Don Atwood & David Small IGARSS July 2011 1 USE OF RADIOMETRIC TERRAIN CORRECTION TO IMPROVE POLSAR LAND COVER CLASSIFICATION Don Atwood 1 and David Small 2 1) University of Alaska Fairbanks 2) University of Zurich, Switzerland
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Don Atwood & David SmallIGARSS July 2011 1
USE OF RADIOMETRIC TERRAIN
CORRECTION TO IMPROVE
POLSAR LAND COVER CLASSIFICATION
Don Atwood1 and David Small2
1) University of Alaska Fairbanks
2) University of Zurich, Switzerland
Don Atwood & David SmallIGARSS July 2011 2
Presentation Overview
• Introduce Boreal Land Cover Classification project
• Focus on species differentiation in boreal environment
• Introduce reference data for land cover classification
• Introduce method of Radiometric Terrain Correction (RTC)
• Terrain-flattened Gamma Naught Backscatter
• Perform RTC on polarimetric parameters to address topography
• Demonstrate synergy of PolSARpro and MapReady Tools
• Compare results for RTC-corrected and non-corrected classification
• Characterize optimal classification approach for Interior Alaska
g T
0
Don Atwood & David SmallIGARSS July 2011 3
Study Region
Boreal environment of Interior Alaska
Characterized by:
• rivers
• wetlands
• herbaceous tundra
• black spruce forests (north facing)
• birch forests (south facing)
• low intensity urban areas
Don Atwood & David SmallIGARSS July 2011 4
Land Cover Reference
Don Atwood & David SmallIGARSS July 2011 5
Study Data
Quad-Pol data selected:
• ALOS L-band PALSAR
• 21.5 degree look angle
• Of April, May, July, and Nov dates,
July 12 2009 selected
• Post-thaw
• Leaf-on
• Coverage includes Fairbanks and
regional roads
Pauli Image
Don Atwood & David SmallIGARSS July 2011 6
Problem of Topography
Span (Trace of T3 Matrix) Wishart Segmentation
Don Atwood & David SmallIGARSS July 2011 7
Normalized Radar
Cross Sections
Nadir
Sensor
Far
Near
Aσ & σ0
Aγ & γ0
Aβ & β0
Let’s compute Normalized Radar Cross Sections
• for an Ellipsoidal Earth
• for Topography
Don Atwood & David SmallIGARSS July 2011 8
sE
0 = b0 × Ab / As = b0 ×sinq
g E
0 = b0 × Ab / Ag = b0 × tanq
Relationships between cross sections
for ellipsoidal surfaces
Backscatter Reference Areas
For an Ellipsoidal Earth
Don Atwood & David SmallIGARSS July 2011 9
Terrain-flattening
We need to move beyond the ellipsoidal Earth to the hills and valleys of
the Fairbanks region:
• Address the layover and foreshortening of geometric distortions
• Correct the radiometric variations associated with topography.
To improve our radiometry:
➡use local area contributing to backscatter at each location in the
SAR scene
Don Atwood & David SmallIGARSS July 2011 10
Terrain-flattening
Convention 1 2 3 4 5
Earth Model None Ellipsoid Terrain
Reference Area
Area
Derivation
Normalisation
Product GTC NORLIM RTC
b 0 s E
0 g E
0 g T
0
Ab AsAg Ag
dr ×dadg ×da d p ×da
d p ×da
DHM
ò
b 0 =b
Ab
b 0 ×Ab
As
= b 0 × sinqE
b 0 ×Ab
Ag
= b 0 × tanqE
b 0 ×Ab
Ag
s E
00 ×As
As
= s E
00 ×sinqLIM
sinqE
sT
0
dg ×da
As
Don Atwood & David SmallIGARSS July 2011 11
Ref.: Small, D., Flattening Gamma: Radiometric Terrain Correction for SAR
Imagery, IEEE Transactions on Geoscience and Remote Sensing, 13p (in press).
Terrain-flattening
=
Aγ /Aβ
β0
-20dB 5dB
g T
0 = b 0iAb
Ag
X
Solution: Use simulated image to Normalize β0
g T
0
Don Atwood & David SmallIGARSS July 2011 12
b 0 -20dB 5dB
Vancouver
GTC (Sept 2008) Integrated contributing area
ENVISAT ASAR WSM data courtesy ESA (based on SRTM3)
Terrain Correction
in Coastal BC
Don Atwood & David SmallIGARSS July 2011 13
Terrain Correction
in Coastal BC
b 0 -20dB 5dB
Zoom
GTC (Sept 2008) Integrated contributing area
ENVISAT ASAR WSM data courtesy ESA (based on SRTM3)
Don Atwood & David SmallIGARSS July 2011 14
ASAR WSM GTCg E
0
Coastal BC: GTC
Don Atwood & David SmallIGARSS July 2011 15
ASAR WSM RTCg T
0
Coastal BC: RTC
Don Atwood & David SmallIGARSS July 2011 16
ASAR WSM NORLIMsNORLIM
0
Coastal BC: NORLIM
Don Atwood & David SmallIGARSS July 2011 17
Coherency Matrix
YXYX
XYXX
SS
SSS
Scattering Matrix
2**
*2*
**2
3
422
2
2*
2
1
XYYYXXXYYYXXXY
XYYYXXYYXXYYXXYYXX
XYYYXXYYXXYYXXYYXX
SSSSSSS
SSSSSSSSS
SSSSSSSSS
T
22T11T 33T: “Single Bounce” : “Double Bounce” : “Volume Scattering”
Don Atwood & David SmallIGARSS July 2011 18
Radiometric Terrain Correctionof Coherency Matrix
• Radiometric Terrain Correction:
Area Normalization
terrain corrected
Coherency Matrix
333231
232221
131211
3
TTT
TTT
TTT
T
Coherency Matrix
333231
232221
131211
3
TTT
TTT
TTT
T
• Scale all matrix elements by Area Normalization
Don Atwood & David SmallIGARSS July 2011 19
For a given class, the ratio of Surface, Double Bounce, and Volume
scattering components depend on incidence angle
POLARIMETRIC IMPLICATIONS OF INCIDENCE ANGLE VARIABILITY FOR UAVSAR
Guritz, Atwood, Chapman, and Hensley
But Wait…..
