Investigation of Noise and Dimensionality Reduction ...
Transcript of Investigation of Noise and Dimensionality Reduction ...
Investigation of Noise and Dimensionality Reduction Investigation of Noise and Dimensionality Reduction Transforms on Hyperspectral Data Transforms on Hyperspectral Data
as Applied to Target Detectionas Applied to Target Detection
Brian StaabEmmett Ientilucci
Digital Imaging and Remote Sensing LaboratoryCenter for Imaging Science
Rochester Institute of Technology
OverviewOverviewBackgroundBackground
Dimension & Noise Reduction TransformsDimension & Noise Reduction TransformsPrincipal Components Analysis (PCA)Principal Components Analysis (PCA)Maximum Noise Fraction (MNF)Maximum Noise Fraction (MNF)
Target Detection AlgorithmsTarget Detection AlgorithmsSpectral Angle Spectral Angle MapperMapper (SAM)(SAM)Orthogonal Subspace Projection (OSP)Orthogonal Subspace Projection (OSP)
Experiment Experiment
ResultsResultsA small sampleA small sample
ConclusionsConclusions
The TransformsThe Transforms
VarianceVariance
σσ22
PCA/MNF BandsPCA/MNF Bands1 2 3 4 5 …. 200
““NoiseNoise”” BandsBands
Too Much DataToo Much DataSimplify and Speed up AnalysisSimplify and Speed up Analysis
Without Throwing Away Information?Without Throwing Away Information?
Principal Components Analysis (PCA)Principal Components Analysis (PCA)Compress Variance to Few BandsCompress Variance to Few BandsDeDe--correlation via linear transformscorrelation via linear transforms
The TransformsThe Transforms
PCA maximizes variance of bandsPCA maximizes variance of bandsSignal & Signal & NoiseNoise MaximizedMaximized
Maximum Noise Fraction (MNF)Maximum Noise Fraction (MNF)Attempt to Attempt to ““separateseparate”” Noise and SignalNoise and SignalMaximize Signal to Noise RatioMaximize Signal to Noise RatioRequires Estimate of NoiseRequires Estimate of Noise
Dimension ReductionDimension Reduction
Noise ReductionNoise ReductionPCA/MNF SpacePCA/MNF Space
Zero the Noise BandsZero the Noise Bands
Inverse Transform to Original SpaceInverse Transform to Original Space““Noise Cleaned DataNoise Cleaned Data””
Only use the First Few PCA/MNF BandsOnly use the First Few PCA/MNF BandsContains the Most Variance or Information (compressed)Contains the Most Variance or Information (compressed)
Remaining Bands = Noise (Low Variance)Remaining Bands = Noise (Low Variance)
ConsiderConsider
How many dimensions do you keep?How many dimensions do you keep?
How will it affect target detection How will it affect target detection performance?performance?
Target Detection AlgorithmsTarget Detection AlgorithmsSpectral Angle Spectral Angle MapperMapper –– SAMSAM
Detection Statistic = Angle Detection Statistic = Angle
,
,
cos( )T
i j
i j
θ =t xt x
t
Target spectra
Xi,j
Test pixel
Band 1
Band 2
Target Detection AlgorithmsTarget Detection AlgorithmsOrthogonal Subspace Projection (OSP)Orthogonal Subspace Projection (OSP)
Geometrical ApproachGeometrical Approach
Matrix operatorMatrix operatorEliminate undesired spectral signatures (background) Eliminate undesired spectral signatures (background) contributioncontributionHow unlike the BackgroundHow unlike the Background
Vector operatorVector operatorMaximizes the contribution of desired target spectraMaximizes the contribution of desired target spectraHow TargetHow Target--LikeLike
HypothesisHypothesis
Dimension and Noise reduction Dimension and Noise reduction transforms on hyperspectral imagery transforms on hyperspectral imagery have no effect on target detection have no effect on target detection performanceperformance
Testing an AssumptionTesting an Assumptionthat there is no Effectthat there is no Effect
Radiance orRadiance orDC SpaceDC Space
ReflectanceReflectanceSpaceSpace
PCA/MNFPCA/MNFSpaceSpace
ELMELM PCAPCA
MNFMNF
ExperimentExperimentHyperspectral Hyperspectral Image CubeImage Cube
145
Band
s
145
Band
s
Atm
osph
eric
Atm
osph
eric
Com
pens
ation
Com
pens
ation
Experiment:Experiment:Band ReductionBand Reduction
Trim PCA/MNFTrim PCA/MNFDataData
PCA/MNFSpace
PCA/MNFsame transform
TargetTargetReflectanceReflectance
Target Detection-SAM-OSP
Record Detection Statistic
Iterate Number of Dimensions Trimmed
Experiment:Experiment:Noise ReductionNoise Reduction
Zero PCAZero PCADataData
Noise Noise Reduced Reduced ReflectanceReflectanceInverse
PCA
Target ReflectanceTarget Reflectance
Target Detection-SAM-OSP
Record StatisticIterate Number of Bands Zeroed
Experiment:Experiment:Band/Noise ReductionBand/Noise Reduction
Vary No. Bands Kept or ZeroedVary No. Bands Kept or ZeroedExamine Trends:Examine Trends:
Low and High Contrast TargetsLow and High Contrast TargetsSAM & OSPSAM & OSP
DetectionDetectionStatisticStatisticθθ
Number of Dimensions Kept Number of Dimensions Kept Or % Variance KeptOr % Variance Kept
100 % = All BandsLess % or Bands
Reference: Detection using all bands
AB
HydiceHydice Forest Radiance SceneForest Radiance Scene
141 x 116 Subset141 x 116 Subset145 Bands145 Bands.4 to 2.4 .4 to 2.4 µµmmELM AppliedELM Applied9 9 mbmb
High Contrast30 Fully Resolved pixels
Low Contrast15 Fully Resolved pixels
Zoom
Noise EstimateNoise EstimateNeeded for MNF TransformNeeded for MNF Transform20x17 pixel calibration panel20x17 pixel calibration panel
Homogeneous RegionHomogeneous RegionDifference of Original and Spatially Difference of Original and Spatially Shifted Version = NoiseShifted Version = Noise
Noise Std. Dev.
