Critical Class Oriented - PowerPoint Presentation

Slide1

Critical Class Oriented Active Learning for Hyperspectral Image Classification

School of Civil Engineering, Purdue UniversityandLaboratory for Applications of Remote SensingEmail: {wdi@purdue.edu1, mcrawford2}@purdue.eduJuly 28, 2011IEEE International Geoscience and Remote Sensing Symposium

Wei Di and Melba CrawfordSlide2

Outline BackgroundCritical Class Oriented Active Learning(AL) Proposed Methods (SVM-CC, SVM-CCMS)Guided & Active Learning

Critical Class OrientedMargin Sampling Based Experimental Results Conclusions & Future WorkSlide3

I. BACKGROUNDSlide4

MotivationSampling StrategyDL Pool

Training DataTarget HSupervised Classifier

Achieve better

performance

Higher utility, low redundancy

Economically

allocate resources for labeling

Focus on

a specific

task

or requirement

Intelligent sampling strategy Slide5

Active Learning

Query Strategy

D

L

Pool

D

U

Pool

Supervised

Classifier

New

x

L

Output Classifier

Training

x

U

f(

x

u

)

Passive Learning

Active

Learning (AL) -

Iterative learning circle

Uncertainty & Critical ClassSlide6

Introduction Active Learning in remote sensingClassification: Tuia et al. [2009], Patra and Bruzzone [2011] Demir et al. [2011], Di and Crawford [2011], .Segmentation: Jun et al. [2010]

Critical Class oriented Active Learning- Shifting hyperplane by pair-wise SVMIdentify “Difficult” ClassesCategory based query & margin sampling GoalProvide concept level guidance for building training set favoring “difficult” classesSlide7

II. PROPOSED METHODSlide8

Key Idea: Shifting HyperplanePair-wise Class A and B

Class AClass B

Hyperplane

w

Margin

Support Vectors

Hyperplane

Margin

New Samples

Changing

HyperplaneSlide9

Critical Class Identification Query-based Regularizerwk - hyperplane vector by SVM for kth binary class at the t th query.

Accumulated Margin Instability Measure the cumulative changes

Order Statistic



Rank class pairs:



Prob. of the

k

th

class pair at critical level

C

L

:Slide10

Critical Class QueryCritical Class SetCritical Class Pair

Higher probability Critical Class Identification

SVM-CC

Random Query From Critical Class Set

SVM-CC

MS

Query Sample within Critical Class set and closest to margin Critical Class Set

QuerySlide11

III. EXPERIMENTAL RESULTSSlide12

Kennedy Space Center & Botswana Data

AVIRIS hyperspectral data Acquired on March, 1996 176 of total 224 bands Spectral bandwidth 10nm Spatial resolution 18m

Data Description

KSC

BOT

Class Name

No.

Class Name

No.

1

Scrub

761

Water

361

2

Willow Swamp

243

Primary Floodplain

308

3

Cabbage Palm Hammock*

256

Riparian*

303

4

Cabbage Palm/Oak Hammock*

252

Firescar

335

5

Slash Pine*

161

Island Interior

370

6

Oak / Broadleaf hammock*

229

Woodlands*

324

7

Hardwood Swamp*

105

Savanna

342

8

Graminoid

Marsh

431

Short

Mopane

299

9

Salt Marsh

419

Exposed Soils

229

10

Water

927

*

Denotes

the hard classesSlide13

Experimental ResultsIndex

Class-Pair KSC

18



(3,4)

Cabbage

Palm

Hammock;

Cabbage Palm/Oak Hammock

26



(4,6)

Cabbage Palm/Oak Hammock

Oak / Broadleaf hammock

BOT

18

(

3,6)

Riparian

Woodlands

Accumulated

Margin Instability (AMI)

18

18

26

10

th

30

th

AMI as learning process

KSC

BOT

18

26

18Slide14

Experimental ResultsDT

KSC at 600th queryBOT at 400th query

Class Index

CC

CC

MS

SVM

MS

CC

CC

MS

SVM

MS

C1

-0.37

-0.41

-0.27

0

0

0

C2

2.57

3.97

1.95

-0.07

0.82

1.10

C3

-1.87

-0.80

-2.02

11.64

12.04

9.08

C4

7.33

9.24

4.84

-0.24

-0.12

-0.06

C5

1.33

4.00

2.11

-5.00

-1.76

-1.91

C6

2.64

6.34

2.71

3.68

3.07

-0.43

C7

4.79

0.30

7.88

0.67

2.50

2.32

C8

-3.86

-2.97

-3.20

0.33

0.60

0.66

C9

-0.28

-0.72

-0.08

-1.06

-1.68

-0.27

C10

0

0

0

D

T

D

U

Learning Curve

Per-Class

Improvement

vs

RSSlide15

Experimental ResultsKSCC1

C2C3C4

C5

C6

C7

C8

C9

C10

CC

0.60

0.30

0.73

0.83

0.85

0.74

0.45

0.25

0.24

0.27

CC

MS

0.50

0.49

0.8

0.89

0.94

0.78

0.38

0.32

0.21

0.23

SVM

MS

0.35

0.60

0.36

0.58

0.67

0.54

0.45

0.54

0.40

0.44

RS

0.48

0.43

0.43

0.45

0.48

0.45

0.42

0.45

0.49

0.45

BOT

C1

C2

C3

C4

C5

C6

C7

C8

C9

CC

0.03

0.31

0.89

0.11

0.04

0.88

0.13

0.04

0.13

CC

MS

0

0.19

0.95

0.07

0.10

0.95

0.15

0.08

0.04

SVM

MS

0.20

0.23

0.39

0.36

0.22

0.44

0.28

0.18

0.20

RS

0.28

0.29

0.29

0.30

0.28

0.26

0.27

0.25

0.29

Per-Class

Sampling Ratio

Per-class Sampling Ratio

Ratio of Support Vectors

RS

SVM

MS

CC

CC

MS

KSCSlide16

IV. CONCLUSIONS AND FUTURE WORKSlide17

Conclusions & Future WorkConclusionsShifting Hyperplane – Provides valuable information for identifying difficult classes. Critical Class Oriented Margin Sampling – Focuses on difficult classes, as well as informative samples, improve performance in multi-class problem.

Support Vectors - Concentrate on samples likely to be support vectors. Future workInvestigation of feature subspaces for identifying the critical classes.Design proper sample-wise utility score to enhance the category based query.Slide18

IV. CONCLUSIONS AND FUTURE WORKSlide19

Conclusions & Future WorkConclusionsShifting Hyperplane – Provides valuable information for identifying difficult classes. Critical Class Oriented Margin Sampling – Focuses on difficult classes, as well as informative

samples; improves performance in multi-class problem. Support Vectors - Concentrate on samples likely to be support vectors. Future workInvestigation of the feature subspace for identifying the critical classes.Design proper sample-wise utility score to enhance the category based query.Slide20

Thanks very much!Slide21

Critical Class Identification Process

Accumulative Margin Instability Critical Class Probability Heat MapSlide22

Experimental Results(a) KSC: RS (b) KSC: SVMMS(c) KSC: SVM-CC (d) KSC: SVM-CCMS

Per-class Learning PerformanceSlide23

Experimental ResultsRS

SVMMSSVM-CC SVM-CCMS

BOT

Ratio of Support Vectors