3 IGARSS2011_Pasolli_Final.pptx
- 1. A Novel Hybrid Approachto the Estimation of Biophysical Parametersfrom Remotely Sensed DataLuca Pasolli1,2Lorenzo Bruzzone1Claudia Notarnicola2E-mail:[email protected]@eurac.eduWeb page: http://rslab.disi.unitn.ithttp://www.eurac.edu
- 2. 2OutlineIntroduction and Motivation1Aim of the Work2Proposed Hybrid Estimation Approach3Experimental Analysis4Discussion and Conclusion5IEEE International Geoscience and Remote Sensing Symposium IGARSS 2011 Vancouver, Canada – 24-29 July, 2011
- 3. 3Introduction and MotivationInvestigatedTopic: EstimationofBiophysicalParametersfromRemotelySensed DataESTIMATIONSYSTEMTarget Biophysical Parameter EstimatesRemotely Sensed DataPrior InformationIMPORTANCE:Efficient and effective way forspatially and temporallymappingbiophysicalparameters at local, regional and global scale
- 6. Climatechange and environmentakriskassessmentCHALLENGES:Complexity and non-linearityof the relationship (mapping) betweenremotelysensed data and output target parameter
- 8. FieldreferencesamplesIEEE International Geoscience and Remote Sensing Symposium IGARSS 2011 Vancouver, Canada – 24-29 July, 2011
- 9. 4Introduction and MotivationThe EstimationProblemimplies the Definitionof a MappingFunction:InputRemotely Sensed VariablesContinuous Target Biophysical VariableMapping FunctionTheoretical Forward ModelInversionEmpirical ModelDevelopmentTheoretical Forward ModelInversion TechniqueReference SamplesRegression TechniqueIterative Methods
- 10. Look Up Tables
- 11. Machine LearningModelization of the Physical ProblemParametric / Non-ParametricRegressionStrength: Good robustness and generalization abilitysolid physical foundation
- 12. ideally no reference samples requiredWeakness:Limited accuracy in specificdomainssimplifications due to analytical modelization
- 13. no modelization of specific application issuesStrength:Good accuracy in specificdomainsideally no analytical simplifications
- 14. implicit modelization of specific application issuesWeakness:Limitedrobustness and generalization abilitywell representative reference samples required
- 15. site and sensor dependencyIEEE International Geoscience and Remote Sensing Symposium IGARSS 2011 Vancouver, Canada – 24-29 July, 2011
- 16. 5Aimof the WorkToDevelop a Novel Hybrid Approachto the Estimation of BiophysicalVariablesfrom Remote Sensing DataThe proposedapproachaims at improvingboth the accuracy and the robustnessof the estimates
- 17. isbased on the integrationoftheoreticalforwardmodel and available (few) referencesampesREFERENCESAMPLESTHEORETICALFORWARD MODELAccuracyin specificdomainsHYBRIDESTIMATION APPROACHRobustness and Generalization AbilityIEEE International Geoscience and Remote Sensing Symposium IGARSS 2011 Vancouver, Canada – 24-29 July, 2011
- 18. 6Proposed Approach: Problem FormulationGeneral Estimation ProblemContinuous Target Biophysical VariableInputRemotely Sensed VariablesTHEORETICAL FORWARD MODEL+INVERSION TECHNIQUEDesired Mapping FunctionDeviation FunctionREFERENCESAMPLESHybridEstimationFunctionIEEE International Geoscience and Remote Sensing Symposium IGARSS 2011 Vancouver, Canada – 24-29 July, 2011
- 19. 7Proposed Approach: ProblemFormulationExample:EstimationProblemwithtwo Input Variables(x1,x2)Goal: To associate a target parameter estimate ŷ to each position of the input spaceTheoreticalForwardModel+InversionTechniqueAvailable (few) ReferenceSamples2-dimensional input spaceIEEE International Geoscience and Remote Sensing Symposium IGARSS 2011 Vancouver, Canada – 24-29 July, 2011
