integration of sensors applied on south african ecosystems (isafe) · south african ecosystems...
Post on 17-Aug-2020
9 Views
Preview:
TRANSCRIPT
INTEGRATION OF SENSORS APPLIED ONSOUTH AFRICAN ECOSYSTEMS (ISAFE)
Presented by Philip FrostMain Authors: Else Swinnen & Karen Wentzel
Slide 2
Objectives
• Reprocessing of the 1km NOAA AVHRR data set forSouthern Africa (1985 – 2000)
• Spectral library of selected South African land coverclasses (low, medium and high NDVI)
• Estimation of correction functions for the integration ofAVHRR, Spot VEG and MODIS sensors
• Establishment of long term archive AVHRR-VGT,AVHRR-MODIS
Slide 3
Pre-processing of AVHRR imagery
•• VGTVGT AVHRRAVHRR
• Atmospheric correction (SMAC)• H2O: 6-hourly measurements from
MeteoFrance• O3: climatology based on TOMS
data• Aod: empirical function / calculated
from B0• Interpolation of inputs in time and
space• Geometric accuracy:
• < 0.5 pixel
• Resampling: bicubic convolution
• MVC: max value TOA NDVI
• BDC: Roujean model; unlimited timewindow
• Atmospheric correction (SMAC)• H2O: 6-hourly 1 degree
measurements from ECMWF• O3: climatology based on TOMS
data• Aod: empirical function
• Interpolation of inputs in time andspace
• Geometric accuracy:• mean RMSE: 1.04 pixels stdev
0.07 pixels• Resampling nearest neighbour
• MVC: max value TOA NDVI + contrainton VZA
• BDC: Roujean model; time windowlimited to 2 months
Slide 4
CalibrationCalibrationValidation on Namib desertValidation on Namib desert : NOAA-9, 11 and 14Vermote and El Saleous for NOAA-9 and NOAA-14, Mitchell for NOAA-11.As recommended by Calwatch
VES CP - MIT VES
RC RC RC VES MIT VES
VES VES VES
Slide 5
BRDF
• NTAM = Non-linear Temporal Angular Model (Latifovic)• Very low R² between actual values and model fit for low
vegetation classes and the R² decreased as the gapfraction increased.
• The method works well on densely vegetated areas.• Conclusion: for BRDF-correction, if the model does
not fit well to your actual values, you only introduceadditional noise (is so for BDC-algorithm and NTAMon this dataset). Then it’s better not to correct forBRDF.
Slide 6
Integration of AVHRR and VEGETATIONIntegration of AVHRR and VEGETATIONarchivearchive
• Based on overlapping year 1998
• Sources of inconsistency:• BRDF-effects : different overpass time• Spectral Response Function (SRF)• Point Spread Function (PSF)• Mis-registration errors
Slide 7
Slide 8
Methodology
• Acquisition of ASD spectra (400-1200nm) of various land coverclasses
• Convolution of SRF and ASD measured spectral signatures perland cover class in order to simulate sensor response
• Calculation of NDVI
• Calculate correction functions based on polynomial fit for absoluteas well as relative differences between sensors responses• All sensors are compared relatively to each sensors
n
nn
!
"!"!"!"
#
++$= 2211
dNIR
dNIRNDVI
Re
Re
!!
!!