Don Atwood & David SmallIGARSS July 2011 20
Radiometric Terrain Correctionof Coherency Matrix
Span: No Normalization Span: Terrain-model Normalization
Don Atwood & David SmallIGARSS July 2011 21
Radiometric Terrain Correctionof Coherency Matrix
Pauli: No Normalization Pauli: Terrain-model Normalization
Don Atwood & David SmallIGARSS July 2011 22
Ingest PALSAR data Terrain-correct Perform Wishart Export to GIS
Generate T3 with MapReady decomposition Cluster-busting
Radiometric correction using area
Lee Sigma Speckle Filter
POA compensation
Integration of PolSARpro
and MapReady
Don Atwood & David SmallIGARSS July 2011 23
Radiometric Terrain Correctionof Coherency Matrix
Wishart - No Normalization Wishart - Radiometric Correction
Don Atwood & David SmallIGARSS July 2011 24
Radiometric Terrain Correctionof Coherency Matrix
USGS Reference Wishart – Radiometric Correction
Don Atwood & David SmallIGARSS July 2011 25
Classification Results
No Normalization USGS Reference RTC
Don Atwood & David SmallIGARSS July 2011 26
Classification Results
Urban areas missed / Identified as Open Water
Don Atwood & David SmallIGARSS July 2011 27
Classification Results
Inability to distinguish Mixed Forests and Shrub / Scrub
Don Atwood & David SmallIGARSS July 2011 28
Accuracy AssessmentNo Normalization
No NormalizationOpen Water
Developed Land
Barren Land
Deciduous Forest
Evergreen Forest
Mixed Forest
Shrub/ Scrub
Woody Wetlands
Herbaceous Wetlands
User Accuracy
Open Water 42402 22539 15229 2168 1512 99 1024 6299 498 46%
Developed Land 836 27431 1304 3130 903 458 123 2663 64 74%
Barren Land 0 0 0 0 0 0 0 0 0 NA
Deciduous Forest 11217 50614 1795 390417 228454 112888 12687 52712 528 45%
Evergreen Forest 13734 69849 6849 162366 323079 49803 12643 94157 617 44%
Mixed Forest 0 0 0 0 0 0 0 0 0 NA
Shrub/ Scrub 0 0 0 0 0 0 0 0 0 NA
Woody Wetlands 7062 15611 4924 56052 135667 12103 30585 480635 11594 65%
Herbaceous Wetlands 0 0 0 0 0 0 0 0 0 NA
Producer Accuracy 56% 15% 0% 64% 47% 0% 0% 76% 0% 51%
Don Atwood & David SmallIGARSS July 2011 29
Accuracy AssessmentWith RTC
Normalized T3Open Water
Developed Land
Barren Land
Deciduous Forest
Evergreen Forest
Mixed Forest
Shrub/ Scrub
Woody Wetlands
Herbaceous Wetlands
User Accuracy
Open Water 45570 33695 17297 3595 2188 165 1616 9905 739 40%
Developed Land 942 27464 1320 4717 1547 608 148 1878 27 71%
Barren Land 0 0 0 0 0 0 0 0 0 NA
Deciduous Forest 10161 59438 1461 482548 234568 128097 10344 30375 147 50%
Evergreen Forest 10614 50149 4409 53025 335583 30621 13520 138224 527 53%
Mixed Forest 0 0 0 0 0 0 0 0 0 NA
Shrub/ Scrub 0 0 0 0 0 0 0 0 0 NA
Woody Wetlands 7964 15298 5614 70248 115729 15860 31434 456084 11861 64%
Herbaceous Wetlands 0 0 0 0 0 0 0 0 0 NA
Producer Accuracy 61% 15% 0% 79% 49% 0% 0% 72% 0% 54%
Don Atwood & David SmallIGARSS July 2011 30
Accuracy AssessmentComparison
Producer Class RTC No RTC Improvement
Open Water 61% 56% 5%
Developed Land 15% 15% 0%
Deciduous Forest 79% 64% 15%
Evergreen Forest 49% 47% 2%
Woody Wetlands 72% 76% -4%
• RTC yields improved accuracy (particularly for Deciduous Forest)
Don Atwood & David SmallIGARSS July 2011 31
Impact of RTC on forest classification
No Normalization USGS Reference RTC
Don Atwood & David SmallIGARSS July 2011 32
Conclusions
• In general, PolSAR classification is difficult!
• Data fusion provides greatest hope for accurate classification results
• Radiometric variability caused by topography dominates PolSAR classification
• Area-based RTC offers effective way to “flatten” SAR radiometry
• RTC of Coherency Matrix shown to improve classification accuracy:
• Impact most pronounced for Deciduous Forests
• Although not complete, RTC approach is simple and effective
• Different scattering mechanisms (SB, DB, Volume) have different
sensitivities to topography. RTC does not address this
• However, RTC is very effective first order correction for segmenting
polarimetric data by phenology rather than topography
Don Atwood & David SmallIGARSS July 2011 33
Discussion
Don Atwood
(907) 474-7380
Photo Credit: Don Atwood