Temporal sensorfusion for the classification of bioacoustic time series [Elektronische Ressource] / Christian R. Dietrich

Sujets

• Informatik
• Biologie

Informations

• Publié par: universitat_ulm
• Publié le: 01 janvier 2004
• Nombre de lectures: 10
• Langue: English
• Poids de l'ouvrage: 5 Mo

Extrait

AbteilungNeuroinformatik
Prof. Dr. Gunther¨ Palm
TemporalSensorfusionforthe
ClassificationofBioacousticTimeSeries
DissertationzurErlangungdesDoktorgradesDr.rer.nat.
derFakultat¨ fur¨ InformatikderUniversitat¨ Ulm
von
ChristianR.Dietrich
ausHirschegg
2003ii
AmtierenderDekan: Prof.Dr. FriedrichW.vonHenke
ErsterGutachter Prof.Dr. Gunther¨ Palm
Zweiter PDDr. AlfredStrey
TagderPromotion 25.06.04Abstract
Classifyingspeciesbytheirsoundsisafundamentalchallengeinthestudy
of animal vocalisations. Most of existing studies are based on manual
inspection and labelling of acoustic features, e.g. amplitude signals and
soundspectra,whichreliesontheagreementbetweenhumanexperts. But
during the last ten years, systems for the automated classification of ani
malvocalisationshavebeendeveloped.
InthisthesisasystemfortheclassificationofOrthopteraspeciesbytheir
sounds is described in great detail and the behaviour of this approach is
demonstratedonalargedatasetcontainingsoundsof53differentspecies.
Thesystemconsistsofmultipleclassifiers,sinceinpreviousstudiesithas
been shown, that for many applications the classification performance of
a single classifier system can be improved by combining the decisions of
multipleclassifiers.
To determinefeatures forthe individualclassifiers thesefeatures have
been selected manually and automatically. For the manuale selec
tion, pattern recognition and bioacoustics are considered as two coher-
entinterdisciplinaryresearchfields. Herebythesoundproductionmecha
nisms of the Orthoptera reveals significant features for the classification to
family and to species level. Nevertheless, we applied a wrapper feature
selection method, the sequential forward selection, in order to determine
furtherdiscriminativefeaturesetsfortheindividualclassifiers.
In particular, this thesis deals with classifier ensemble methods for
time series classification applied to bioacoustic data. Hereby a set of lo
cal features is extracted inside a sliding time window which moves over
the whole sound signal. The temporal combination of local features and
thecombinationoverthefeaturespaceisstudied.
Staticcombiningparadigmswheretheclassifieroutputsaresimplycom
bined through a fixed fusion mapping, and adaptive combining paradigms
where an additional fusion layer is trained through a second supervised
learning procedure are proposed and discussed. The decision template
(DT) fusion scheme is an intuitive approach for such a trainable fusioniv Abstract
scheme, which is typically applied to recognise static objects. During the
second supervised learning step the DT algorithm uses confusion matrix
data to adapt the fusion layer. In several empirical studies it has been
shownthattheclassificationperformanceofadaptivefusionschemes,par-
ticularlyforthesocalleddecisiontemplate,issuperior.
Many linear trainable fusion mappings, e.g. the combination
withthelinearassociativememory,thepseudoinversematrixandthenaive
Bayes fusion scheme are based on the same idea. Links between these
methods are given. However, regarding the classification of temporal se
quencesthesemethodsdonotconsiderthetemporalvariationoftheclas
sifieroutputs.
In order to deal with variations of classifier decisions within time se
riesweproposetocalculatemultipledecisiontemplates(MDTs)perclass.
Two new methods called temporal d t (TDTs) and clustered
decision templates (CDTs) are introduced and the behaviour of these new
methods is discussed on real data from the field of bioacoustics and arti
ficially generated data. In contrast to the combination with decision tem
plates the multiple decision template approaches lead to an increased ex
pressionpowerofthefusionlayer. Thisenhancestheclassificationperfor-
manceforclassificationproblemswheretheoutputsoftheindividualclas
sifiers show different characteristic patterns, which is a typical behaviour
intimeseriesapplications. Herebyseveralcharacteristicpat
ternsmayexistforthewholetimeseriesbelongingtothesameclassorfor
the individual local decisions of a single time series. Whereas the TDT
approach offers the advantage to learn several characteristic patterns for
wholetimeseries,theCDTapproachisalsoabletolearndifferentcharac
teristicpatternsovertime. Suchpatternshavebeenfoundinthetemporal
domainoftheOrthoptera.Acknowledgements
This thesis was supported by the help and the assistance of many people
whoseeffortsIherebywouldliketogratefullyacknowledge.
