classification of hyperspectral data using spectral...
TRANSCRIPT
Classification of Hyperspectral Data UsingSpectral-Spatial Approaches
Yuliya Tarabalka
GIPSA-Lab, Grenoble Institute of Technology, FranceDepartment of Electrical and Computer Engineering, University of Iceland, Iceland
Thesis committee
Melba CRAWFORD PresidentJean-Yves TOURNERET OpponentXiuping JIA OpponentAntonio PLAZA OpponentPaolo GAMBA OpponentJohannes R. SVEINSSON ExaminatorJón Atli BENEDIKTSSON Thesis advisorJocelyn CHANUSSOT Thesis advisor
June 14, 2010
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Hyperspectral image
Every pixel contains a detailed spectrum(>100 spectral bands)
+ More information per pixel
− Dimensionality increases
⇓Advanced algorithms are required!
λ pixel vector xi
xi xix1 x2 x3x2i
xn
lines
samples
Tens or hundreds of bands wavelength λ
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 2
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Supervised classification problem
ROSIS imageSpatial resolution: 1.3m/pixSpectral resolution: 103 bands
Ground-truth data Task
Nine classes: alphalt, meadows, gravel, trees, metal sheets, bare soil,bitumen, bricks, shadows
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 3
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Supervised classification problem
AVIRIS imageSpatial resolution: 20m/pix
Spectral resolution: 200 bands
Ground-truthdata
Task
16 classes: corn-no till, corn-min till, corn, soybeans-no till, soybeans-mintill, soybeans-clean till, alfalfa, grass/pasture, grass/trees,grass/pasture-mowed, hay-windrowed, oats, wheat, woods,
bldg-grass-tree-drives, stone-steel towers
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 4
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Classification approaches
Only spectral information
Pixelwise approach
Spectrum of each pixel is analyzedVariety of methods
Support Vector Machines (SVM) → goodclassification results [Camps-Valls05]
⇒
alphaltmeadowsgraveltrees
metal sheetsbare soilbitumenbricks
shadows
Overall accuracy (OA) = 81.01%
Average accuracy (AA) = 88.25%
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 5
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Classification approaches
Only spectral information
Pixelwise approach
Spectrum of each pixel is analyzedVariety of methods
Support Vector Machines (SVM) → goodclassification results [Camps-Valls05]
Spectral + spatial information
Info about spatial structures is includedSince neighboring pixels are related
How to define spatial structures?
How to combine spectral and spatialinformation?
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 6
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
State of the art: Approaches for extracting spatial info
1 Closest fixed neighborhoodsMarkov Random Field [Pony00, Jackson02,Farag05]Contextual features [Camps-Valls06]+ Simplicity− Imprecision at the border of regions
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 7
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
State of the art: Approaches for extracting spatial info
1 Closest fixed neighborhoodsMarkov Random Field [Pony00, Jackson02,Farag05]Contextual features [Camps-Valls06]+ Simplicity− Imprecision at the border of regions
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 7
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
State of the art: Approaches for extracting spatial info
1 Closest fixed neighborhoodsMarkov Random Field [Pony00, Jackson02,Farag05]Contextual features [Camps-Valls06]+ Simplicity− Imprecision at the border of regions
2 Morphological and area filteringMorphological profiles [Pesaresi01, Dell’Acqua04, Benediktsson05]Self-complementary area filtering [Fauvel07]+ Neighborhoods are adapted to the structures− Neighborhoods are scale dependent ⇒ imprecision in the spatial info
Closing − Original − Opening
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 7
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
State of the art: Approaches for extracting spatial info
1 Closest fixed neighborhoodsMarkov Random Field [Pony00, Jackson02,Farag05]Contextual features [Camps-Valls06]+ Simplicity− Imprecision at the border of regions
2 Morphological and area filteringMorphological profiles [Pesaresi01, Dell’Acqua04, Benediktsson05]Self-complementary area filtering [Fauvel07]+ Neighborhoods are adapted to the structures− Neighborhoods are scale dependent ⇒ imprecision in the spatial info
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 7
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
State of the art: Approaches for extracting spatial info
3 Segmentation map = exhaustive partitioning of the image intohomogeneous regions
Extraction and Classification of Homogeneous Objects [Kettig76]+ Has become a standard spectral-spatial classification technique− Statistical approach ⇒ not well adapted for hyperspectral data
Recent works:
Multiscale segmentation + SVM classification [Linden07, Huang09]− Computationally demanding
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 8
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
State of the art: Approaches for extracting spatial info
3 Segmentation map = exhaustive partitioning of the image intohomogeneous regions
Extraction and Classification of Homogeneous Objects [Kettig76]+ Has become a standard spectral-spatial classification technique− Statistical approach ⇒ not well adapted for hyperspectral data
Recent works:
Multiscale segmentation + SVM classification [Linden07, Huang09]− Computationally demanding
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 8
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
State of the art: Approaches for extracting spatial info
3 Segmentation map = exhaustive partitioning of the image intohomogeneous regions
Extraction and Classification of Homogeneous Objects [Kettig76]+ Has become a standard spectral-spatial classification technique− Statistical approach ⇒ not well adapted for hyperspectral data
Recent works:
Multiscale segmentation + SVM classification [Linden07, Huang09]− Computationally demanding
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 8
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Thesis objective: combine spectral and spatial info
Proposed spectral-spatial classification approaches
Using closest fixed neighborhoods
SVM and MRF-based classification
Using adaptive neighborhoods
Using spatial neighborhoods derived from unsupervised
segmentation
Segmentation Classification Using marker-based region growing segmentation
Marker selection ClassificationWatershed Partitional
clustering Hierarchical
segmentation Object-based
Composite kernels
Segmentation + pixelwise
classification
Using ML discriminant
values
Using probabilistic
SVM
Multiple classifier approach
Marker-controlled watershed
Construction of a Minimum
Spanning Forest
Strategy 1
Strategy 2
Strategy 3
High-performance computing
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 9
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Thesis objective: combine spectral and spatial info
Proposed spectral-spatial classification approaches
Construction of a Minimum
Spanning Forest
Using closest fixed neighborhoods
SVM and MRF-based classification
Using adaptive neighborhoods
Using spatial neighborhoods derived from unsupervised
segmentation
Segmentation Classification Using marker-based region growing segmentation
Marker selection ClassificationWatershed Partitional
clustering Hierarchical
segmentation Object-based
Composite kernels
Segmentation + pixelwise
classification
Using ML discriminant
values
Using probabilistic
SVM
Multiple classifier approach
Strategy 1
Strategy 2
Marker-controlled watershed
Strategy 3
High-performance computing
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 9
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Outline
1 Introduction
2 Classification using segmentation-derived adaptive neighborhoodsSegmentationSpectral-spatial classificationConcluding discussion
3 Segmentation and classification using automatically selected markersMarker selection
Using probabilistic SVMMultiple classifier approach
MSF-based marker-controlled region growingConcluding discussion
4 Conclusions and perspectives
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 10
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
SegmentationSpectral-spatial classificationConcluding discussion
Objective
1 Automatically segment a hyperspectral image into homogeneousregions
2 Spectral info + segmentation map → classify image
+ → Classificationmap
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 11
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
SegmentationSpectral-spatial classificationConcluding discussion
Outline
1 Introduction
2 Classification using segmentation-derived adaptive neighborhoodsSegmentationSpectral-spatial classificationConcluding discussion
3 Segmentation and classification using automatically selected markersMarker selection
Using probabilistic SVMMultiple classifier approach
MSF-based marker-controlled region growingConcluding discussion
4 Conclusions and perspectives
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 12
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
SegmentationSpectral-spatial classificationConcluding discussion
1. Watershed segmentation
gradient⇒
Region growing method:
Minimum of a gradient = core of a homogeneous region
1 region = set of pixels connected to 1 local minimum of the gradient
Watershed lines = edges between adjacent regions
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 13
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
SegmentationSpectral-spatial classificationConcluding discussion
1. Watershed segmentation
Originalimage
Y. Tarabalka, J. Chanussot, and J. A. Benediktsson, “Segmentation and classification of hyperspectralimages using watershed transformation,” Pattern Recognition, vol. 43, no. 7, pp. 2367-2379, July2010.