- 20. 8Proposed Approach: Characterization of δ(.)Hypothesis:pointsclose in the input spacehavesimilarvaluesofδ(.)Idea:to exploit the deviationassociatedwith the availableReferenceSamples2-dimensional input spaceIEEE International Geoscience and Remote Sensing Symposium IGARSS 2011 Vancouver, Canada – 24-29 July, 2011
- 21. 8Proposed Approach: Characterization of δ(.)Hypothesis:pointsclose in the input spacehavesimilarvaluesofδ(.)Idea:to exploit the deviationassociatedwith the availableReferenceSamplesCase I: VeryFewReferenceSamplesGlobal DeviationBias (GDB) Strategyδ(.) isapproximatedwith a constantvalue in the whole input space2-dimensional input spaceIEEE International Geoscience and Remote Sensing Symposium IGARSS 2011 Vancouver, Canada – 24-29 July, 2011
- 22. 8Proposed Approach: Characterization of δ(.)Hypothesis:pointsclose in the input spacehavesimilarvaluesofδ(.)Idea:to exploit the deviationassociatedwith the availableReferenceSamplesCase II: More ReferenceSamplesLocalDeviationBias (LDB) Strategyδ(.) isassumedvariablewithin the input spacebutlocallyconstantFordefiningN(x):Fixedlocalneighborhood2-dimensional input spaceFixed quantization of the input space according to and IEEE International Geoscience and Remote Sensing Symposium IGARSS 2011 Vancouver, Canada – 24-29 July, 2011
- 23. 8Proposed Approach: Characterization of δ(.)Hypothesis:pointsclose in the input spacehavesimilarvaluesofδ(.)Idea:to exploit the deviationassociatedwith the availableReferenceSamplesCase II: More ReferenceSamplesLocalDeviationBias (LDB) Strategyδ(.) isassumedvariablewithin the input spacebutlocallyconstantFordefiningN(x):Fixedlocalneighborhood2-dimensional input spaceAdaptivelocalneighborhoodK-Nearest Neighborhood according to IEEE International Geoscience and Remote Sensing Symposium IGARSS 2011 Vancouver, Canada – 24-29 July, 2011
- 24. 9Proposed Approach: ImplementationTraining PhaseREFERENCESAMPLESCharacterizationofδ(.)Operational Estimation Phase+IEEE International Geoscience and Remote Sensing Symposium IGARSS 2011 Vancouver, Canada – 24-29 July, 2011
- 25. Experimental Analysis: Context and DatasetApplication Domain: SoilMoistureEstimationfromMicrowaveRemotelySensed DataChallenging and complexestimationproblem
- 26. High spatial and temporalvariabilityof the target parameter
- 27. Sensitivityof the microwavesignaltomanydifferent target properties
- 28. LimitedavailabilityofreferencesamplesStudy Area:bare agriculturalfieldsnear Matera, ItalyMedium/dry soilmoistureconditions
- 29. High variabilityofroughnessconditions due toplowingpracticeDataset:17 referencesamplesBackscatteringmeasurementswith a fieldscatterometer
- 30. C-Band (5.3 GHz)
- 31. Dual-polarization (HH and VV)
- 32. Multi-angle (23° - 40°)
- 34. Soilmoisture/dielectricconstant (5 < ε< 15)
- 35. Soilroughness (1.3 < σ< 2.5 cm)IEEE International Geoscience and Remote Sensing Symposium IGARSS 2011 Vancouver, Canada – 24-29 July, 201110
- 36. 11Experimental Analysis: SetupEstimationof the SoilMoistureContentperformedaccordingtoTheoreticalForwardModelInversionIntegralEquationModel (IEM)
- 37. Inversionperfomedbymeansof the SupportVectorRegressiontechniquewithGaussian RBF kernelfunctionaccordingto [1]Correctionof the deviationtermaccordingto the proposedapproach in two operative scenarios:Experiment 1: VeryfewreferencesamplesavailableGlobal DeviationBias(GDB) strategyExperiment 2: More referencesamplesavailableLocalDeviationBias(LDB) strategywithfixellocalneighborhoodEstimation Performance AssessmentComparisonwiththeoreticalForwardModelinversionwithoutdeviationtermcorrection
