igarss04.hyperspectral

7/29/2019 igarss04.hyperspectral

http://slidepdf.com/reader/full/igarss04hyperspectral 1/164

1

United

States

Canada

Australia

Argentina

Minerals

Forests

Agriculture

Grasslands GlaciersDeserts

SaharaAntarctica

Hyperspectral Image Processing and Analysis

for Land Cover Characterization

Dr. Jay Pearlman and Dr. Melba CrawfordSeptember 19, 2004



2

Course Overview

• Introduction to EO Remote Sensing

– Photon end to end flow

– Alternative sensor configurations

• Hyperion Instrument – design and trades

• Sensor Calibration• Data Processing and Corrections

• Classification of Hyperspectral Data

– Input space reduction• Feature selection and extraction

– Output space reduction

• Multiclassifier Systems

• A View Forward



3

EO Remote Sensing Process Overview

0

50

100

150

200

250

400 700 1000 1300 1600 1900 2200 2500

Wavelength (nm)

I r r a d i a n c e ( µ W / c m

2 / n m )

0

0.2

0.4

0.6

0.8

1

400 700 1000 1300 1600 1900 2200 2500

Wavelength (nm)

T r a n s m i t t a n c e

O3

O2H2O

H2O

H2O

CO2

H2O

CO2

H2O

CH4

O2

H2O

Atmospheric

Scattering

O2H2O

0

1

2

3

4

5

6

7

8

400 700 1000 1300 1600 1900 2200 2500

W avelength (nm)

R a d i a n c e ( µ W / c m

2 / n m / s r )

Path Radiance

Solar Irradiance

Sensor Datadownlink

Path Radiance

Surf. Reflectance

0.0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

400.0 700.0 10 00.0 1300.0 1600.0 1900.0 2200.0 2500 .0

Wavelength (nm)

R e f l e c t a n c e

Co ni fer Gr ass

Broad Leaf Sage_Brush

NPV

Calibrated Signal

0

2

4

6

8

1 0

1 2

1 4

1 6

4 0 0 7 0 0 1 0 0 0 1 3 0 0 1 6 0 0 1 9 0 0 2 2 0 0 2 5 0 0

W a v e le n g t h ( n m )


2 / n m / s r )

A t m o s p h e r i c O x y g e n

L e a f W a t e r

A t m o s p h e r i c

C a r b o n D i o x i d e L

e a f C h l o r o p h y l l

L e a f S u g a r

a n d C e ll u l o s e

P h o t o s y n t h e s i s

6 H 2 O + 6 C O 2 + p h o t o n = > C 6 O 6 H 1 2 + 6 O 2

Measured signal

0

100

200

300

400

500

600

700

800

900

1000

400 700 1000 1300 1600 1900 2200 2500

Wavelength (nm)

R e c o r d e d S i g n a l ( D N )

Signal

0.0

50.0

100.0

150.0

200.0

250.0

300.0

350.0

400.0

450.0

500.0

550.0

600.0

650.0

700.0

400.0 700.0 1000.0 1300.0 1600.0 1900.0 2200.0 2500.0

Wavelength(nm)

6 00 K

8 00 K

1 00 0 K

1 20 0 K

Sun

Emitted Radiance

0.00

0.10

0.20

0.30

0.40

0.50

0.60

0.70

0.80

0.90

1.00

400 600 800 1000 1200 1400 1600 1800 2000 2200 2400

Wavelength (nm)


Atmospheric Correction

to Surface Reflectance

Reflectance

At Sensor Radiance

0

2

4

6

8

10

12

14

16

18

20

400 600 800 1000 1200 1400 1600 1800 2000 2200 2400 2600 2800

Wavelength (nm)


2 n m s r )

Upwelling

GroundStation

AtmosphericCorrection

DataExploitation

Atm. Transmittance

CalibrationParameters

Courtesy of R. Green



4

Solar Spectrum and Atmospheric Transmission

0

10

20

30

40

50

60

70

400 700 1000 1300 1600 1900 2200 2500

Wavelength (nm)

R a d i a n c e ( µ W

/ c m

2 / n m / s r )

0.0

0.10.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0

Atmosphere

Solar w/ 1.0 Reflectance

T r a n s mi t t a n c e



5

Solar Irradiance



6

Reflection from a Metal Roof

San Francisco: January 17, 2001: Roof TopHypGain_revA

0

20

40

60

80

100

120

140

160

350 850 1350 1850 2350

Wavelength (nm)

R a

d i a n c e ( W / m 2

/ u m / s r )

Oxygen

Line

water

absorption

band

water

absorption

band

water

absorption

band

water absorption

band

VNIR-SWIRSpectral

Overlap C0

non-utilized

spectral

Reflection of Solar Irradiance of aroof top, a relfective surface.

Some atmospheric features are

indicated.

non-utilized

spectral

channels



7

Radiance To Reflectance Inversion

0

5

10

15

20

25

400 700 1000 1300 1600 1900 2200 2500

Wavelength (nm)

R a d i a n c e ( u W / c m

2 / n m

/ s r ) S pectrum 1 Spectrum 2

Spectrum 3 Spectrum 4

Spectrum 5

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

400 700 1000 1300 1600 1900 2200 2500

Wavelength (nm)


Grass Pavement

Water Roof

Tree

Lt = µF0ρa/π + µF0TdρsTu/π

Lt is the at sensor radiance

µ is the cosine of the solar zenith angle

F0 is the exo atmospheric irradianceρa is the upward reflectance of the atmosphere

Td is the downward transmittance

ρs is the reflectance of the surface

Tu is the upward transmittance of the atmosphere



8

Alternative Sensor Configurations

• Multispectral vs. Hyperspectral

• Whiskbroom vs. Pushbroom



9

Hyperspectral and Multispectral Perspectives

Wavelength

M e a s u r e d R e f l e c t a n c e

Band 3

Band 4

Band 2

Band 1

Wavelength

M e a

s u r e d R e f l e c t a n c e

Wavelength


Spectral characteristic of scene

Hyperspectral Imaging

Hundreds of bands

Multispectral Imaging Few bands



10

Landsat Picks the Windows – But Misses a Lot

L1 L5 L7

L2 L3 L4

Pan



11

Hyperion surface composition map agreeswith known geology of Mt. Fitton in South

Australia

(1) Published Geologic Survey Map

(2) Hyperion three color image (visible) showing regions of interest(3) Hyperion surface composition map using SWIR spectra above

(1) (2) (3)

Hyperion Spectra Reference Spectra

(a) (b)

Hyperion-based apparent

reflectance compares with

library reference spectra

Courtesy of CSIRO, Australia

Hyperion Maps Mt. Fitton Geology



12

Mt. Etna Sample Spectra

veg.

lava

smoke

Hyperion Spectra of Mt Etna Scene July 13th 2001

0

20

40

60

80

100

400 600 800 1000 1200 1400 1600 1800 2000 2200 2400

Wavelength

R a d i a n c e ( W / m 2

- u m - s r )

lava smoke vegetation

1639 nm 2226 nm1234 nm



13

Hyperspectral Sensor Configurations

Similar Similar Bandwidth

Space and AirborneAirborneApplication domain

Harder EasierCalibration

Harder EasierSpatial uniformity

LongerShorter Relative pixel dwell

time

2-dimensional arrayLinear arrayDetectors

PushbroomWhiskbroomCharacteristic



14

Hyperion Configuration

• Pushbroom configuration with common fore optics – 256 field-of-view locations, defines 7.7 km swath width, 30 meters

each

– 6925 frames of data, defines 185 km swath length, 30 meters each.(sample based on length of data), frame rate timed with spacecraft

velocity

Time

6925

frames

field of view256 pixels

spectral range

242 bandsHyperion Data CubeX-swath width

Y-swath length

Z-spectra signature



15

The Hyperion Sensor



16

Hyperion Imaging Spectrometer

Hyperion is a push-broom

imager

• 220 10nm bands covering 400nm -

2500nm

• 6% absolute rad. accuracy

• Swath width of 7.5 km

• IFOV of 42.4 µradian

• GSD of 30 m

• 12-bit image data• Orbit is 705km alt (16 day repeat)