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 14
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
SegmentationSpectral-spatial classificationConcluding discussion
1. Watershed segmentation
Originalimage
⇒
Robust ColorMorpho Gradient
Y. Tarabalka, J. Chanussot, and J. A. Benediktsson, “Segmentation and classification of hyperspectralimages using watershed transformation,” Pattern Recognition, vol. 43, no. 7, pp. 2367-2379, July2010.
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 14
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
SegmentationSpectral-spatial classificationConcluding discussion
1. Watershed segmentation
Originalimage
⇒
Robust ColorMorpho Gradient
⇒
Watershed11802 regions
Y. Tarabalka, J. Chanussot, and J. A. Benediktsson, “Segmentation and classification of hyperspectralimages using watershed transformation,” Pattern Recognition, vol. 43, no. 7, pp. 2367-2379, July2010.
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 14
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
SegmentationSpectral-spatial classificationConcluding discussion
1. Watershed segmentation
Originalimage
⇒
Robust ColorMorpho Gradient
⇒
Watershed11802 regions
⇒
Edges → toadjacent regions
Y. Tarabalka, J. Chanussot, and J. A. Benediktsson, “Segmentation and classification of hyperspectralimages using watershed transformation,” Pattern Recognition, vol. 43, no. 7, pp. 2367-2379, July2010.
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 14
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
SegmentationSpectral-spatial classificationConcluding discussion
2. Partitional clustering (EM)
1 Clusteringpixels are grouped into C clustersin each cluster → pixels drawn froma Gaussian distributiondistribution parameters → EM algorithm
2 Labeling of connected components
10 clusters⇒
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 15
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
SegmentationSpectral-spatial classificationConcluding discussion
2. Partitional clustering (EM)
1 Clusteringpixels are grouped into C clustersin each cluster → pixels drawn froma Gaussian distributiondistribution parameters → EM algorithm
2 Labeling of connected components
10 clusters⇒
21450regions
same cluster,but differentregions!
XXXXXXXz
HHHHHj
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 15
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
SegmentationSpectral-spatial classificationConcluding discussion
3. Hierarchical SEGmentation (HSEG,[Tilton98])
Region growing + Spectral Clustering
Dissimilarity criterion (DC):Spectral Angle Mapper (SAM)between the region mean vectors ui and uj
SAM(ui , uj) = arccos(ui · uj
‖ui‖2‖uj‖2)
1 Each pixel - one region2 Find DCmin between adjacent regions3 Merge adjacent regions with DC = DCmin4 Merge non-adjacent regions withDC ≤ DCmin ·SpectralClusterWeight
5 If not converge, go to 2
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 16
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
SegmentationSpectral-spatial classificationConcluding discussion
3. Hierarchical SEGmentation (HSEG,[Tilton98])
Region growing + Spectral Clustering
Dissimilarity criterion (DC):Spectral Angle Mapper (SAM)between the region mean vectors ui and uj
SAM(ui , uj) = arccos(ui · uj
‖ui‖2‖uj‖2)
1 Each pixel - one region2 Find DCmin between adjacent regions3 Merge adjacent regions with DC = DCmin4 Merge non-adjacent regions withDC ≤ DCmin ·SpectralClusterWeight
5 If not converge, go to 2
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 16
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
SegmentationSpectral-spatial classificationConcluding discussion
3. Hierarchical SEGmentation (HSEG,[Tilton98])
Region growing + Spectral Clustering
Dissimilarity criterion (DC):Spectral Angle Mapper (SAM)between the region mean vectors ui and uj
SAM(ui , uj) = arccos(ui · uj
‖ui‖2‖uj‖2)
1 Each pixel - one region2 Find DCmin between adjacent regions3 Merge adjacent regions with DC = DCmin4 Merge non-adjacent regions withDC ≤ DCmin ·SpectralClusterWeight
5 If not converge, go to 2
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 16
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
SegmentationSpectral-spatial classificationConcluding discussion
3. Hierarchical SEGmentation (HSEG,[Tilton98])
Region growing + Spectral Clustering
Dissimilarity criterion (DC):Spectral Angle Mapper (SAM)between the region mean vectors ui and uj
SAM(ui , uj) = arccos(ui · uj
‖ui‖2‖uj‖2)
1 Each pixel - one region2 Find DCmin between adjacent regions3 Merge adjacent regions with DC = DCmin4 Merge non-adjacent regions withDC ≤ DCmin ·SpectralClusterWeight
5 If not converge, go to 2
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 16
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
SegmentationSpectral-spatial classificationConcluding discussion
3. Hierarchical SEGmentation (HSEG,[Tilton98])
Region growing + Spectral Clustering
Dissimilarity criterion (DC):Spectral Angle Mapper (SAM)between the region mean vectors ui and uj
SAM(ui , uj) = arccos(ui · uj
‖ui‖2‖uj‖2)
1 Each pixel - one region2 Find DCmin between adjacent regions3 Merge adjacent regions with DC = DCmin4 Merge non-adjacent regions withDC ≤ DCmin ·SpectralClusterWeight
5 If not converge, go to 2
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 16
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
SegmentationSpectral-spatial classificationConcluding discussion
3. Hierarchical SEGmentation (HSEG,[Tilton98])
Region growing + Spectral Clustering
Dissimilarity criterion (DC):Spectral Angle Mapper (SAM)between the region mean vectors ui and uj
SAM(ui , uj) = arccos(ui · uj
‖ui‖2‖uj‖2)
1 Each pixel - one region2 Find DCmin between adjacent regions3 Merge adjacent regions with DC = DCmin4 Merge non-adjacent regions withDC ≤ DCmin ·SpectralClusterWeight
5 If not converge, go to 2
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 16
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
SegmentationSpectral-spatial classificationConcluding discussion
3. Hierarchical SEGmentation (HSEG,[Tilton98])
Region growing + Spectral Clustering
Dissimilarity criterion (DC):Spectral Angle Mapper (SAM)between the region mean vectors ui and uj
SAM(ui , uj) = arccos(ui · uj
‖ui‖2‖uj‖2)
1 Each pixel - one region2 Find DCmin between adjacent regions3 Merge adjacent regions with DC = DCmin4 Merge non-adjacent regions withDC ≤ DCmin ·SpectralClusterWeight
5 If not converge, go to 2
7575 regions
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 16
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
SegmentationSpectral-spatial classificationConcluding discussion
Outline
1 Introduction
2 Classification using segmentation-derived adaptive neighborhoodsSegmentationSpectral-spatial classificationConcluding discussion
3 Segmentation and classification using automatically selected markersMarker selection
Using probabilistic SVMMultiple classifier approach
MSF-based marker-controlled region growingConcluding discussion
4 Conclusions and perspectives
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 17
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
SegmentationSpectral-spatial classificationConcluding discussion
Spectral-spatial classification: majority vote
Hyperspectral image (B bands)
Pixelwise classification Segmentation
Spectral-spatial classification
(majority voting)
Final classification map
Pixelwise classification map (dark blue, white
and light grey classes)
Segmentation map
(3 spatial regions)
Result of spectral-spatial classification (classification
map after majority vote)
Combination of unsupervised
segmentation and pixelwise
classification results (majority vote within
3 spatial regions)
1 1 1 1 1 1 1 1 1 2 2 2 1 1 2 2 2 2 1 1 2 2 2 2 1 1 3 3 3 2 1 3 3 3 3 3 3 3 3 3 3 3
Y. Tarabalka, J. A. Benediktsson, and J. Chanussot, “Spectral-spatial classification of hyperspectralimagery based on partitional clustering techniques,” IEEE Trans. on Geoscience and Remote Sensing,vol. 47, no. 8, pp. 2973-2987, Aug. 2009.