- 42. Slope and Interceptof the lineartendencylinebetweenestimated and measured target values[1]L. Pasolli, C. Notarnicola and L. Bruzzone, “EstimatingSoilMoisturewith the SupportVectorRegressionTechnique,” IEEE Geoscience and Remote SensingLetters, in pressIEEE International Geoscience and Remote Sensing Symposium IGARSS 2011 Vancouver, Canada – 24-29 July, 2011
- 43. 12Results: Experiment 1HP:VeryFewReferenceSamplesInfluenceof the # ofReferenceSamplesAvailableProposedHybridEstimationApproach(GDB Strategy)Standard TheoreticalForwardModelInversion2-dimensional Input SpaceIEEE International Geoscience and Remote Sensing Symposium IGARSS 2011 Vancouver, Canada – 24-29 July, 2011
- 44. ProposedHybridEstimationApproach(LDB Strategywithfixedlocalneighborhood)Standard TheoreticalForwardModelInversion13Results: Experiment 2HP:FewReferenceSamples2-dimensional Input SpaceIEEE International Geoscience and Remote Sensing Symposium IGARSS 2011 Vancouver, Canada – 24-29 July, 2011
- 45. 14DiscussionThe experimentalresultspresented are in agreement withthoseobtainedwithotherdatasets in different operative conditionsactive (scatterometer) and passive (radiometer) C-bandmicrowave data over bare areas
- 46. P-band SAR data overvegetatedareasThe potential and effectivenessof the methodisshownespeciallywhenchallenging operative conditions are addressedHigh level and variabilityofsoilroughness
- 47. PresenceofvegetationMore advanced and complexstrategies can bedefinedfor the characterizationof the deviationfunctionδ(.)MachineLearning (ML) methodsIEEE International Geoscience and Remote Sensing Symposium IGARSS 2011 Vancouver, Canada – 24-29 July, 2011
- 48. ConclusionA novelhybridapproachto the estimationofbiophysicalparametershasbeenpresentedItisbased on the inversionof a theoreticalforwardmodelforperforming the estimation
- 49. Itexploitsavailable (few) referenesamplestocorrectapproximationsintrinsic in the forwardmodelformulaitonThe proposedapproachispromising and effectivetoaddress the estimationofbiophysicalparametersfrom remote sensing dataItallowsonetoincrease the estimationaccuracy
- 50. Itiscapabletohandle the variabilityof the deviationδ(.)in the input space domain
- 51. Itisgeneral, simple, easytoimplement and fastduring the processingFuture ActivitiesDevelopmentofnoveladaptivestrategiesfor the characterizationofδ(.)
- 52. Investigationof the proposedappraoch in otherchallengingapplicationdomainsIEEE International Geoscience and Remote Sensing Symposium IGARSS 2011 Vancouver, Canada – 24-29 July, 201115
- 53. A specialThanktoDr. Claudia Notarnicola and Prof. Lorenzo BruzzoneThankyoufor the [email protected]@eurac.eduIEEE International Geoscience and Remote Sensing Symposium IGARSS 2011 Vancouver, Canada – 24-29 July, 2011
- 54. 16Results: Experiment P-Band SARStudy Area:VegetatedAgriculturalFields (SMEX O2 Experiment)Dataset: 35 referencesamplesAirborne SAR data (AirSAR)
- 56. Dual-polarization (HH and VV)
- 59. Soilmoisture/dielectricconstant (5 < ε< 16)
- 60. Soilroughness (1.3 < σ< 2.5 cm)Standard TheoreticalForwardModelInversionProposedHybridApproach (LDB)ProposedHybridApproach (GDB)IEEE International Geoscience and Remote Sensing Symposium IGARSS 2011 Vancouver, Canada – 24-29 July, 2011