17

Hyperion Technologies

SPECTROMETER

(curved grating)

Pulse Tube

CRYOCOOLER

CALIBRATION

(spectral/

pushbroom)

Reflecting

TELESCOPE

DATARATES



18

Hyperion Sensor Assembly

Baffle /

CALIBRATION

Reflecting

TELESCOPE

Signalprocessor

Spectro-

meter

Pulse Tube

CRYOCOOLER



19

Hyperion Optical System

Three mirror anastigmat (TMA) foreoptic

Grating

Spectrometer FPA

Entrance slit

Focal

Plane

Array

Earth’s

Surface



20

Hyperion Subassemblies

Hyperion

Electronics

Assembly

(HEA)

Cryocooler Electronics

Assembly

(CEA) Hyperion Sensor Assembly (HSA)



21

Hyperion Block Diagram

VNIRGIS

SWIRFPA &FPE

VNIRCCD &

FPE

VNIR ASP

SWIR ASP

Radiator

Cryo-cooler

Cooler Radiator

Cryo-cooler Electronics

(CEA)

(HEA)

Processor,Telemetry

SignalConditioning & Acquisition,

Data Formatter,

& Power Conversion

Hyperion Instrument Electronics

1773Command &Telemetry

SWIRImage DataRS-422

VNIRImage DataRS-422

S/C 28 VDC_1

S/C 28 VDC_1

Heater S/C 28 VDC_ Htr

SWIRGIS

Thermally Connected to S/CThermally Isolated from S/C

Al Honeycomb StructureHeater &

Temp Sensor

Hyperion Sensor

Fore-optics

In-flightCalibration

Source

Aperture Cover

TemperatureSensors (4)

To Spacecraft

Heater Power

Radiator

RS-422



22

Hyperion Key Characteristics

Characteristic Pre-launch On-orbitGSD (m) 29.88 30.38

Swath (km) 7.5 7.75VNIR MTF @ 630nm 0.22-0.28 0.23-0.27

SWIR MTF @ 1650nm 0.25-0.27 0.28

VNIR X-trk Spec. Error 2.8nm@655nm 2.2nm

SWIR X-trk Spec. Error 0.6nm@1700nm 0.58

Abs. Radiometry(1Sigma) <6% 3.40%

VNIR SNR (550-700nm) 144-161 140-190

SWIR SNR (~1225nm) 110 96

SWIR SNR (~2125nm) 40 38

No. of Spectral Channels 220 200 (L1)VNIR Bandwidth (nm) 10.19-10.21 **

SWIR Bandwidth (nm) 10.08-10.09 **

** not measured

O P T I C A L

R A D I O -

M E T R I C

S P E C T R A L



23

Hyperion SNR

0

50

100

150

200

400 700 1000 1300 1600 1900 2200 2500Wavelength (nm)

S N R

Measured

VNIR ModelSWIR Model

Radiometric performance model based

on 60o Solar zenith angle, 30% albedo,

standard scene

550 nm 650 nm 700 nm 1025 nm 1225 nm 1575 nm 2125 nm161 144 147 90 110 89 40

Hyperion Measured SNR



24

Spectrometer Overlays

V

SLevel 1 data: 438-926nm and 892-2406nm

Bands 9-57 and 75 - 225;

SWIR is West of VNIR and rotated CCW by one pixel

S W I R

V N I R

Band Center(nm) FWHM(nm)

50 854.66 11.27

51 864.83 11.2752 875 11.28

53 885.17 11.29

54 895.34 11.3

55 905.51 11.31

56 915.68 11.31

57 925.85 11.31

58 936.02 11.31

59 946.19 11.31

71 852 11.17

72 862.09 11.17

73 872.18 11.17

74 882.27 11.17

75 892.35 11.17

76 902.44 11.17

77 912.53 11.17

78 922.62 11.17

79 932.72 11.1780 942.81 11.17



25

Polarization - VNIR

AVERAGED OVER FOVIn Units of Percent Polarization

Polarizer Angle [degrees]

W a v e l e n g t h

[ n m ]

0 50 100 150 200 250 300 350

500

550

600

650

700

750

800

850

900

950

1000

1050 -5

-4

-3

-2

-1

0

1

2

3

4

5



26

Polarization - SWIR

Polarizer angle [degrees]

S p e c t r a l C h a

n n e l

0 50 100 150 200 250 300

20

40

60

80

100

120

140

160

-5

-4

-3

-2

-1

0

1

2

3

4

5

AVERAGED OVER FOVIn Units of Percent Polarization



27

Echo Variation Map

• Ratio images for each intensity

level were interpolated to create

maps of echo variation across

the entire FPA.

• Based on this map a technique

for echo removal was

implemented.

• The removal algorithm consists

of subtracting a shifted andscaled version of each frame

from itself. Scaling is a

multiplication of the image by

the echo variation map.



28

Profiles Before(red)/After(blue) Echo Removal



29

Hyperion Images Before / After Echo Removal

A l Pi l C ll



30

Anomalous Pixel Collage – Possibilities including striping

Band Swath Pix Band Swath Pix Band Swath Pix Band Swath Pix Band Swath Pix

5 114 12 6 32 114 99 92 168 256

5 141 12 114 33 114 116 127 168 256

6 6 13 114 34 114 116 127 169 12

6 68 14 114 39 177 116 138 169 12

6 172 14 247 40 13 116 138 169 23

7 7 15 114 48 20 119 240 169 23

7 68 16 114 49 20 119 240 169 23

7 179 17 114 50 20 120 240 190 113

7 185 18 114 51 20 120 240 190 113

8 7 19 114 52 13 121 196 190 113

8 12 20 114 52 33 125 54 200 8

8 68 21 114 53 33 125 115 200 88 114 22 114 57 13 125 160 200 8

8 121 23 114 61 93 125 164 201 8

9 6 24 114 72 95 127 40 201 8

9 68 25 114 72 95 127 66 201 8

9 114 26 114 75 2 127 213 203 104

10 6 27 47 94 82 128 30 203 10410 114 27 114 94 82 128 96 203 115

10 131 28 47 94 93 128 126 203 115

10 199 28 114 94 93 165 148 203 116

11 6 29 114 99 81 165 248 222 98

11 114 30 114 99 81 168 245 225 186

11 199 31 114 99 92 168 245

Bad Pix 3 - L1B1; CSIRO – Jupp; Dark Image – Pearlman; PFC - Han



31

Hyperion Nominal Data Modes

Mode Cover

Position

Data collect Comment

Standby Closed None Default mode for active stateDark

calibration

Closed Minimum 100 frames Performed as close as possible to

imaging, before and after

Lamp

calibration

Closed Minimum 100 frames Performed after second dark

calibration; two radiance levels

Solar calibration

Open 37degrees

Minimum 1 second Nominal 1 cube

Performed over North Pole only tokeep cover out of ALI keep-out zone;

yaw maneuvers required

Lunar

calibration

Fully open

(135o)

Minimum: 1 second

Nominal: 1 cube

Performed on dark side of earth; off-

track spacecraft pointing required

Groundcalibration

Fully open(135o)

Minimum 1 second Nominal 1 cube

Ground target selected

Imaging Fully open

(135o)

Minimum 1 second

Nominal 9 cubes

Nominal data collect is equivalent to

Landsat scene, and takes 27 seconds.