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 18
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
SegmentationSpectral-spatial classificationConcluding discussion
Spectral-spatial classification
Originalimage
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 19
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
SegmentationSpectral-spatial classificationConcluding discussion
Spectral-spatial classification
Originalimage
⇒
SVMclassification
OA = 81.01%AA = 88.25%
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 19
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
SegmentationSpectral-spatial classificationConcluding discussion
Spectral-spatial classification
SVMclassification
OA = 81.01%AA = 88.25%
⇒
SVM +Watershed
OA = 85.42%AA = 91.31%
SVM +Partit.clustering
OA = 94.00%AA = 93.13%
SVM +HSEG
OA = 93.85%AA = 97.07%
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 19
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
SegmentationSpectral-spatial classificationConcluding discussion
Classification accuracies (%)
SVM +Watersh. +Part.Cl. +HSEG EMP1 ECHO
Overall Acc. 81.01 85.42 94.00 93.85 85.22 87.58Average Acc. 88.25 91.31 93.13 97.07 90.76 92.16Kappa Coef.κ 75.86 81.30 91.93 91.89 80.86 83.90Asphalt 84.93 93.64 90.10 94.77 95.36 87.98Meadows 70.79 75.09 95.99 89.32 80.33 81.64Gravel 67.16 66.12 82.26 96.14 87.61 76.91Trees 97.77 98.56 85.54 98.08 98.37 99.31Metal sheets 99.46 99.91 100 99.82 99.48 99.91Bare soil 92.83 97.35 96.72 99.76 63.72 93.96Bitumen 90.42 96.23 91.85 100 98.87 92.97Bricks 92.78 97.92 98.34 99.29 95.41 97.35Shadows 98.11 96.98 97.36 96.48 97.68 99.37
1A. Plaza et al., "Recent advances in techniques for hyperspectral image processing," RemoteSensing of Environment, vol. 113, Suppl. 1, 2009.
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 20
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
SegmentationSpectral-spatial classificationConcluding discussion
Outline
1 Introduction
2 Classification using segmentation-derived adaptive neighborhoodsSegmentationSpectral-spatial classificationConcluding discussion
3 Segmentation and classification using automatically selected markersMarker selection
Using probabilistic SVMMultiple classifier approach
MSF-based marker-controlled region growingConcluding discussion
4 Conclusions and perspectives
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 21
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
SegmentationSpectral-spatial classificationConcluding discussion
1 Spectral-spatial classification improves accuracies when compared topixel-wise classification
2 Several segmentation techniques are investigated3 The HSEG segmentation map leads to the best classification4 Obtained classification accuracies > all previous results
However...
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 22
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
SegmentationSpectral-spatial classificationConcluding discussion
Unsupervised segmentation
Unsupervised segmentation = exhaustive partitioning into homogeneousregionsHow to define a measure of homogeneity?
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 23
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
SegmentationSpectral-spatial classificationConcluding discussion
Unsupervised segmentation
Unsupervised segmentation = exhaustive partitioning into homogeneousregionsHow to define a measure of homogeneity?
Hierarchical SEGmentation (HSEG, [Tilton98])
Image
⇒
2381reg. 1389reg. 280reg.
Oversegmentation! Undersegmentation?@@I ��� 6
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 23
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
SegmentationSpectral-spatial classificationConcluding discussion
Unsupervised segmentation
Unsupervised segmentation = exhaustive partitioning into homogeneousregionsHow to define a measure of homogeneity?
Hierarchical SEGmentation (HSEG, [Tilton98])
Image
⇒
2381reg. 1389reg. 280reg.
Oversegmentation Undersegmentation?⇓
Preferred in previous works → not to lose objects
���
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 23
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
SegmentationSpectral-spatial classificationConcluding discussion
Watershed segmentation
Originalimage
⇒
Robust ColorMorpho Gradient
⇒
Watershed1277 regions
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 24
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
SegmentationSpectral-spatial classificationConcluding discussion
Watershed segmentation
Originalimage
⇒
Robust ColorMorpho Gradient
⇒
Watershed1277 regions
�������3
Severe oversegmentation!
Every local minimumof the gradient
↓one region
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 24
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
SegmentationSpectral-spatial classificationConcluding discussion
Marker-controlled segmentation
Reduce oversegmentation ⇐ incorporate an additional knowledge intosegmentation
We propose to use markers
Determine markers foreach region of interest
⇒
Segmentan image
One regionin a segmentation map
⇑One marker
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 25
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Marker selectionMSF-based marker-controlled region growingConcluding discussion
Objective
Determine markers automatically ← using probabilisticclassification results
Marker-controlled region growing→ segment and classify ahyperspectral image
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 26
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Marker selectionMSF-based marker-controlled region growingConcluding discussion
Outline
1 Introduction
2 Classification using segmentation-derived adaptive neighborhoodsSegmentationSpectral-spatial classificationConcluding discussion
3 Segmentation and classification using automatically selected markersMarker selection
Using probabilistic SVMMultiple classifier approach
MSF-based marker-controlled region growingConcluding discussion
4 Conclusions and perspectives
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 27
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Marker selectionMSF-based marker-controlled region growingConcluding discussion
Marker selection using probabilistic SVM
B-band hyperspectral imageX = {xj ∈ RB, j = 1, 2, ..., n}B ∼ 100
Hyperspectral image (B bands)
Probabilistic pixelwise
classification classification map
Selection of the most probability map
map of Marker-controlled
region growing markers reliable classified
pixels
Segmentation map + classification map
Y. Tarabalka, J. Chanussot, and J. A. Benediktsson, “Segmentation and classification of hyperspectralimages using Minimum Spanning Forest grown from automatically selected markers,” IEEE Trans. onSystems, Man, and Cybernetics, Part B: Cybernetics, 2010, DOI 10.1109/TSMCB.2009.2037132.
