1. A physics-based atmospheric and BRDF
correction for Landsat data over
mountainous terrain
Presentation by Fuqin Li1, David Jupp2, Medhavy Thankappan1,
Leo Lymburner1, Norman Mueller1, Adam Lewis1 and Alex Held23
1
National Earth Observation Group, Geoscience Australia
2
CSIRO, Marine and Atmospheric Research
3
TERN AusCover

2. Contents
• Why we need to do the correction?
• Method
• Results
• Discussions: Implications for multi-temporal
land cover mapping and effects of DSM
quality

3. GA current standard processing product
(atmospheric and BRDF correction)
Landsat orthocorrected images Atmospheric state: aerosol
and metadata, e.g., geographical optical depth, CO2, ozone and
coordinators, time, day, year etc water vapour etc
MODTRAN 5 or 6S radiative
View and solar zenith, transfer models/software
view and solar azimuth
Atmospheric parameters:
ρm, L0, tS, tV, S, td(θS), td(θV), Eh, Ehdir, Ehdif
A coupled BRDF and atmospheric model BRDF shape
for both flat and sloping surface function
Final product:
BRDF corrected surface reflectance

4. Landsat image over mountainous area
why do we need the extra correction?
Slopes facing sun Slopes away from sun
Uncorrected image shows
slopes facing sun are brighter
than slopes away from the sun
even if the vegetation cover is
the same. This will cause
problems for land cover
mapping and other applications
Two corrections are needed:
(1) Detect deep shadows
(self and cast shadows)
(2) Remove terrain shadows
Landsat 5 image over Victorian
Deep shadows
Alps (May 11, 2007)

5. An Australian terrain image
Map of terrain roughness for Australia, yellow is very
low relief with green and red high relief where
correction is essential

6. Self and cast shadow diagram
Sun
Zd D
H=Z0+d x tan(90-θS)
A
C 90-θS
Z0
d
B
Cast Self
shadow shadow

7. Methods: BRDF on flat and sloping surfaces
et it
θV θS
θt
Flat surface Sloping surface slope angle θt
 E h [ fV ρ S (θ S , θV , δϕ ) + (1 − fV ) ρ '] + 
dir  E dir [ fV ρ S (it , et , δϕ t ) + (1 − fV ) ρ '(it )] + 

TV 
 TV  
LTOA = L0 +
2
 LTOA = L0 +  S(ρ )  2
π  E h [ fV ρ + (1 − fV ) ρ ] + E h
dif Sρ  π  E dif [ fV ρ ( et ) + (1 − fV ) ρ ] + E 
  1− Sρ 
 1− Sρ    

8. Terrain correction flowchart
Landsat orthocorrected images Atmospheric state: aerosol
and metadata, e.g., geographical optical depth, CO2, ozone
DSM coordinators, time, day, year etc and water vapour etc
View and solar zenith, MODTRAN 5 or 6S radiative
Slope and aspect view and solar azimuth transfer models/software
Incident, exiting angles Atmospheric parameters:
and their relative azimuth ρm, L0, tS, tV, S, td(θS), td(θV), Eh, Ehdir, Ehdif
angles, cast shadow
A coupled BRDF and atmospheric model BRDF shape
for both flat and sloping surface function
Final product:
BRDF and terrain illumination
corrected surface reflectance

9. DSM and DEM data
According to definitions used by GA and CSIRO
(1) DSMs (Digital Surface Models) provide surface
height above sea level and may include effects of
forests and other local surface roughness features.
(2) DEMs (Digital Elevation Models) estimate the
elevation of the soil surface free of fine scale
roughness elements such as trees and buildings.
(3) For topographic correction, we used the GA-CSIRO
SRTM based DSM product with some pre-
processing, e.g. smoothing and filtering to remove
remaining artefacts.

10. Study Areas
Victoria Alps Blue Mountains

11. Results: visual assessment
Deep Shadow Deep Shadow
11/05/2007
Victorian Alps
25/02/2009
08/11/2009

12. Results: visual assessment
Deep Shadow Deep Shadow
21/08/2006
South 22/09/2006
Blue Mountains
06/01/2005

13. Decorrelation-indirect validation
6000 6000
Surface reflectance factor * 10000
Surface reflectance factor * 10000
R=0.02
4500 R=0.81 4500
3000 3000
1500 1500
0 0
0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00
Cosine incident angle Cosine incident angle
xy plot between cosine incident angle and Landsat band 4
reflectance value at selected target area where x-axis is cosine
incident angle and y-axis is band 4 reflectance factor for
February 25, 2009, (a) standard corrected product (b) terrain
corrected product.

14. Decorrelation-indirect validation
Correlation coefficient
Date
(day/month/year) Before terrain
After terrain correction
correction
06/01/2005 0.5859 -0.0723
15/04/2006 0.8134 0.0485
21/08/2006 0.8214 -0.0290
22/09/2006 0.7288 -0.0920
08/03/007 0.7092 0.0058
11/05/2007 0.6781 -0.0613
25/02/2009 0.8056 -0.0156
08/11/2009 0.4752 0.0946

15. Discussions: How the correction impact on Landcover
classfication
(a) (b)
Mean surface reflectance factors for Classes 2 and 3 (a) BRDF and atmospheric
correction only, (b) BRDF and atmospheric correction plus terrain correction. c2_NE is
the mean surface reflectance factor for the NE slopes of class 2, c2_SW is the mean
surface reflectance factor for the SW of class 2, c3_NE is the mean surface reflectance
factor for the NE slopes of class 3 and c3_SW is the mean surface reflectance factor for
the SW of class 3

16. Discussions
Correction quality and DSM
Wrong deep Shadows Miss deep Shadows
The impact of DSM artefacts on the accuracy of terrain
correction for the south Blue Mountains image of Sept.
22, 2006

17. Correction quality with co-registration accuracy
Feb 25, 2009
Sep 22, 2006
Correct co-registration 2 pixels shifted
The impact of co-registration between DSM and Landsat images on
the accuracy of terrain correction.

18. Correction quality with co-registration accuracy
Deep Shadow
Correct co-registration Wrong co-registration
The impact of co-registration between DSM and Landsat
images on the accuracy of terrain correction for the
South coast area

19. Correction accuracy with DSM spatial resolution
1 sec DSM 3 sec DSM
Feb 25, 2009
Sep 22, 2006
The impact of DSM spatial resolution on the accuracy
of terrain correction

20. Conclusions
• A physics-based BRDF and atmospheric correction
model can remove most of the topographic effect for
Landsat images and detect deep shadows.
• The method is independent of the image data but
requires a DSM/DEM
• The model can be applied to other similar resolution
satellite images.
• The correction quality depends on the DSM/DEM
quality, co-registration accuracy and both satellite and
DSM/DEM resolution.

21. Future work
• Further validation of the combined correction
algorithm using field work at different times
• Further testing with multi-temporal land
cover mapping applications
• Implement the algorithm into the GA
automatic processing system

22. Acknowledgements
• Aerosol data were provided by Ross
Mitchell’s group at CSIRO
• Access to MODIS BRDF data has been
facilitated by Edwards King’s group at
CSIRO
• The Geoscience Australia provided the
satellite images