32

Performance Tradeoffs

• Spatial Resolution

• Spectral resolution



33

Landsat

SAC-C

MODIS

MODIS

250m

1000m

Coleambally Image Collection – 30m to 1km

C i f H l I



34

Comparison of Hyperspectral Instruments

Parameter Lewis HSI Hyperion

Volume (L x W x H, cm) 43x69x94 39x75x66

Weight (Kg) 39.5 49

Avg Power (W) 66 51

Peak Power (W) 126

Aperture (cm) 12 12

IFOV (mrad) 0.057 0.043

Crosstrack FOV (deg) 0.84 0.63

Wavelength Range (nm) 380 - 2450 400 - 2450

Spectral Resolution (nm) 5.1/6.45 10

No. Spectral Bands 384 220

Digitization 12 12

Frame Rate (Hz) 237 225

Typical SNR 100 - 200 65 - 130

Spectral Calibration (nm) 1 1

Radiometric Calibration <6% <6%

Hyperion Calibration



35

Hyperion CalibrationGround Test and On-Orbit Validation

• System performance assessment strategy

• Present pre-flight and on-orbit measurement techniques

– Absolute Radiometric Calibration

– Spectral Calibration

– Image Quality Characterization



36

Strategy

• Pre-Flight:

– Establish fundamental characteristics of the instrument and

assess requirement compliance

– Establish instrument performance through the build,environmental test and spacecraft integration phases

– Provide solid foundation for on-orbit comparison

• On-orbit:

– Determine on-orbit performance and compare with pre-flight

performance

– Define data collects that can be used to assess identified performance parameters

– Acquire and analyze data collections; and assess accuracy of

technique

– Compare on-orbit results with pre-flight



37

Special Targets for Characterization

Searchlights

-California

Planets

-Venus

90 deg

Yaw

Gas Flares

-Moomba

D t Sit d f Vi i C lib ti



38

Desert Sites used for Vicarious Calibration

Lake Frome RR Valley Arizaro/Barreal Blanco



39

the good

the bad

and

the gotchas

The world of real life operations

Images of Lake Frome



40

Images of Lake Frome

The Salt Surface



41

The Salt Surface

Impact of Moisture on Salt Signature



42

Lake Frome SamplesLab Spectra - Dry with increasing amounts of water

0

0.2

0.4

0.6

0.8

1

1.2

1.4

400 600 800 1000 1200 1400 1600 1800 2000 2200 2400

Wavelength (nm)

R e l a t i v e R e f l e c t a n c e

dry 10ml 15ml 20ml 25ml saturated

Dry

Increasing Water

Content

Approximate lab analysis indicates

changes of spectral signature

based on amount of water added

Impact of Moisture on Salt Signature

The Locusts That Did Not Make It Over



43

The Locusts That Did Not Make It Over

Calibration signature

Or

New culinary delicacy?

Arizaro Dry Lake bed



44

Arizaro Dry Lake bed

Ground Calibration



45

Ground Calibration

Coleambally Collection History



46

Coleambally Collection History

Date(GMT)

Julian

Day(GMT)

Path/

Row

X-track

Position

In-track

Position Comments

23-Dec-00 358 93/84 100 3148 Cum(South);VNIR

1-Jan-01 001 92/84 82 2131 Clear

9-Jan-01 009 93/84 - - UARDRY

17-Jan-01 017 92/84 - - Cloudy

25-Jan-01 025 93/84 95** 2481** High Clouds(North)

2-Feb-01 034 92/84 36 2599 Clear

9-Feb-01 041 93/84 - - Cloudy

18-Feb-01 049 92/84 41 3782 Shadows(East)

25-Feb-01 056 93/84 - -

6-Mar-01 065 92/84 95 3977 Clear

13-Mar-01 072 93/84 104 2926 Clear

22-Mar-01 082 92/84 - - Cloudy

ARIZARO, Argentina



47

g

High altitude (~12,000ft) dry salt

lakebed.

Surface: extremely rough, locally

uniform, and bright.

Some years no rain at all!

It rained the day we arrived

The good news is that conditions

were good for the experiment on the

7th of February 2001.

Calibration & Validation Can Be Fun



48

Calibration & Validation Can Be Fun



49

Let’s Address Calibration

System Performance Verification Strategy



50

y gy

VNIR/SWIR

Spatial

Coregistration

GSD MTF

Geometric

Bandwidth

Center

Wavelength

Spectral

Vicarious

Calibration

Repeatability

Applying

Calibration

Solar

Cal.

Artifact

Removal

Dark

Removal

Defining

Calibration

Absolute

Radiometry

PerformanceVerification

(end-to-end)



51

Absolute Radiometric Calibration

Key Factors Impacting Calibration



52

y p g

No intermediate

properties.

Constant

Spacecraft scans

moon. Relative

moon, sun, sat

angle

noneBased on Lunar

models

Lunar

Calibration

User oriented

effort

Depends on

surface

Atmospheric

effects must be

modeled

Based on ground

truth

measurements

Lake Frome

(vicarious)

Uniform acrossfield-of-view

Constant

Critical tomodeling

intermediate

properties

Diffuse reflectanceof Hyperion cover

Models avail tocommunity

VNIR more

accurate then

SWIR

SolarCalibration

StrengthsSpacecraft

Pointing

Intermediate

Properties

Absolute

Knowledge

On-Orbit Radiometric Calibration



53

• Solar Calibration

– Absolute Comparison: VNIR within 2%, SWIR 5-8% low; SWIR has

larger uncertainty due to solar model and BRDF model of cover surface

– Used to correct for pixel-to-pixel corrections

– Included in repeatability assessment 0.6% for VNIR, 1.6% for SWIR

– Used to define noise level as a function of signal level to determinedSNR

• Lunar Calibration

– Used to eliminate artifacts of diffuser

• Vicarious Calibration and Cross Calibration

– Calibration accuracy in 4-10% range

Solar Calibration Trending



54

History of VNIR Response to Solar Calibration

0.95

0.96

0.97

0.98

0.99

1

1.01

1.02

1.03

1.04

1.05

40 60 80 100 120 140 160

Solar Cal Day of Year 2001

R a t i o t o S o l a r - 0 4 7

, W i t h D i s t . C o r r e c t i o

n

550 nm 651 nm 753 nm 855 nm

History of SWIR Response to Solar Calibration

0.95

0.96

0.97

0.98

0.99

1

1.01

1.02

1.03

1.04

1.05

40 60 80 100 120 140 160

Solar Cal Day of Year 2001

R a t i o t o S o l a r - 0 4 7 , W i t h

D i s t . C o r r e c t i o n

1044 nm 1346 nm 1649 nm 2153 nm

Lunar Calibration Trending



55

Day 038 Day 069 Day 097 Day 128 Day 156

Ratio of Profiles to Lunar 128 for the Different Collects

0.9

1

1.1

1.2

1.3

1.4

1.5

400 900 1400 1900 2400

Wavelength (nm)

R

a t i o

Lunar 38 Lunar 69 Lunar 128 Lunar 97 Lunar 156

Intensity Dependent on

Moon-Sun-Spacecraft

Relative Angles

VNIR and SWIR track

intensity changes

Description of the 90 Degree Yaw Collect



56

Gr o un d T r a c k

1 2256

3 4

123

4

256

4

3

2

1

T=0; Frame#1

Frame#256

4

3

2

1

4

3

2

1

T=0;

Frame#1

Spacecraft rotates such

that the ifov direction is

parallel with the flight path

-90deg

The target spot is

viewed by pixel 256

during frame 1, and

then by pixel 255

during frame 2, and

so on…

255

2 5 5

Frame 1Frame 2

2 5 6

IFOV

Analysis for Yaw Data



57

Level 1R Transposed 45 Deg FlatfieldedFlatfielded(using 2001197 coefficients)

Differences between the sets of correction

coefficients for the entire vnir focal plane are

within ±2.5%

Hyperion Band 20 (VNIR)

Analysis for Yaw Data



58

Transposed 45 Deg FlatfieldedFlatfielded(using 2001197 coefficients)

Hyperion Band 91 (SWIR)



59

Spectral Calibration

Pre-Flight Spectral Calibration



60

Pre Flight Spectral Calibration

Spectral response modeled as agaussian with a center

wavelength and full width half

max

Monochrometer profiles usedto define center wavelength

and bandwidth at discrete

locations

Pre-Flight Spectral Calibration



61

g p

• Center wavelength and bandwidth

– Measured at discrete locations 20 VNIR locations, 25 SWIR locations.