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 28
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Marker selectionMSF-based marker-controlled region growingConcluding discussion
Marker selection using probabilistic SVM
SVM classifier* → well suited forhyperspectral images
Output:
classification map probability map
-
Hyperspectral image (B bands)
Probabilistic pixelwise
classification classification map
Selection of the most probability map
map of Marker-controlled
region growing markers reliable classified
pixels
Segmentation map + classification map
probability estimate for each pixel tobelong to the assigned class
*C. Chang and C. Lin, "LIBSVM: A library for Support Vector Machines," Software available at
http://www.csie.ntu.edu.tw/∼cjlin/libsvm, 2001.
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 29
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Marker selectionMSF-based marker-controlled region growingConcluding discussion
Marker selection using probabilistic SVM
Analysis of classification and probability maps:classification map probability map
1 Perform connected components labeling ofthe classification map
2 Analyze each connected component:
If it is large (> 20 pixels) → use P%(5%) of its pixels with the highestprobabilities as a markerIf it is small → its pixels withprobabilities > T% (90%)are used as a marker
Hyperspectral image (B bands)
Probabilistic pixelwise
classification classification map
Selection of the most probability map
map of Marker-controlled
region growing markers reliable classified
pixels
Segmentation map + classification map
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 30
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Marker selectionMSF-based marker-controlled region growingConcluding discussion
Marker selection using probabilistic SVM
Analysis of classification and probability maps:classification map probability map
1 Perform connected components labeling ofthe classification map
2 Analyze each connected component:
If it is large (> 20 pixels) → use P%(5%) of its pixels with the highestprobabilities as a markerIf it is small → its pixels withprobabilities > T% (90%)are used as a marker
Hyperspectral image (B bands)
Probabilistic pixelwise
classification classification map
Selection of the most probability map
map of Marker-controlled
region growing markers reliable classified
pixels
Segmentation map + classification map HH
HY
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 30
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Marker selectionMSF-based marker-controlled region growingConcluding discussion
Marker selection using probabilistic SVM
Analysis of classification and probability maps:classification map probability map
1 Perform connected components labeling ofthe classification map
2 Analyze each connected component:
If it is large (> 20 pixels) → use P%(5%) of its pixels with the highestprobabilities as a markerIf it is small → its pixels withprobabilities > T% (90%)are used as a marker
Hyperspectral image (B bands)
Probabilistic pixelwise
classification classification map
Selection of the most probability map
map of Marker-controlled
region growing markers reliable classified
pixels
Segmentation map + classification map
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 30
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Marker selectionMSF-based marker-controlled region growingConcluding discussion
Marker selection using probabilistic SVM
Analysis of classification and probability maps:classification map probability map
1 Perform connected components labeling ofthe classification map
2 Analyze each connected component:
If it is large (> 20 pixels) → use P%(5%) of its pixels with the highestprobabilities as a markerIf it is small → its pixels withprobabilities > T% (90%)are used as a marker
Hyperspectral image (B bands)
Probabilistic pixelwise
classification classification map
Selection of the most probability map
map of Marker-controlled
region growing markers reliable classified
pixels
Segmentation map + classification map
Must contain a marker!
HHHHH
HHHHHH
HHHHHj
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 30
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Marker selectionMSF-based marker-controlled region growingConcluding discussion
Marker selection using probabilistic SVM
Analysis of classification and probability maps:classification map probability map
1 Perform connected components labeling ofthe classification map
2 Analyze each connected component:
If it is large (> 20 pixels) → use P%(5%) of its pixels with the highestprobabilities as a markerIf it is small → its pixels withprobabilities > T% (90%)are used as a marker
Hyperspectral image (B bands)
Probabilistic pixelwise
classification classification map
Selection of the most probability map
map of Marker-controlled
region growing markers reliable classified
pixels
Segmentation map + classification map
Must contain a marker!
HHHHH
HHHHHH
HHHHHj
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 30
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Marker selectionMSF-based marker-controlled region growingConcluding discussion
Marker selection using probabilistic SVM
Analysis of classification and probability maps:classification map probability map
1 Perform connected components labeling ofthe classification map
2 Analyze each connected component:
If it is large (> 20 pixels) → use P%(5%) of its pixels with the highestprobabilities as a markerIf it is small → its pixels withprobabilities > T% (90%)are used as a marker
Hyperspectral image (B bands)
Probabilistic pixelwise
classification classification map
Selection of the most probability map
map of Marker-controlled
region growing markers reliable classified
pixels
Segmentation map + classification map
Has a marker only if it isvery reliable
HHHH
HHHHHH
HHHHHHj
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 30
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Marker selectionMSF-based marker-controlled region growingConcluding discussion
Marker selection using probabilistic SVM
Analysis of classification and probability maps:classification map probability map
Each connected component → 1 or 0 marker(2250 regions → 107 markers)
Marker is not necessarily a connected set ofpixels
Each marker has a class label
Hyperspectral image (B bands)
Probabilistic pixelwise
classification classification map
Selection of the most probability map
map of Marker-controlled
region growing markers reliable classified
pixels
Segmentation map + classification map
map of 107 markers
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 30
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Marker selectionMSF-based marker-controlled region growingConcluding discussion
Multiple classifier approach for marker selection
Previous method: strong dependence on the performances of theselected probabilistic classifier
Objective: mitigate this dependence→ using multiple classifiers
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 31
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Marker selectionMSF-based marker-controlled region growingConcluding discussion
Multiple classifier approach for marker selection
Previous method: strong dependence on the performances of theselected probabilistic classifier
Objective: mitigate this dependence→ using multiple classifiers
Multiple classifier marker selection approach
1 Classify an image by severalindependent classifiers
2 Pixels assigned by all theclassifiers to the same class
⇓Map of markers
classifiers
Hyperspectral image
Map ofmarkers
Marker selection
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 31
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Marker selectionMSF-based marker-controlled region growingConcluding discussion
Multiple spectral-spatial classifier marker selection
Take into account spatial context
Pixelwise classification
Hyperspectral image (B bands)
Majority voting within segmentation
regions
Watershed segmentation
Marker selection Map of markers Segmentation by EM for Gaussian mixture resolving
Hierarchical SEGmentation
Majority voting within segmentation
regions
Majority voting within segmentation
regions
classification map
classification map
classification map
Y. Tarabalka, J. A. Benediktsson, J. Chanussot, and J. C. Tilton, “Multiple spectral-spatialclassification approach for hyperspectral data,” submitted to IEEE Trans. on Geoscience andRemote Sensing (under review).