Vs. Band
Noise Covariance Matrix
ResultsResults
To compare with the trends of target pixelsTo compare with the trends of target pixels3 background classes were used as test 3 background classes were used as test pixelspixels
8 Graphs per Experiment: Dimension and Noise Reduction8 Graphs per Experiment: Dimension and Noise ReductionPCA & MNFPCA & MNFSAM & OSPSAM & OSPHigh & Low Contrast TargetsHigh & Low Contrast Targets
Results: PCA Band ReductionResults: PCA Band Reduction
High ContrastHigh ContrastVery Little EffectVery Little EffectMean of 30 PixelsMean of 30 Pixels
BackgroundBackgroundGrassGrassRoadRoadDirtDirt
PCA Band Reduction: SAM: Mean High Contrast
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
0 5 10 15 20
Number of Dimensions
Ang
le (R
adia
ns) Mean Target
GrassRoadDirtO
| 94.6% | 99.15% | 99.6 % | 99.69% % Variance Kept
Results: MNF Band ReductionResults: MNF Band Reduction
High ContrastHigh ContrastMore Separation More Separation with Background with Background ClassesClasses
Mean of 30 PixelsMean of 30 Pixels
MNF Band Reduction: SAM: Mean High Contrast
0.1
0.3
0.5
0.7
0.9
1.1
1.3
1.5
1.7
0 20 40 60 80 100 120 140
Number of Dimensions
Angl
e (r
adia
ns)
Target MeanGrassRoadRock/DirtG
| 73.6% | 97.99% | 99.37 % | 99.89% % Variance Kept
Results: MNF Noise ReductionResults: MNF Noise Reduction
Low ContrastLow Contrast
Improvement in Improvement in AngleAngle
Increase in Increase in background background separationseparation
IdealIdeal
MNF Noise Reduction: SAM: Mean Low Contrast
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
0 20 40 60 80 100 120 140
Number of Bands Kept
Angl
e (r
adia
ns)
MeanGrassRoadDirt/Rock
| 73.6% | 97.99% | 99.37 % | 99.89% % Variance Kept
Summary: SAMSummary: SAM
+ Improves / Better target Angle till significant noise reduction then:
++ Improves separation from background (Target Improves/ Background Worsens)
+ Target angle improves+ Improves separation from background
Low Contrast
+ Improves / Better target angle till significant noise reduction then:
+ Improves separation from background slightly
- Improves background angles
- Worsens target angle High Contrast
MNFPCANoiseReduction
- Worsens target angle- Confusion with background
+ Improves target Angle and+ Improves separation from background
slightly
Low Contrast
+ Improves target angle+ Improves separation from
background
+ Improves target angle- Improves background angles
High Contrast
MNFPCABandReduction
Target relatively constant till significant Target relatively constant till significant band/noise reduction then:band/noise reduction then:
Summary: OSPSummary: OSP
- Worsens target till 99.8% variance (~110 bands) kept then:
• Constant target/backgrounds till significant noise reduction then:
- Worsens target statistic+ Small area of improvement of target &
background separation
Low Contrast
- Worsens target statistic
+ Small area of improvement of target & background separation
• Target Relatively Constant till significant noise reduction then:
- Worsens Target statistic
High Contrast
MNFPCANoiseReduction
- Confusion with background• Target relatively constant till significant
band reduction then:- Worsens Target Angle
+ Improves Target and + Improves Separation from background
till significant band reduction then:- Background Increase but at lower rate
Low Contrast
+ Improves target slightly + Improves Separation from Background
till significant band reduction
+ Improves target slightly + Improves separation from Background
till significant band reduction
High Contrast
MNFPCABand
Reduction
EndmemberEndmember changes cause drastic fluctuations, some changes cause drastic fluctuations, some cause background confusioncause background confusion
ConclusionConclusion
Band ReductionBand ReductionDetection statistic remains relatively Detection statistic remains relatively constant or improvesconstant or improvesSome cases: Increase in separationSome cases: Increase in separation
Noise ReductionNoise ReductionSAM angle constant or improvesSAM angle constant or improves
Background Separation increasesBackground Separation increases
OSP statistic constant or deterioratesOSP statistic constant or deterioratesDrastic fluctuations (Drastic fluctuations (EndmembersEndmembers))
ConclusionConclusion
Future Considerations:Future Considerations:Use of varyingUse of varying
Data setsData setsTarget Detection AlgorithmsTarget Detection AlgorithmsTargetsTargets
Better Noise EstimatesBetter Noise Estimates
QuestionsQuestions