– Used to define the center wavelength and bandwidth for every VNIR andSWIR pixel, 256 field-of-view locations and 242 spectral bands.

• Dispersion (nm/pixel) – Spacing of spectral channels, Hyperion dispersion (~10nm/pixel) closely matches

the bandwidth (10 nm)

• Cross-track spectral difference – Maximum wavelength difference across field-of-view for a single spectral

channel,

– VNIR = 2.6-3.6 nm – SWIR = 0.40- 0.97 nm.

Pre-Flight Calibration Algorithm Development



620.000

0.100

0.200

0.300

0.4000.500

0.600

300 500 700 900 1100

VNIR Measurement

Model calculation (LLS fit with 4 parameters)

R e l a t i v e S e n s o r R e s p o n s e

Reflectance Spectra of Doped Spectralon

S p e c t r a l o n R e f l e c t i v i t y

0.4

0.6

0.8

1.0

1.2

1.4

1.6

200 400 600 800 1000 1200 1400 1600 1800 2000Wavelength (nm)

Erbium (offset by 0.5)

Holmium

Wavelength (nm)

High resolution scans of Holmium and Erbium Oxide doped Spectralon

Process:1. Image contains reference

spectra uniform across the

field of view. (pre-flight:

doped spectralon)

2. High resolution reference

spectra convolved with

sensor spectral response

function

3. Resulting reference spectraaligned with Hyperion

measured spectra to

determine spectral

calibration.

On-Orbit Spectral Calibration



63

Early solar cal, sun is rising through earth’slimb during collect.

View sun, off cover, through atmosphere

Extension of Pre-flight algorithm development

Spectral Calibration



64

800 1000 1200 1400 1600 1800 2000 2200 2400 260010

-1

10 0

10

1

10 2

10 3

Identification of Spectral Features

Wavelength (nm)

S

C

AL

E

D

S

PE

C

T

R

UM

Hyperion Spectra – red

Atmospheric Reference – black

Diffuse Reflectance of cover – blue

Process:

1.) Create Pseudo-Hyperion Spectra

from reference: Modtran-3 for

atmosphere, and Cary 5 & FTS

measurements for diffuse

reflectance of the cover

2.) Correlate Spectral Features:

band number units of Hyperionmax/min correlated with reference

wavelength of max/min

3.) Calculate Band to Wavelength

map: apply low order polynomial

to fit the data over the entire SWIR regime

SWIR

VNIR Spectral calibration based on two lines:

one solar line (520 nm) and an oxygen line (762.5nm)

Spectral-Spatial-Uniformity

Cross Track Sample



65

W a v e l e n

g t h

Requirement Failure by Frown(aka smile)

Depiction- Grids are the detectors

- Spots are the IFOV centers

- Colors are the wavelengths

Failure by Twist Failure by Spectral-IFOV-Shift

VNIR Spectral Variation Across the Field of View

Characteristics of Spectral Calibration



66

SWIR Spectral Variation Across the Field of View

-3.5

-3.0

-2.5

-2.0

-1.5

-1.0

-0.5

0.0

0.5

0 50 100 150 200 250

Pixel Field of View

W

a v e l e n g t h R e l a t i v e t o

F O V 1 2 8 ( n m )

SWIR Band 75: 892.35 nm SWIR Band 150: 1648.96 nm SWIR Band 225: 2405.63 nm

-3.5

-3.0

-2.5

-2.0

-1.5

-1.0

-0.5

0.0

0.5

0 50 100 150 200 250

Pixel Field of View

W a v e l e n g t h R e l a t i v e t o F O V 1 2 8 ( n m )

VNIR Band 10: 447.892 VNIR Band 30: 651.278 VNIR Band 55: 905.511

2.6 to 3.6 nm

.40 to .97 nm

Understanding Frown



67

• Why is this important if it is only a few nm? – Some spectral effects such as movement of the chlorophyll edge

due to stress is of the same order

– Reproducibility of results due to variations in pointing

The Oxygen A Bands at 1 cm-1 & 1 nm Steps



68

Different Answers – Continuum?



69

Oxygen Line from Lake Frome Image



70



71

Image Quality

Image QualityCape Canaveral



72

p

Jan. 13 2001• Modulation Transfer Function

– Pre-flight used knife edge and slit to measure Cross track

direction, Along-track was Cross-Track*2/pi

– On-orbit used Ice Shelf & Lunar Limb (knife edge) and bridge (slit) to measure Cross-Track and Along-Track

directly.

• Co-registration of VNIR and SWIR

– Pre-flight used test bed to project a slit with a broadspectrum at multiple locations

– On-orbit used combination of edges (Lunar, Ross), point

sources (clouds, flares), ground control points

• Ground Sample Distance – Pre-flight measured IFOV using test bed

– On-orbit triangulated marked features in well mapped scene

MTF Approach



73

• Calculate cross-track and in-track MTF using a stepresponse and impulse response example

• Results of on-orbit analysis give good agreement with the

pre-launch laboratory measurements

VNIR Imaging of “Starburst Pattern”



74

Pattern used to validate

optics performance

and level 1 processing

Level 0 data Level 1 data

MTF Example: Cross-track Bridge



75

Dec 24, 2000. Bridge is the Mid-bay bridge near Destin, Florida.

Bridge width (13.02 m) acquired and utilized in the MTF processing.

Bridge angle small, every 5th line used to develop high resolution bridge image.MTF result at Nyquist is between 0.39 to 0.42; pre-flight measurement was 0.42.

-10 -5 0 5 100

50

100

150

200

250

Interlaced bridge images and curve-fit from band 30

line 220line 225line 230line 235line 240Curve-Fit

Cross-track pixel

1 0 * R a d i a n c e

Image Quality Example



76

Vertical and Horizontal MTF can be calculated from diagonal edge

MTF Calculation Process Summary



77

• MTF measurement on-orbit requires collection of scenes

– step response: Ross Ice Shelf, Lunar Collect

– impulse response: Bridge (Eglin, Cape Canaveral)

• Data Processing steps (simplified):

1.) Define Edge Spread Function (ESF): Interlace adjacent lines from an

object that is at a slight angle to the spacecraft motion

- allows over sampling (sub-pixel) sampling

2.) Calculate Line Spread Function (LSF):

- edge technique: calculate an error function curve-fit to the ESF to derive the

LSF, OR take the band-limited derivative of the ESF with a Tukey window.

- slit technique: LSF is obtained from interlacing adjacent lines and de-

convolving the profile with the slit width (in the frequency domain)

3.) MTF is the Fourier Transform of the LSF

What’s Next



78

• We have looked at instrument characteristics and calibration

• Now what do we do to “correct” the data?

– Radiometric

– Geometric – Atmospheric

– Uniformity (striping)

– Other nuances

• How well can all this work?

Hyperion Data Flow



79

Science Data: Level 0 or

Level 1 (radiometrically

corrected) data products

with VNIR and SWIR data

frames combined.