Firstofall,IwouldliketothankmyadviserProf. Dr. Gunther¨ Palmfor
hissteadysupport,forgivingmeindispensablemotivationandforgiving
metherightideasattherighttime. Hisvaluableadviceandguidancewill
always be appreciated. I am greateful to PD Dr. Alfred Strey for serving
as a reading committee member, who supported this work with valuable
directionsandameticulousreviewofthismanuscript.
Especially, I would like to thank to my thesis adviser Dr. Friedhelm
Schwenkerwhocontributedtothisthesiswithhisguidance,generousas
sistanceandmanyfruitfuldiscussions. Despiteofhisowntightschedule,
he always found a large amount of time for writing many papers in the
field of classifier fusion together with me, and for carefully reading this
manuscript.
ManythankstothemembersoftheDORSAprojectwhoimprovedthis
work in the field of bioacoustics, significantly. In particular, I would like
to thank PD Dr. Klaus Riede, Dr. Sigfrid Ingrisch, Dr. Klaus G. Heller
and Dr. Frank Nischk for providing their sound recordings, suggestions,
fruitfuldiscussionsandthesuccessfulcooperation.
Special thanks to all my present and former colleagues at the depart
ment of neural information processing for their friendliness and helpful
ness that makes working each single day a real pleasure for me. Valuable
discussionswithmycolleaguesimprovedthisworkconsiderably.
I also want to thank Barbara Previer for carefully reading this manu
AscriptandtoDr. AxelBauneforhissupportregardingLT X.E
Last, but most important I am grateful to my family for all their pa
tience, support and loving. My son Daniel who makes the nights shorter
but my days much brighter and my marvellous wife Dr. Katrin Dietrich
forherencouragementandnever endinglove.vi AcknowledgementsContents
Abstract iii
Acknowledgements v
ListofFigures xi
ListofTables xv
Abbreviation xvii
SomeCommonAcronyms xix
1 Introduction 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 RelatedWork . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 MajorContributionofthisThesis . . . . . . . . . . . . . . . . 8
1.4 OutlineoftheDissertation . . . . . . . . . . . . . . . . . . . . 9
2 ClassificationMethods 11
2.1 NeuralNetworks . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.1.1 RadialBasisFunctionNetworks . . . . . . . . . . . . 12
2.1.2 LearningVectorQuantisation . . . . . . . . . . . . . . 19
2.1.3 Fuzzy K NearestNeighbourClassifiers . . . . . . . . 23
2.1.4 SummaryandDiscussion . . . . . . . . . . . . . . . . 25
2.2 TimeAlignmentandPatternMatching . . . . . . . . . . . . . 26
2.2.1 Pre ProcessingofFeatures . . . . . . . . . . . . . . . . 26
2.2.2 DynamicTimeWarping . . . . . . . . . . . . . . . . . 28
2.2.3 TimeDelayNeuralNetworks . . . . . . . . . . . . . . 30
2.2.4 PartiallyRecurrentNeuralNetworks . . . . . . . . . 30
2.2.5 HiddenMarkovModels . . . . . . . . . . . . . . . . . 32
2.2.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . 32viii Contents
3 MultipleClassifierSystems 35
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.2 MCSDesignMethodologyandTerminology . . . . . . . . . 42
3.3 FeatureExtractioninTimeSeries . . . . . . . . . . . . . . . . 44
3.4 ThreeArchitecturesfortheClassificationofTimeSeries . . . 45
3.4.1 AverageFusion . . . . . . . . . . . . . . . . . . . . . . 49
3.4.2 ProbabilisticFusion. . . . . . . . . . . . . . . . . . . . 49
3.4.3 Percentiles . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.4.4 Voting . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.5 DecisionTemplatesforTimeSeriesClassification . . . . . . . 52
3.5.1 DecisionTemplates . . . . . . . . . . . . . . . . . . . . 52
3.5.2 MultipleDecisionTemplates . . . . . . . . . . . . . . 55
3.6 LinkstoNeuralNetworkTrainingSchemes . . . . . . . . . . 60
3.6.1 LinearAssociativeMemory . . . . . . . . . . . . . . . 60
3.6.2 DecisionTemplates . . . . . . . . . . . . . . . . . . . . 62
3.6.3 Pseudo InverseSolution . . . . . . . . . . . . . . . . . 63
3.6.4 NaiveBayesDecisionFusion . . . . . . . . . . . . . . 64
3.6.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.7 StatisticalEvaluation . . . . . . . . . . . . . . . . . . . . . . . 66
3.8 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4 OrthopteraBioacoustics 69
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.1.1 AutomatedSongAnalysis . . . . . . . . . . . . . . . . 71
4.1.2 SoundProductionandMorphologyoftheOrthoptera 72
4.1.3 AcousticStructureoftheOrthopteraSound . . . . . . 75
4.1.4 TheAcousticReceptorSystem . . . . . . . . . . . . . 78
4.1.5 AcousticCharacteristics . . . . . . . . . . . . . . . . . 80
4.1.6 NormalisationoftheTemperatureInfluence . . . . . 81
4.2 HierarchicalFeatureExtractioninBioacousticTimeSeries . 85
4.2.1 Pre processing