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 32
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Marker selectionMSF-based marker-controlled region growingConcluding discussion
Multiple spectral-spatial classifier marker selection
Take into account spatial context
Pixelwise classification
Hyperspectral image (B bands)
Majority voting within segmentation
regions
Watershed segmentation
Marker selection Map of markers Segmentation by EM for Gaussian mixture resolving
Hierarchical SEGmentation
Majority voting within segmentation
regions
Majority voting within segmentation
regions
classification map
classification map
classification map
Y. Tarabalka, J. A. Benediktsson, J. Chanussot, and J. C. Tilton, “Multiple spectral-spatialclassification approach for hyperspectral data,” submitted to IEEE Trans. on Geoscience andRemote Sensing (under review).
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 32
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Marker selectionMSF-based marker-controlled region growingConcluding discussion
Multiple spectral-spatial classifier marker selection
Take into account spatial context
Pixelwise classification
Hyperspectral image (B bands)
Majority voting within segmentation
regions
Watershed segmentation
Marker selection Map of markers Segmentation by EM for Gaussian mixture resolving
Hierarchical SEGmentation
Majority voting within segmentation
regions
Majority voting within segmentation
regions
classification map
classification map
classification map
Y. Tarabalka, J. A. Benediktsson, J. Chanussot, and J. C. Tilton, “Multiple spectral-spatialclassification approach for hyperspectral data,” submitted to IEEE Trans. on Geoscience andRemote Sensing (under review).
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 32
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Marker selectionMSF-based marker-controlled region growingConcluding discussion
Multiple spectral-spatial classifier marker selection
Take into account spatial context
Pixelwise classification
Hyperspectral image (B bands)
Majority voting within segmentation
regions
Watershed segmentation
Marker selection Map of markers Segmentation by EM for Gaussian mixture resolving
Hierarchical SEGmentation
Majority voting within segmentation
regions
Majority voting within segmentation
regions
classification map
classification map
classification map
Y. Tarabalka, J. A. Benediktsson, J. Chanussot, and J. C. Tilton, “Multiple spectral-spatialclassification approach for hyperspectral data,” submitted to IEEE Trans. on Geoscience andRemote Sensing (under review).
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 32
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Marker selectionMSF-based marker-controlled region growingConcluding discussion
Multiple spectral-spatial classifier marker selection
Take into account spatial context
Pixelwise classification
Hyperspectral image (B bands)
Majority voting within segmentation
regions
Watershed segmentation
Marker selection Map of markers Segmentation by EM for Gaussian mixture resolving
Hierarchical SEGmentation
Majority voting within segmentation
regions
Majority voting within segmentation
regions
classification map
classification map
classification map
Y. Tarabalka, J. A. Benediktsson, J. Chanussot, and J. C. Tilton, “Multiple spectral-spatialclassification approach for hyperspectral data,” submitted to IEEE Trans. on Geoscience andRemote Sensing (under review).
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 32
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Marker selectionMSF-based marker-controlled region growingConcluding discussion
Outline
1 Introduction
2 Classification using segmentation-derived adaptive neighborhoodsSegmentationSpectral-spatial classificationConcluding discussion
3 Segmentation and classification using automatically selected markersMarker selection
Using probabilistic SVMMultiple classifier approach
MSF-based marker-controlled region growingConcluding discussion
4 Conclusions and perspectives
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 33
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Marker selectionMSF-based marker-controlled region growingConcluding discussion
Construction of a Minimum Spanning Forest (MSF)
Hyperspectral image (B bands)
Classification
map of Construction of minimum spanning
forest
markers Marker selection
Majority voting within connected
components
Segmentation map + classification map
Y. Tarabalka, J. Chanussot, and J. A. Benediktsson, “Segmentation and classification of hyperspectralimages using Minimum Spanning Forest grown from automatically selected markers,” IEEE Trans. onSystems, Man, and Cybernetics, Part B: Cybernetics, 2010, DOI 10.1109/TSMCB.2009.2037132.
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 34
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Marker selectionMSF-based marker-controlled region growingConcluding discussion
Construction of a Minimum Spanning Forest (MSF)
B bands x 9x 8x 7
x 6x 5x x1 x2 x3
4
map of markers
1 1 0 0 0 0 0 0 2
1) Map an image onto a graphWeight wi ,j indicates the degree of dissimilarity between pixels xi and xj .Spectral Angle Mapper (SAM) distance can be used:
wi ,j = SAM(xi , xj) = arccos
( ∑Bb=1 xibxjb
[∑B
b=1 x2ib]1/2[∑B
b=1 x2jb]1/2
)
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 35
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Marker selectionMSF-based marker-controlled region growingConcluding discussion
Construction of a Minimum Spanning Forest (MSF)
B bands x 9x 8x 7
x 6x 5x x1 x2 x3
4
map of markers
1 1 0 0 0 0 0 0 2 image graph
1) Map an image onto a graphWeight wi ,j indicates the degree of dissimilarity between pixels xi and xj .Spectral Angle Mapper (SAM) distance can be used:
wi ,j = SAM(xi , xj) = arccos
( ∑Bb=1 xibxjb
[∑B
b=1 x2ib]1/2[∑B
b=1 x2jb]1/2
)
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 35
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Marker selectionMSF-based marker-controlled region growingConcluding discussion
Construction of a Minimum Spanning Forest (MSF)
B bands x 9x 8x 7
x 6x 5x 4
x3 x2 x1
map of markers
1 1 0 0 0 0 0 0 2
w12
image graph
1) Map an image onto a graphWeight wi ,j indicates the degree of dissimilarity between pixels xi and xj .Spectral Angle Mapper (SAM) distance can be used:
wi ,j = SAM(xi , xj) = arccos
( ∑Bb=1 xibxjb
[∑B
b=1 x2ib]1/2[∑B
b=1 x2jb]1/2
)
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 35
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Marker selectionMSF-based marker-controlled region growingConcluding discussion
Construction of a Minimum Spanning Forest (MSF)
B bands x 9x 8x 7
x 6x 5x 4
x3 x2 x1
map of markers
1 1 0 0 0 0 0 0 2 image graph
0 8 14 2
8 5 15
9 5164
1211 12 314 10
511
8
1) Map an image onto a graphWeight wi ,j indicates the degree of dissimilarity between pixels xi and xj .Spectral Angle Mapper (SAM) distance can be used:
wi ,j = SAM(xi , xj) = arccos
( ∑Bb=1 xibxjb
[∑B
b=1 x2ib]1/2[∑B
b=1 x2jb]1/2
)
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 35
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Marker selectionMSF-based marker-controlled region growingConcluding discussion