Includes solar, lunar

calibrations, earth

images, dark and light

calibrations

Metadata: Data about thescience data. Information

to support higher level

processing, e.g., pre-flight

characterization data

Ancillary Data:

Supporting data derived

from spacecraft telemetry

during image collection

Ancillary data in

engineering units

Final product:level 1 data,

metadata file

attached

L0 Science data

Level 1

Science data

L0 Proc.

METADATA

Level 1 Data Processing Flow



80

Background removal:

>Subtract interpolated dark

file from echo and smear

corrected image file

Echo removal

> Remove echo from smear

corrected files

> Both dark and image files

> Log file generated (MD8)

Average dark files

> Average frames of echo

and smear corrected pre-

and post-image dark files.

Provide averaged pre-

image (MD3A) and post-

image (MD3B) dark files> Compute average value

of corrected pre- and post-

image dark files. Log file

generated reporting

average values (MD5A for

pre-image and 5B for post-

image)

Smear correction

> Apply smear correction to SWIR data

> Both dark and image files


.L1_A

data

Step 0

Step 1

Step 2

Apply calibration:

> Multiply dark subtracted

image by lab gain file (MD7) to

obtain radiometrically corrected

(Level 1) data

>Multiply VNIR radiance by 40

>Multiply SWIR radiance by 80

>Log file generated (MD2)

Fixstripes :

> Repair known bad pixels and

output level 1_A file in signed

integer format (.L1_A)


Step 6

Step 5

Step 4

Step 3

Flag pixels >4095:

> Pre- & post-image darks> Image

> Output to log file (MD15)

Dark cal interpolation:

> Linearly interpolate

between preimgdk .avg (MD3A)

and postimgdk .avg files (MD3B)

QA .L1_A image:

> Display image in ENVI

> Complete QA form during

evaluation (MD11)

> Add center wavelengths to

ENVI header (. hdr ) fi le (MD12)

Step 7

L0 data

Hyperion Processing: One Step Further



81

User Receive

‘L1A’

VNIR bands 1:70

divide by 40.0

SWIR bands 71:242

divide by 80.0

Untar tapePerform Quick

Checks and Subset

the Data

L1A subset

Return to Absolute

Radiance, float

Align VNIR to SWIR

Using swir_to_vnir.pts

Or -1 cross track

And 0-1 along track 0-1

Recombine Datafiles

‘L1P’

Subset data file, absoluteradiance in float, SWIR

aligned to VNIR, in a

single file SWIR aligned to

VNIR, absolute

radiance, float

Data Processing Levels



82

Level 2G:

Geometric and

VNIR/SWIR merge

Level 1A:

delivered data

Level 1P:

Radiometric and

VNIR/SWIR overlay

Data Evolution - Processing L1A –> L2G



83

Level 1A

Level 1P

Level 2G

Level 1A:

delivered data

Level 1P:

Radiometric andVNIR/SWIR overlay

Level 2G: Geometric

and VNIR/SWIR mergeVNIR/SWIR merge:

Bands 9-55 & 77-220

Geo-Correction: L1A Æ L2G



84

Methods / Geo-correction



85

A time series of polynomials allowing for geo-correction were developed by

simultaneously fitting Ground Control Points (GCPs) collected in multiple

images (date and sensor) using the MOSMOD module of microBRIAN.

These polynomials allow the images to be resampled to geographiccoordinates.

Fifty GCPs were collected between the base Aerial photography, an ETM

image acquired 02-January-2001 (ETMjan), Hyperion VNIR, acquired 02-January-2001, 03-February-2001, 07-March-2001 (HypVNIRjan,

HypVNIRfeb, HypVNIRmar), and Hyperion SWIR for those same dates

(HypSWIRjan, HypSWIRfeb, HypSWIRmar).

MOSMOD both optimizes the images to geographic coordinates and the

images internal registration.

GIS Data



86

High resolution (2m) digital aerialphotographs acquired January 2001

used for creating positionally accurate

field and rice bay GIS datasets.

Over 466km of linear road network

and 129 well-defined points digitisedwith a Differential Global Positioning

System (DGPS). These datasets can

be used for geo-referencing Hyperion

time series

Coleambally Geo - Correction

Si H i ‘i ’ (VNIR d SWIR f 02 J 03 F b



87

• Six Hyperion ‘images’ (VNIR and SWIR for 02 Jan; 03 Feb;and 07 Mar 2001); GCPs collected in each Hyperion ‘image’

and 2m air photos

• Prior to fitting, outlier GCPs were identified and redefined; a

linear polynomial was selected and a transitive chain method

(MOSMOD) was used

Pixel Size

Avg X = 30.77m; Avg RMS error = 2.52 m

Avg Y = 30.49m; Avg RMS error = 6.87 mRegistration

Cross-track 12.9m (SD 0.6m)

Along-track 11.6m (SD 2.1m)

Presented at IGARSS 01

Preliminary Results / Geo-correction



88



Aligning the VNIR and SWIR

VNIR d SWIR i d d t t t



90

• VNIR and SWIR are independent spectrometers

and focal plane arrays

• For the user to align the SWIR to the VNIR apply- constant cross track shift of –1 pixel

- linear along track shift of 0 pixels at field-of-view 0 and 1 pixel at

field of view 256.

VNIR SWIR

SWIR

VNIR

Reflectance ProcessingHyperion Band 223 ( 2385 nm )

Input Radiance Image



91

2 . A t m

o s p h

e r i c C o r r e

c t i o n

( F L A A

S H - b a

s e d )

3. MNF Smoothing

Output Reflectance Image

1. De-striping

Should Atmospheric Correction be Performed?



92

• Hyperspectral remote sensing has effects of atmosphere aswell as earth materials

• The result is easily recognized and its worst effects can beavoided

• Should you atmospherically correct? – Some applications do not need it (classification, MNF, CVA,

exploration)

– Some can benefit from it (standardization) – Some need it (modeling)

• The data are there to use – do not be put off nor wait until

all the “problems” have been solved!

A Simple Atmospheric Model



93

1

1

:

:

t t g p

t

t t

t t

L E T L s

odelling L

Atmospheric Correction L

ρ

π ρ

ρ

ρ

= +−

→

→

Irradiance

Transmittance

Path Radiance

Atmosphere

Reflectance

Ground Reflectance

At-Sensor

Radiance

Not so frightening after all?! (DLBJ)



94

Chlorophyll

Liquid Water

Clay

Red Edge

Atmospheric Correction Codes

Hatch FLAASH ACORN



95

Hatch, FLAASH, ACORN

Atmospheric Correction Codes

Hatch FLAASH ACORN



96

Hatch, FLAASH, ACORN

FLAASH – ASD Comparison

Soil1 Means



97

Soil1 Means

0.00

0.05

0.10

0.15

0.20

0.25

0.30

350.00 850.00 1350.00 1850.00 2350.00

Wavelength (nm)


Soil1.mn

f_Soil1.mn Stubble1 Means

0.00

0.05

0.10

0.15

0.20

0.25

0.30

0.35

0.40

0.45

0.50

350.00 850.00 1350.00 1850.00 2350.00

Wavelength (nm)


Stubb1.mn

f_Stubb1.mn

Soil

Stubble

Atmospheric Avoidance



98

“Stable” set of bands can be considered if application permits

Bands Wavelength (nm)

10-57 448-92681-97 953-1114

101-119 1155-1336

134-164 1488-1790

182-221 1972-2365

Yerington, NV: pixel 62, 35



990.4 0.9 1.4 1.9 2.4Wavelength (µm)

R e f l e

c t a n c e

HATCHATREM

.