+
"=
Slide 9
NDVI values calculated from simulatedsensor responses
NDVI
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
AVHRR 09
AVHRR 11
AVHRR 14
AVHRR 16
VGT01
Mod Terra
Mod Aqua
Sugar cane_site1117
White Maize_site1516
Pumpkin1_site415
Open Woodland_site914
Open Bushland2_site1013
Young Grass_site1412
Green Grass_site1311
Dense Woodland_site310
Dry Grass _site129
OGrass1_siteB38
Oschrub2_siteB27
DegradedWoodland2_site66
OGrass2_siteB45
Oschrub_siteB14
Bare Soil1_site53
Bare Soil2_site82
Bare Rock_site71
ASD Sites
Slide 10
Methodological approach
Southern African LandCoverBoreal regionAgricultural sites
2nd Oder Polynomial2nd Oder PolynomialLinear transformation
Abs & Relative differencecompared all with all
Abs & Relative difference inrelation to AVHRR-9
Red, NIR,NDVI TOCRed, NIR, NDVI TOC &TOANDVI TOC
ASD (Field)Spectral curves form(ASTER & PROBE)ASD (Field)
SRFSRFSRF
ISAFETrischenko et al. 2002Steven et al. 2003
Slide 11
Results• Comparison of AVHRR sensors
• Analysis of 3 families of AVHRRsensors indicate difference betweenresponses is NOT random but is relatedto amount of green vegetation
• High correlation between NOAA-9 &NOAA-11 therefore no correlationfunctions required between them
• NDVI of NOAA-14 (for vegetatedsurfaces) is 0.3% lower than NOAA-9and NOAA-11
• NDVI of NOAA-16 is about 5% higherthan the NDVI of NOAA-9, NOAA- 11 &NOAA-14: A simple correction of 5%suggested,
• However there is NO imagery overlapbetween NOAA-16 with AVHRRsensors onboard NOAA-9, 11 & 14
Red
y = -20.267x2 - 7.3555x + 0.4556
R2 = 0.9854
-15
-10
-5
0
5
10
15
0 0.2 0.4 0.6 0.8
NDVI NOAA-14
NOAA-16 - NOAA-14
Rel Diff
Poly. (NOAA-16 -
NOAA-14 Rel Diff)
NIR
y = 5.2443x2 - 1.7773x + 0.1357
R2 = 0.8773
-15
-10
-5
0
5
10
15
0 0.2 0.4 0.6 0.8
NDVI NOAA-14
NOAA-16 - NOAA-14
Rel Diff
Poly. (NOAA-16 -
NOAA-14 Rel Diff)
NDVI
y = -21.112x2 + 22.152x + 0.8725
R2 = 0.7737
-15
-10
-5
0
5
10
15
0 0.2 0.4 0.6 0.8
NDVI NOAA-14
NOAA-16 - NOAA-14
Rel Diff
Poly. (NOAA-16 -
NOAA-14 Rel Diff)
Slide 12
ResultsRed
y = -51.066x2 + 15.784x + 1.378
R2 = 0.9157
-15
-10
-5
0
5
10
15
0 0.2 0.4 0.6 0.8
NDVI NOAA-14
VGT01 - NOAA-14
Rel Diff
Poly. (VGT01 -
NOAA-14 Rel Diff)
NIR
y = 28.137x2 - 14.89x + 0.9394
R2 = 0.7675
-15
-10
-5
0
5
10
15
0 0.2 0.4 0.6 0.8
NDVI NOAA-14
VGT01 - NOAA-14
Rel Diff
Poly. (VGT01 -
NOAA-14 Rel Diff)
NDVI
y = -71.98x2 + 93.46x - 24.978
R2 = 0.7859
-15
-10
-5
0
5
10
15
0 0.2 0.4 0.6 0.8
NDVI NOAA-14
VGT01 - NOAA-14
Rel Diff
Poly. (VGT01 -
NOAA-14 Rel Diff)
• NDVI is generally higher for VGT thanfor AVHRR sensors because ofnarrow width and exclusion ofwavelength subject to water vapor
• NDVI difference between VGT andAVHRR sensors depend on surfacethat is measured – Polynomialcorrection coefficients is required forthis relationship
• Relative correction shows betterresults than absolute comparison
Slide 13
Results Red
y = 6.7139x2 + 4.2061x + 0.5488
R2 = 0.9452
-15
-10
-5
0
5
10
15
0 0.2 0.4 0.6 0.8
NDVI Mod Terra
VGT01 - Mod Terra
Rel Diff
Poly. (VGT01 - Mod
Terra Rel Diff)
NIR
y = 8.6364x2 - 8.7742x - 0.3685
R2 = 0.2222
-15
-10
-5
0
5
10
15
0 0.2 0.4 0.6 0.8
NDVI Mod Terra
VGT01 - Mod Terra
Rel Diff
Poly. (VGT01 - Mod
Terra Rel Diff)
NDVI
y = -15.87x2 + 25.469x - 13.501
R2 = 0.7438
-15
-10
-5
0
5
10
15
0 0.2 0.4 0.6 0.8
NDVI Mod Terra
VGT01 - Mod Terra
Rel Diff
Poly. (VGT01 - Mod
Terra Rel Diff)
• Similar NDVI trend occurs forMODIS sensors as in the caseof VGT – slightly higher NDVIvalues for green vegetation
Slide 14
RMSEorig - RMSESRF
Lower RMSE
Higher RMSE
S
Tabs
Trel
Iabs
Irel
Slide 15
Conclusions
• SR correction functions are required to inter-calibratebetween different sensors especially VGT & AVHRR, VGT& MODIS and AVHRR & MODIS
• Correction functions are not linear, but related tovegetation greenness
• Correction functions derived from Relative comparisonbetween sensors provide higher R2 values than Absolutecomparison and should be used in inter-calibration
Slide 16
VGT time series 1985 – 2005SRF corrections not applied
Slide 17
Acknowledgement
• GLOVEG-project: scientific support of the VEGETATIONinstrument
top related