Construction of a Minimum Spanning Forest (MSF)
1 1
2
B bands x 9x 8x 7
x 6x 5x 4
x3 x2 x1
map of markers
1 1 0 0 0 0 0 0 2
80 0 14 2
8 5 15
image graph
0 0 9 5
164 0 1211
12 314 10 5
11 80 0
1) Map an image onto a graphWeight wi ,j indicates the degree of dissimilarity between pixels xi and xj .Spectral Angle Mapper (SAM) distance can be used:
wi ,j = SAM(xi , xj) = arccos
( ∑Bb=1 xibxjb
[∑B
b=1 x2ib]1/2[∑B
b=1 x2jb]1/2
)
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 35
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Marker selectionMSF-based marker-controlled region growingConcluding discussion
Construction of a Minimum Spanning Forest (MSF)
80
1 1
2
02 14
5 8 15 9
image graph
5164 0 0 0
11 12 14 12 3 10 11
8 5 0 0
Given a graph G, a MSF F ∗ rooted on vertices {r1, ..., rm} is:a non-connected graph without cycles
contains all the vertices of G
consists of connected subgraphs, each subgraph (tree) contains (isrooted on) one root risum of the edges weights of F ∗ is minimal
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 36
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Marker selectionMSF-based marker-controlled region growingConcluding discussion
Construction of a Minimum Spanning Forest (MSF)
0
1 1
2
modified graph
0
0 0
0
0
0 0
3
8
164
8 5
5
14
5
2
9
14
10 1111 12
8
12
15
r2
r1 0
0
2) Add m extra vertices ri , i = 1, ..., m corresponding to m markers
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 37
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Marker selectionMSF-based marker-controlled region growingConcluding discussion
Construction of a Minimum Spanning Forest (MSF)
0
1 1
2
0
0 0
0
0
0 0
3
8
164
8 5
5
14
5
2
9
14
10 1111 12
8
12
15
r2
r1 0
0
3) Construct a MSF F ∗ = (V ∗, E∗)
Initialization: V ∗ = {r1, r2, ..., rm} (roots are in the forest)1 Choose edge of the modified graph ei j with minimal weight such thati ∈ V ∗ and j /∈ V ∗
2 V ∗ = V ∗ ∪ {j}, E∗ = E∗ ∪ {ei ,j}3 If V ∗ 6= V , go to 1
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 37
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Marker selectionMSF-based marker-controlled region growingConcluding discussion
Construction of a Minimum Spanning Forest (MSF)
0
1 1
2
0
0 0
0
0
0 0
3
8
164
8 5
5
14
5
2
9
14
10 1111 12
8
12
15
r2
r1 0
0
3) Construct a MSF F ∗ = (V ∗, E∗)
Initialization: V ∗ = {r1, r2, ..., rm} (roots are in the forest)1 Choose edge of the modified graph ei j with minimal weight such thati ∈ V ∗ and j /∈ V ∗
2 V ∗ = V ∗ ∪ {j}, E∗ = E∗ ∪ {ei ,j}3 If V ∗ 6= V , go to 1
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 37
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Marker selectionMSF-based marker-controlled region growingConcluding discussion
Construction of a Minimum Spanning Forest (MSF)
0
1 1
2
0
0 0
0
0
0 0
3
8
164
8 5
5
14
5
2
9
14
10 1111 12
8
12
15
r2
r1 0
0
3) Construct a MSF F ∗ = (V ∗, E∗)
Initialization: V ∗ = {r1, r2, ..., rm} (roots are in the forest)1 Choose edge of the modified graph ei j with minimal weight such thati ∈ V ∗ and j /∈ V ∗
2 V ∗ = V ∗ ∪ {j}, E∗ = E∗ ∪ {ei ,j}3 If V ∗ 6= V , go to 1
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 37
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Marker selectionMSF-based marker-controlled region growingConcluding discussion
Construction of a Minimum Spanning Forest (MSF)
0
1 1
2
0
0 0
0
0
0 0
3
8
164
8 5
5
14
5
2
9
14
10 1111 12
8
12
15
r2
r1 0
0
3) Construct a MSF F ∗ = (V ∗, E∗)
Initialization: V ∗ = {r1, r2, ..., rm} (roots are in the forest)1 Choose edge of the modified graph ei j with minimal weight such thati ∈ V ∗ and j /∈ V ∗
2 V ∗ = V ∗ ∪ {j}, E∗ = E∗ ∪ {ei ,j}3 If V ∗ 6= V , go to 1
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 37
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Marker selectionMSF-based marker-controlled region growingConcluding discussion
Construction of a Minimum Spanning Forest (MSF)
0
1 1
2
0
0 0
0
0
0 0
3
8
164
8 5
5
14
5
2
9
14
10 1111 12
8
12
15
r2
r1 0
0
3) Construct a MSF F ∗ = (V ∗, E∗)
Initialization: V ∗ = {r1, r2, ..., rm} (roots are in the forest)1 Choose edge of the modified graph ei j with minimal weight such thati ∈ V ∗ and j /∈ V ∗
2 V ∗ = V ∗ ∪ {j}, E∗ = E∗ ∪ {ei ,j}3 If V ∗ 6= V , go to 1
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 37
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Marker selectionMSF-based marker-controlled region growingConcluding discussion
Construction of a Minimum Spanning Forest (MSF)
0
1 1
2
0
0 1
0
0
0 0
3
8
164
8 5
5
14
5
2
9
14
10 1111 12
8
12
15
r2
r1 0
0
3) Construct a MSF F ∗ = (V ∗, E∗)
Initialization: V ∗ = {r1, r2, ..., rm} (roots are in the forest)1 Choose edge of the modified graph ei j with minimal weight such thati ∈ V ∗ and j /∈ V ∗
2 V ∗ = V ∗ ∪ {j}, E∗ = E∗ ∪ {ei ,j}3 If V ∗ 6= V , go to 1
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 37
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Marker selectionMSF-based marker-controlled region growingConcluding discussion
Construction of a Minimum Spanning Forest (MSF)
0
1 1
2
0
0 1
0
2
0 0
3
8
164
8 5
5
14
5
2
9
14
10 1111 12
8
12
15
r2
r1 0
0
3) Construct a MSF F ∗ = (V ∗, E∗)
Initialization: V ∗ = {r1, r2, ..., rm} (roots are in the forest)1 Choose edge of the modified graph ei j with minimal weight such thati ∈ V ∗ and j /∈ V ∗
2 V ∗ = V ∗ ∪ {j}, E∗ = E∗ ∪ {ei ,j}3 If V ∗ 6= V , go to 1
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 37
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Marker selectionMSF-based marker-controlled region growingConcluding discussion
Construction of a Minimum Spanning Forest (MSF)
0
1 1
2
0
1 1
1
2
2 2
3
8
164
8 5
5
14
5
2
9
14
10 1111 12
8
12
15
r2
r1 0
0
3) Construct a MSF F ∗ = (V ∗, E∗)
Initialization: V ∗ = {r1, r2, ..., rm} (roots are in the forest)1 Choose edge of the modified graph ei j with minimal weight such thati ∈ V ∗ and j /∈ V ∗
2 V ∗ = V ∗ ∪ {j}, E∗ = E∗ ∪ {ei ,j}3 If V ∗ 6= V , go to 1
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 37
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Marker selectionMSF-based marker-controlled region growingConcluding discussion
Construction of a Minimum Spanning Forest (MSF)
0
1 1
2
0
1 1
1
2
2 2
3
8
164
8 5
5
14
5
2
9
14
10 1111 12
8
12
15
r2
r1 0
0
3) Construct a MSF F ∗ = (V ∗, E∗)
4) Class of each marker → class of the corresponding region(of all the pixels grown from this marker)
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 37
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Marker selectionMSF-based marker-controlled region growingConcluding discussion
Construction of a Minimum Spanning Forest (MSF)
Pixelwiseclassification map
⇒
Map of107 markers
⇒
MSF-basedclassification map
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 38
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Marker selectionMSF-based marker-controlled region growingConcluding discussion
Construction of a Minimum Spanning Forest (MSF)
Pixelwiseclassification map
⇒
Map of107 markers
⇒
MSF-basedclassification map
If a markeris classified
to the wrong class⇒
The whole region grownfrom this marker
risks to bewrongly classified!