10

12

14

16

n m / s r )

Cirrus Cloud Detection Over Mojave Desert



100

0

2

4

6

8

400 700 1000 1300 1600 1900 2200 2500

Wavelength (nm)

R a d i a n c e ( µ W / c m ^ 2 / n

0 m altitude

3000 m altitude

10000 m altitude

Visible Image Image from 1380 nm

Accounting For Water Vapor

18 0

20.0



101

0.0

2.0

4.0

6.0

8.0

10.0

12.0

14.0

16.0

18.0

400.0 700.0 1000.0 1300.0 1600.0 1900.0 2200.0 2500.0

Wavelength (nm)

R a d i a n c

e ( µ W / c m ^ 2 / n m / s r )

0.00 mm PW

0.10 mm PW

1.00 mm PW

10.0 mm PW

50.0 mm PW

ModeledMeasured

-1

0

1

2

3

4

5

6

7

8

9

1 0

8 40 86 0 8 80 900 92 0 9 40 9 60 9 8 0 1 000 1 02 0 1 0 40 1 06 0

W aveleng th (nm )


2 / n m / s r )

M easu red

M odele d

R e s id u a l

15 .92 p rec ip i tab le mm

Water Vapor Parameter map

Yerington, NV: Water Vapor Images



102

ATREM HATCH

0.68 cm 1.45 cm 0.60 cm 1.10 cm

Stripes in the Image



103

• Stripes may not

impact analyses

• Three spatial

frequency scales

occur:

•High

•Medium

•Low

• Is there a

temporal effect?

RGB (42, 21,15)

Approaches for De-Striping



104

1. Set a column mean (intensity) equal to the global mean

2. Set the column mean and standard deviation equal to theglobal mean and standard deviation

3. Use a locally-based mean and standard deviation in spatialdimensions

4. Use a locally-based mean and standard deviation in spatialand spectral dimensions

Effects of ‘Global’ and ‘Local’ De-Striping



105

•Typical spatial scales are pixel, surface features and swath

•‘Global’ removes streaks and low-frequency effects (but‘Global’ can bias the results) – it depends on the surface

cover characteristics along the column

•‘Local’ removes spikes – especially in the VNIR but leaves

in the low frequency effects such as the smile radiance effect.

•VNIR and SWIR noise are different. Should be treated

separately.

Striping

VNIR SWIR



106

De-streaking Approach

( , )ij ijm s are column averages for sample i band j



107

( , )

( , )

ij ij

ij ij

ijk ij ijk ij

ij

ij ij ij ij ij

ij

m s are column averages for sample i band j

m s are reference values for them

x x sample i band j line k

sm m

s

α β

α β α

→ +

= = −

Selection of reference values determines whether

the de-striping is local or global

Global De-striping MNF1 and 15

Original Data De-Streaked Data



108

MNF 1 Radiance MNF 15 Radiance MNF 15 RadianceMNF 1 Radiance

Effects of ‘Global’ and ‘Local’ De-Striping

Change Analysis•Global De-streaked minus Original Radiance Local De-streaked minus Original Radiance



109

Global De streaked minus Original Radiance Local De streaked minus Original Radiance

•MNF Band 1 MNF Band 5 MNF Band 1 MNF Band 5

•Global De-streaking removes ‘smile’ effect and streaks Local de-streaking removes stripes but leaves

•But also alters mid frequency spatial effects in the data the ‘smile’ effect in



110

Other Thoughts

Other Effects - Directional Illumination



111Midday

Early Morning

Through the Glass Darkly – or “no free lunch”



112

• Hyperspectral data is fighting for photons• Small pixels & many narrow bands lead to lower SNR

• The technology is good but there is a limit

• Landsat FWHM ranges from 70 in VNIR to 300-400 in the

SWIR

• Hyperion Bands have an FWHM and step of 10 nm

Sensor Characteristics



113

SAC-C

CWL

(nm)

FWHM

(nm) ETM

CWL

(nm)

FWHM

(nm) ALI CWL (nm)

FWHM

(nm)

MODIS

Land

CWL

(nm)

FWHM

(nm)

Band 1p 441.6 18.9

Band 1 490.5 19.3 Band 1 478.7 67.2 Band 1p 484.8 52.6 Band 3 465.7 17.6

Band 2 548.8 19.2 Band 2 561.0 77.6 Band 2 567.2 69.7 Band 4 553.7 19.7



Band 4p 865.6 44.0

Band 5p 1244.4 88.1 Band 5 1242.3 24.7


Band 7 2208.1 251.3 Band 7 2225.7 272.5 Band 7 2114.2 52.2

Pan 719.9 319.5 Pan 591.6 144.3

SAC-C ETM+ ALI MODIS

SAC-C VNIR Bands vs ALI, ETM+

ALI ETM



114

ALI ETM+

Red: SAC-C

Blue: Other Instrument

Binned Hyperion vs SAC-C



115

Hyperspectral ImagingApplications & Benefits



Existing Satellite

Capabilities (SPOT,

Application LandSat) Hyperion Capability Perceived Benefits

Mining/Geology Land cover classification Detailed mineral Accurate remote mineral

mapping exploration

Forestry Land cover classification Species ID Forest health/infestations

Detail stand mapping Forest productivity/yieldFoliar chemistry analysis

Tree stress Forest inventory/harvest

planning

Agriculture Land cover classification Crop differentiation Yield prediction/commodities

Limited crop mapping Crop stress crop health/vigor Soil mapping

Environmental Resource meeting Chemical/mineral Contaminant Mapping

Management Land use monitoring mapping & analysis Vegetation Stress

Hyperspectral Image Provides Forestry Detail

LandSat Analysis Hyperspectral Analysis



Open field

No Data

Red Maple

Red Oak

MixedHardwood

Legend

Hardwood/Conifer Mix

White Pine

Hemlock/

Hardwood Mix

Mixed Conifer

Norway Spruce

Red Pine

Spruce Swamp

Hardwood Bog

Analysis by Mary Martin University of New Hampshire

N o D a t a

H a r d w o o d

S o f t w o o d

G r a s s / F i e ld s

L e g e n d

Hyperspectral Image Provides Geological Data

GEOTHERMAL AREA(no specific mineral information)

CALCITE(gold bearing quartz)



Red: Hydro-thermal alterationMULTISPECTRAL ANALYSIS

Red: Calcite

Blue and Green: MuscoviteHYPERSPECTRAL ANALYSIS Analysis courtesy AIG Limited Liability Company

Analysis of Red Edge Shift- sample process



119



120

Now, Let’s Classify the Data!



121

Pixel-Based Supervised Classification Methods

for Hyperspectral Data

Important Issues for Classification of Hyperspectral

Data

• Impact of noise and calibration

L SNR h di l i l d



122

– Lower SNR than corresponding multispectral data

– Striping in data acquired by pushbroom sensors

– Data susceptible to atmospheric artifacts and sensor anomalies

• Large input space

– Enormous number of parameters to estimate – Sparse data in hyperspectral domain

– Adjacent bands are often highly correlated

– Quantity of training data typically limited

• Potentially large output space

– Possible overlapping spectral signatures

Approaches to Resolve Issues

• High dimensional input spaces



123

– Simplify or reduce dimension of input space via extraction

• Subset feature selection

• Projection of original features on a lower dimensional space

– Regularize covariance matrix of input space

– Utilize unlabelled samples via semi-supervised learning

– Develop ensembles of classifiers

• High dimensional output spaces

– Output space decomposition

• Utilization of nonparametric classifiers



124

Dealing with High Dimensional Input Spaces

Feature Selection/Extraction

G l



125

• Goals

– Reduce no. observations required to train the classifier

– Reduce computational complexity

– Improve classification accuracy

• Desired Properties of Extractors – Class dependent

– Exploit band ordering

– Transformations should maximize discrimination among classes

• Selection vs extraction



126

• Selection vs extraction

– Selection of subset of original features

• Less redundancy, but potentially some information loss

• Better interpretability (domain knowledge)

• Possibly improved generalization

– Extraction

• Typically computationally superior to selection

• No loss of information as all features retained

• Resultant features unrelated to physical phenomena

Approaches to Feature Extraction

• Principal Component Transform [Anderson (1984)] – Orthogonal linear combinations of the p original bands with maximumvariance. At pixel x,



127

p ,

– Formulated as an optimization problem

– Solution is a subset of successively ordered eigenvalues of thecovariance matrix Σ; are corresponding eigenvectors

( ) ( )

subject to

1, 1,....

