies.Incorporatingtheobserveddominantspeciesinthepredictionsimprovedtherootmeansquarederrorsby13.3-16.4%and12.6-28.9%basedonMSNandSUR,respectively,dependingonthespecies.Predictingthedominantspeciesbasedonalineardiscriminantanalysishadanoverallaccuracyofonly76%atbest,whichdegradedtheaccuraciesofthepredictedvolumes.Consequently,thepredictionsthatdidnotconsiderthedominantspeciesweremoreaccuratethanthoserefinedwiththepredictedspecies.TheMSNmethodgaveslightlybetterresultsthanmodelsfittedwithSUR.Conclusions:Accordingtoourresults,incorporatinginformationonthedominantspecieshasaclearpotentialtoimprovethesubsequentpredictionsofspecies-specificforestattributes.DeterminingthedominantspeciesbasedsolelyonALSdataisdeemedchallenging,butimportantinparticularinareaswherethespeciescompositionisotherwiseseeminglyhomogeneousexceptbeingdominatedbycertainspecies.
出版日期
2016年02月12日（中国期刊网平台首次上网日期，不代表论文的发表时间）