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 38
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Marker selectionMSF-based marker-controlled region growingConcluding discussion
Post-processing
Map of connected components
Pixelwise
classification map
Combination of segmentation and
spectral classification results
1 1 1 1 1 1 1 1 1 2 2 2 1 1 2 2 2 2 1 1 2 2 2 2 1 1 3 3 3 2 1 3 3 3 3 3 3 3 3 3 3 3
Final classification map
MSF-based classification map
Hyperspectral image (B bands)
Classification
map of Construction of minimum spanning
forest
markers Marker selection
Majority voting within connected
components
Segmentation map + classification map
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 39
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Marker selectionMSF-based marker-controlled region growingConcluding discussion
Classification accuracies for the Indian Pines data (%)
SVM ECHOStrategy 2 Strategy 3: Marker-based classif.SVM+ SVM- SVM- MSSC-HSEG MSF MSF+MV MSF
Overall Accuracy 78.17 82.64 90.86 88.41 91.80 92.32Average Accuracy 85.97 83.75 93.96 91.57 94.28 94.22Kappa Coef. κ 75.33 80.38 89.56 86.71 90.64 91.19Corn-no till 78.18 83.45 90.46 90.97 93.21 89.74Corn-min till 69.64 75.13 83.04 69.52 96.56 86.99Corn 91.85 92.39 95.65 95.65 95.65 95.11Soybeans-no till 82.03 90.10 92.06 98.04 93.91 91.84Soybeans-min till 58.95 64.14 84.04 81.97 81.97 89.16Soybeans-clean till 87.94 89.89 95.39 85.99 97.16 97.34Alfalfa 74.36 48.72 92.31 94.87 94.87 94.87Grass/pasture 92.17 94.18 94.41 94.63 94.63 94.63Grass/trees 91.68 96.27 97.56 92.40 97.27 97.85Grass/pasture-mow 100 36.36 100 100 100 100Hay-windrowed 97.72 97.72 99.54 99.77 99.77 99.77Oats 100 100 100 100 100 100Wheat 98.77 98.15 98.15 99.38 99.38 99.38Woods 93.01 94.21 98.63 97.59 99.68 99.44Bldg-Grass-Tree-Dr 61.52 81.52 82.12 68.79 68.79 73.64Stone-steel towers 97.78 97.78 100 95.56 95.56 97.78
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 40
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Marker selectionMSF-based marker-controlled region growingConcluding discussion
Classification of the Pavia image
SVMclassification
OA = 81.01%AA = 88.25%
Strategy 2:SVM + HSEG
OA = 93.85%AA = 97.07%
Strategy 3:SVM-MSF+MV
OA = 91.08%AA = 94.76%
Strategy 3:MSSC - MSF
OA = 97.90%AA = 98.59%
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 41
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Marker selectionMSF-based marker-controlled region growingConcluding discussion
Classification accuracies for the Pavia image (%):
SVM ECHOStrategy 2 Strategy 3: Marker-based classif.
SVM+ SVM+ SVM- SVM- MSSC-Waters. HSEG MSF MSF+MV MSF
Overall Accuracy 81.01 87.58 85.42 93.85 84.14 91.08 97.90Average Accuracy 88.25 92.16 91.31 97.07 92.35 94.76 98.59Kappa Coef. κ 75.86 83.90 81.30 91.89 79.71 88.30 97.18Asphalt 84.93 87.98 93.64 94.77 93.05 93.16 98.00Meadows 70.79 81.64 75.09 89.32 72.30 85.65 96.67Gravel 67.16 76.91 66.12 96.14 89.15 89.15 97.80Trees 97.77 99.31 98.56 98.08 87.02 91.24 98.83Metal sheets 99.46 99.91 99.91 99.82 99.91 99.91 99.91Bare soil 92.83 93.96 97.35 99.76 97.11 99.91 100Bitumen 90.42 92.97 96.23 100 98.57 98.57 99.90Bricks 92.78 97.35 97.92 99.29 95.66 99.05 99.76Shadows 98.11 99.37 96.98 96.48 98.36 96.23 96.48
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 42
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Marker selectionMSF-based marker-controlled region growingConcluding discussion
Outline
1 Introduction
2 Classification using segmentation-derived adaptive neighborhoodsSegmentationSpectral-spatial classificationConcluding discussion
3 Segmentation and classification using automatically selected markersMarker selection
Using probabilistic SVMMultiple classifier approach
MSF-based marker-controlled region growingConcluding discussion
4 Conclusions and perspectives
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 43
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Marker selectionMSF-based marker-controlled region growingConcluding discussion
1 Classification using automatically selected markers:significantly decreases oversegmentationimproves classification accuraciesprovides classification maps with homogeneous regions
2 Marker selection: it is advantageous to useSVM classifierspatial informationmultiple classifier approaches
3 Marker-controlled region growingMSF-based method has proven to be efficient and robust
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 44
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Contributions
Three spectral-spatial classification strategies:1 Using SVM and MRF models
2 Using adaptive neighborhoods derived from segmentationSegmentation techniques working both in the spatial and spectral domain→ good performancesPixelwise classification + majority voting within regions→ simple and efficient technique
3 Using marker-based region growing segmentationAnalyzing probabilistic classification results for marker selectionInterest of using spatial information and multiple classifier approachesfor marker selectionMSF-based marker-controlled region growing→ efficient and robust
Possibilities of high-performance parallel computing using commodityprocessors (GPUs)
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 45
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Perspectives
1 Spectral-spatial image analysisAutomatically select results in segmentation hierarchiesDevelop new similarity measuresPerform segmentation and classification concurrently
2 Further explore parallel strategies using commodity processors
3 Apply and adapt the proposed methods for other types ofdata/applications
medical imaging
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 46
Classification of Hyperspectral Data UsingSpectral-Spatial Approaches
Yuliya Tarabalka
GIPSA-Lab, Grenoble Institute of Technology, FranceDepartment of Electrical and Computer Engineering, University of Iceland, Iceland
Thesis committee
Melba CRAWFORD PresidentJean-Yves TOURNERET OpponentXiuping JIA OpponentAntonio PLAZA OpponentPaolo GAMBA OpponentJohannes R. SVEINSSON ExaminatorJón Atli BENEDIKTSSON Thesis advisorJocelyn CHANUSSOT Thesis advisor
June 14, 2010
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
1. Watershed segmentation for hyperspectral image
From B-band image →1-band segmentation map:
Feature extraction (PCA,ICA,...)?