0 , 1,...

t

i i

t

i i

t t

i j

Max a a

a a i p

E a x x a i j j i

= =

= ∀ ≠ =

Σ

Z Ζ

( ) ( ) 1t

i x Z x , i ,...p= = Z

*

ia

– Characteristics of Principal Component Transform



128

Characteristics of Principal Component Transform

• Developed to maintain variance of data in smaller set of orthogonal inputs

• Order of components not necessarily related to data quality• Components have no physical relationship to the targets

• PCT data are image content dependent (poor generalization)

• Does not exploit the adjacency of correlated bands

• Maximum variability has no inherent relationship to criterion for

classification

• Sensitive to outliers

Principal Components of AVIRIS Data



129

• Segmented Principal Component Transformations [Jia andRichards (1999)]



130

– Detect boundaries in correlation matrix,

compute PCT’s and select subset. Further prune using Bhattacharya distance

– Characteristics of SPCT

• Restricted to two-class problems• Utilizes adjacent band correlation structure

• Not related to discrimination measures

• Sensitive to outliers

• Maximum Noise Fraction Transform [Green, Berman, Switzer,

and Craig (1988)]



131

g ( )]

– Uncorrelated linear combinations of bands that maximize “noise

fraction” for each band.

– Define “noise fraction” at each pixel x

( ) ( ) ( ) ( )1

where

known

t

i

Z S N Z N

x Z x , i ,...p x x

,

Ν

Σ Σ Σ Σ Σ

= = = +

= +

Z S

( ) ( ) ( ) 1i i i NF x Var N x Var Z x , i ,...p= =

– Formulated as an optimization problem similar to PCT, except that



132

orthogonal linear combinations are sought to maximize the Noise

Fraction

– Optimal weights:

( )

( ) ( )

subject to

1 1,...

0 , 1,...

t t

i i N i i Z i

t

i Z i

t t

i j

Max NF x a a a a

a a i p

E a Z x Z x a i j j i

= Σ Σ

Σ = =

= ∀ ≠ =

* 1eignvectors of i N

a −= Σ Σ

– Characteristics of MNF Transform



133

• Developed to filter noisy bands, then transform data back to original domain

• Ordering related to quality of tranformed data• Invariant under data rescaling

• Does not exploit the adjacency of correlated bands or class specific

information

• Resultant bands of transformed data may be related subjectively to phenomena, but are not related to discrimination between classes

MNF Bands for AVIRIS Data



134

• Fisher’s Linear Discriminant [Anderson (1984)]



135

– Seeks projection matrix A* to maximize separation between classes

, 1,..i i C ω Ω = =

( )( ) ( )( )

where

generalized eigenvectors of (dim 1)

t

t

t t

x X

Max

X x x

C

ω

ω ω ω ω ω

ω ω

µ µ µ µ µ µ ∈Ω ∈Ω ∈

= − − = − −

= ≤ −

∑ ∑ ∑-1

A BA

A WA

B W

A* W B

– Characteristics of Fisher’s Linear Discriminant



136

• Requires a trained classifier

• Poor discrimination well between classes with similar means• Classes with “different” variance dominate the combination

• Ignores band ordering

• Problematic for hyperspectral data with small training sample due to

required matrix inversion

• Decision Boundary Feature Extraction [Lee and Landgrebe(1997)]– Computes a decision boundary that determines the projection of the



137

Computes a decision boundary that determines the projection of thehyperspectral space for a two-class problem

• C -class problem solved using wtd sum of 2-class decision boundaryfeature matrices

– Requires trained classifier

– Ignores band ordering

• Projection Pursuit [Jimenez and Landgrebe (1999)] – Compute projection for two classes based on Bhattacharya distance

– Subsets constrained to be smaller groups of adjacent bands – One set of features selected for all classes

– C-class problem uses sum of discriminatory power for pairs(problematic for large no. of classes)

• Best Bases Band Aggregation [Kumar et al. (2002); Morgan etal. (2004)]

Adj b d d b d l i ( d f



138

– Adjacent bands merged based on correlation measure (or product of

correlation and Fisher discriminant) – Implemented in Top-Down and Bottom-Up approaches

– Characteristics

• Exploits band adjacent correlations, maintains unique individual bands

• Features are class specific

• Affects signal to noise ratio in a non-uniform way

• Mitigates impact of small training sample problem

• Polyline Feature Extraction [Henneguelle et al. (2003)] – Approximate mean spectrum via piece-wise polynomials (usually linear).



139

– Select break points via top-down or bottom-up methods and determine parameter values (Linear: slope, midpoint of segment)

– Parameters of piece-wise functions become input features to classifier

– Characteristics

E l it b d dj



140

• Exploits band adjacency

• Features are class dependent• Computationally efficient

• Provides approximation to derivatives

• Features robust for generalization

• Sensitive to stopping criterion (no. of bands vs error in

approximation)

• Discrete Fourier Transform

D ib b d i t i t i f ti



141

– Describes band sequence via trigonometric functions

– Subset of modes selected to represent features – Characteristics

• Multiscale representation of spectral signature

• Difficult to determine optimal modes to represent original problem

• Computation unrelated to class discrimination

• Modes may contain domain knowledge

• Discrete Wavelet Transform [Bruce et al. (2001, 2002);

Kaewpijit et al (2003)]



142

Kaewpijit et al. (2003)]

– Multiscale method that accounts for both frequency and position – Describes band sequence via discrete wavelet transform

– Implemented in conjunction with a selection method

– Characteristics• Provides robust features

• No. of inputs must be a power of 2

– Accomplished via padding or interpolation

• Criterion unrelated to class discrimination

Approaches to Feature Selection

• Selection techniques can be implemented individually, or inconjunction with extraction methods



143

j

• Methods that guarantee optimal solutions – Exhaustive search

– Branch & bound (only if objective function is monotonic)

• Heuristic methods – Characteristics

• Typically produce sub-optimal solutions• Developed to reduce computational complexity

• Traditionally implemented via sequential forward selection or backwardelimination

Heuristic Feature Selection Methods

• Sequential Floating Search Methods [Pudil et al. (1994);Serpico and Bruzzone (2001)]



144

– Evaluate a number of features for elimination/forward selection for each feature (group) added/eliminated

• Tabu Search [(Zhang and Sun (2002); Korycinski et al.

(2003)] – Developed for solution of combinatorial optimization problems

– Non-greedy heuristic that adaptively and reactively adjusts searchspace, flexible in number of features evaluated at each iteration

• Various simulated annealing and genetic algorithms [Siedleckiand Sklansky (1988, 1989)]

Regularization of Covariance Matrix

• Methods to regularize estimated covariance matrix



145

– Pseudo-inverse [Fukanaga (1990); Skurichina and Duin (1996);

Raudys and Duin (1998)]

• Poor performance when ratio of amount of training data to input

dimension is small

• Exhibits “peaking” effect at

– Shrinking and pooling covariance matrices [Tadjudin and Landgrebe

(1999)]

• Reduce variance of estimates, but introduce bias

2 X d =

Incorporation of Unlabeled Samples

• Semi-supervised learning

A gment e isting training data ith nlabeled samples [Jeon and



146

– Augment existing training data with unlabeled samples [Jeon and

Landgrebe (1999); Tadjudin and Landgrebe (2000); Jackson andLandgrebe (2001)]

• Classify unlabeled observations and update estimates, usually via EM

algorithm

• Alternatively perturb sample data

– Effective for generalization and knowledge re-use

– Sensitive to original training samples, impacted by outliers



147

Dealing with Large Output Spaces

Multiclassifier Systems

• Multiclassifier Systems – Ensembles of base classifiers that are learned individually to produce anaggregated predictor

– CharacteristicsC l ifi f k



148

• Component classifiers often weak

• Decision boundaries for individual classifiers typically simple• Generalization typically superior, if classifiers are diverse

• Goals – Improve classification accuracy and generalization

– Reduce complexity and improve interpretation – Enhance transferability of classifiers to new problems or beyond spatial

extent of training data – Mitigate the impact of sensor artifacts

– Improved utilization of small quantity of training data• Challenges

– Achieving highly diverse, but relevant classifiers – Maintaining interpretability

Multiclassifiers from Input Space Decomposition

• Selecting sub-samples of original data, creating classifiers,and developing a classifier for each sample. Combine

lt f i di id l l ifi



149

results from individual classifiers – Perform poorly for extremely small sample sizes as ensemble

methods cannot overcome lack of diversity

• Bagging [Brieman (1996)]

– Create multiple samples with replacement; develop individualclassifiers

• Arcing – Adaptively reweighting and combining (e.g. boosting) [Freund and

Schapire (1996); Dietterich (2000)]

– Simple random sub-sampling [Ho (1998); Skurichina and Duin(2002)] (See previous slides for details)

• Random subspace selection methods [Ho (1998); Skurichinaand Duin (2002)] – Utilized in conjunction with multiclassifier systems



150

• Randomly sample input features for each classifier • Combine outputs of multiple classifiers via voting or maximizingaposteriori probabilities

– Characteristics

• Provides diversity and robust classifiers with good generalization

• Reduces band redundancy, but gnores band adjacency, unless implementedwith band aggregation (which reduces diversity)

• Mitigates impact of small sample size problems and stabilizes parameter estimates

• Computationally expensive

• Does not provide domain knowledge on feature characteristics

• Mitigates the impact of anomalous sensor behavior (striping)

Multiclassifier Systems: Output SpaceDecomposition

• C one-vs-rest problems [Anand et al. (1995)] – Classify individual classes vs the mixture of the “rest” – Can result unbalanced priors for Bayesian methods if sample sizes

for some classes extremely small



151

for some classes extremely small

– Does not exploit class affinity, so convergence can be very slow

• Pairwise classifiers [Crawford et al.(1999);

Kumar et al. (2001); Furnkranz (2002)]

– Selects best set of features for class

pairs, performs classification, then combines

results for C-class problem via voting or

maximum aposteriori rule – Can incorporate pairwise feature extraction

– Yields classifiers

– Potential coupling of outputs

2

C

• Error correcting output codes [Dietterich and Bakiri (1995);Rajan and Ghosh (2004)] – C -class problem encoded as binary classifiers which

bi d b ti

C C >>



152

are combined by voting

– Each class assigned according to code book

– Observations labeled as class with code closest to code formed byoutputs of classifiers

– Characteristics• Implemented with user-selected classifiers

• Feature selection/extraction tunable to

pairwise classifiers

• Output dependent on classifier separation• Reduces impact of small sample sizes

• Does not exploit natural class affinities

C C ×

C

• Support vector machines [Vapnik (1995); Hsu and Lin (2002)]

– Maximize the margin between two class samples instead of accuracies

Characteristics



153

– Characteristics

• Formulated as an optimization problem for a binary classifier

• Nonparametric

• Potentially high dimensional decision boundary

• Classification accuracies dependenton typically slow tuning

• Tends not to overtrain, which leads to

high classification accuracies and

good generalization

Support Vector Machine

Margin

Support Vectors

wx+b=0

wx+b<=-1

wx+b>=1

x2

Support Vector Machine

Margin

Support Vectors

wx+b=0

wx+b<=-1

wx+b>=1

x2

• Hierarchical Decomposition Methods

– Characteristics



154

• Adopt a sequential “divide and conquer” approach involvingdecomposition based on input or output spaces

• Maintain natural groupings with useful domain knowlege

– Hierarchical Decision Trees

• Tree constructed sequentially based on series of binary questions and

splitting rule which maximizes decrease in impurity of parent and

child nodes

– CART: Best split based on linear combinations of input features; Gini

impurity index, post pruning [Brieman et al. (1984)] – C4.5: Impurity index based on entropy [Quinlan (1986)]

– Binary Hierarchical Classifier [Kumar et al. (2002); Morgan et al.(2004)]

• Output space decomposition approach

– C leaf nodes, C-1 internal nodes



155

• Top down tree constructed via deterministic simulated annealing• Feature extraction/selection tunable to each internal node

• Classifier specific to each internal node (Fisher discriminant,

SVM)

• Mitigates small sample problems

• Exploits natural class affinities

4A+3 C Ω = 4A+2 3Ω =

2A+1 3,C Ω =

2A 6Ω =

12 5Ω = 13 9Ω =

5 1C Ω = − 4 1Ω =

6 5,9Ω =

3 2, , 2,C C Ω = −K

21, 1C Ω = −

1

Leaf node

2

C-11

3

6

95

2A+1

C 3

1 1, ,C Ω = K

A

A 3,6,C Ω =

6

Internal node



156

Let’s look at an example

Study Site: Okavango Delta, Botswana



157Republic of Botswana

Classification Results: Okavango Delta



158

Classification scheme consists

of 14 landcover classes selectedusing IKONOS imagery,

aerial photography, and field

campaigns

Hyperion Hyperspectral Data Inputs

• Hyperion hyperspectral data- Acquired by Earth Observer-1 Satellite May 31, 2001

- Converted to reflectances, destriped, georeferenced



15995exposed soils14

268mixed mopane13

181short mopane12

305acacia grasslands11

248acacia shrublands10314acacia woodlands9

203island interior 8

259firescar 7

269riparian6

269reeds5 215floodplain grasses24

251floodplain grasses13

101hippo grass2

270water 1

Sample SizeClass NameClass #

Design of Experiment for Random Forests

• Create 10 independent stratified samples of training/test data – Select 75%:25% of training data for training and test

– Subsample from 75% to achieve 50, 15% sample size; test remains

constant



160

constant

– Select spatially removed test sample

• Create BHC tree

– Select bagged sample

– Randomly sample input features

• (D/5) or min(D/5, 20)

– Develop tree

• Grow BHC Random Forest

• Compare to BB-BHC, RS-BHC, BB RS BHC, RF CART

Adequate Destriping is Critical for Classification



161

Classification Results with Poor Destriping

Classification Results: BHC AlgorithmOkavango Hyperion



162

(a) Subset of Hyperion data

(Bands 51, 149, 31),(b) Classified image of Hyperion data

Classification Accuracy, Ind Test Set (%)

66

67

68

69

70

71

72

73

A c c u r a c y

Classification Accuracies, Independent Test Set



163

Std. Deviation, Classification Accuracy

0

0.5

1

1.5

2

2.5

3

3.5

4

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8

Training Sample Fraction

64

65

66

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8

Training Sample Fraction

BB RS BHC RS BHC BB BHC

RF BHC1 RF BHC2 RF CART



164

Next Steps - an interactive discussion

Documents

igarss04.hyperspectral