Vector gradient?
Combine B gradients?
Combine B watershed regions?
Gradient (1 band – 1 band)
Watershed
Hyperspectral image (B bands)
Feature extraction (B bands – 1 band)
Combine gradients (B bands – 1 band)
(1 band – 1 band)
Segmentation map (1 band)
Combine regions (B bands – 1 band)
Y. Tarabalka, J. Chanussot, and J. A. Benediktsson, “Segmentation and classification of hyperspectralimages using watershed transformation,” Pattern Recognition, vol. 43, no. 7, pp. 2367-2379, July2010.
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 48
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
1. Watershed segmentation for hyperspectral image
From B-band image →1-band segmentation map:
Feature extraction (PCA,ICA,...)?
Vector gradient?
Combine B gradients?
Combine B watershed regions?
Gradient (1 band – 1 band)
Watershed
Hyperspectral image (B bands)
Feature extraction (B bands – 1 band)
Combine gradients (B bands – 1 band)
(1 band – 1 band)
Segmentation map (1 band)
Combine regions (B bands – 1 band)
Y. Tarabalka, J. Chanussot, and J. A. Benediktsson, “Segmentation and classification of hyperspectralimages using watershed transformation,” Pattern Recognition, vol. 43, no. 7, pp. 2367-2379, July2010.
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 48
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
1. Watershed segmentation for hyperspectral image
From B-band image →1-band segmentation map:
Feature extraction (PCA,ICA,...)?
Vector gradient?
Combine B gradients?
Combine B watershed regions?
Gradient (B bands – 1 band)
Watershed
Hyperspectral image (B bands)
Feature extraction (B bands – 1 band)
Combine gradients (B bands – 1 band)
(1 band – 1 band)
Segmentation map (1 band)
Combine regions (B bands – 1 band)
Y. Tarabalka, J. Chanussot, and J. A. Benediktsson, “Segmentation and classification of hyperspectralimages using watershed transformation,” Pattern Recognition, vol. 43, no. 7, pp. 2367-2379, July2010.
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 48
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
1. Watershed segmentation for hyperspectral image
From B-band image →1-band segmentation map:
Feature extraction (PCA,ICA,...)?
Vector gradient?
Combine B gradients?
Combine B watershed regions?
Gradient (B bands – B bands)
Watershed
Hyperspectral image (B bands)
Feature extraction (B bands – 1 band)
Combine gradients (B bands – 1 band)
(1 band – 1 band)
Segmentation map (1 band)
Combine regions (B bands – 1 band)
Y. Tarabalka, J. Chanussot, and J. A. Benediktsson, “Segmentation and classification of hyperspectralimages using watershed transformation,” Pattern Recognition, vol. 43, no. 7, pp. 2367-2379, July2010.
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 48
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
1. Watershed segmentation for hyperspectral image
From B-band image →1-band segmentation map:
Feature extraction (PCA,ICA,...)?
Vector gradient?
Combine B gradients?
Combine B watershed regions?
Gradient (B bands – B bands)
Watershed
Hyperspectral image (B bands)
Feature extraction (B bands – 1 band)
Combine gradients
(B bands – B bands)
Segmentation map (1 band)
Combine regions (B bands – 1 band)
Y. Tarabalka, J. Chanussot, and J. A. Benediktsson, “Segmentation and classification of hyperspectralimages using watershed transformation,” Pattern Recognition, vol. 43, no. 7, pp. 2367-2379, July2010.
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 48
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Gradient
Robust Color Morphological Gradient (RCMG):
For each pixel xp, χ = [x1p, x2p, ..., xep] is aset of e vectors within E
Color Morphological Gradient (CMG):
CMGE(xp) = maxi ,j∈χ{‖xip − xjp‖2}
RCMG:xi_maxp , xj_maxp - pixels that define theCMG of xp
RCMGE(xp) = maxi ,j∈[χ−[xi_maxp ,xj_maxp ]]
{‖xip−xjp‖2}
Hyperspectral image X ( B bands)
Gradient
Feature extraction
Combine gradients
Watershed
Spectro-spatial classification
Spectral information
E
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 49
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Gradient
Robust Color Morphological Gradient (RCMG):
For each pixel xp, χ = [x1p, x2p, ..., xep] is aset of e vectors within E
Color Morphological Gradient (CMG):
CMGE(xp) = maxi ,j∈χ{‖xip − xjp‖2}
RCMG:xi_maxp , xj_maxp - pixels that define theCMG of xp
RCMGE(xp) = maxi ,j∈[χ−[xi_maxp ,xj_maxp ]]
{‖xip−xjp‖2}
Hyperspectral image X ( B bands)
Gradient
Feature extraction
Combine gradients
Watershed
Spectro-spatial classification
Spectral information
E
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 49
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Gradient
Robust Color Morphological Gradient (RCMG):
For each pixel xp, χ = [x1p, x2p, ..., xep] is aset of e vectors within E
Color Morphological Gradient (CMG):
CMGE(xp) = maxi ,j∈χ{‖xip − xjp‖2}
RCMG:xi_maxp , xj_maxp - pixels that define theCMG of xp
RCMGE(xp) = maxi ,j∈[χ−[xi_maxp ,xj_maxp ]]
{‖xip−xjp‖2}
Hyperspectral image X ( B bands)
Gradient
Feature extraction
Combine gradients
Watershed
Spectro-spatial classification
Spectral information
E
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 49
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
MC-MSF classification
3-NN
∧
ML
∧
SVM
=
Map of MC markers
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 50
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
MC-MSF classification
3-NN
∧
ML
∧
SVM
=
MC-MSF clas. map
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 50
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
MSSC-MSF classification
WH+MV
∧
EM+MV
∧
RHSEG+MV
=
Map of markers
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 51
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
MSSC-MSF classification
WH+MV
∧
EM+MV
∧
RHSEG+MV
=
MSSC-MSF clas. map
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 51
IntroductionClassification using segmentation-derived adaptive neighborhoods
Segmentation and classification using automatically selected markersConclusions and perspectives
Segmentation and classification of the ROSIS image
SVM classif map13648 regions
RHSEG segm map7575 regions
RHSEG+MV clas map1863 regions
MSSC clas map802 regions
Yuliya Tarabalka ([email protected]) Spectral-Spatial Classification of Hyperspectral Data 52