integration of sensors applied on south african ecosystems (isafe) · south african ecosystems...

INTEGRATION OF SENSORS APPLIED ONSOUTH AFRICAN ECOSYSTEMS (ISAFE)

Presented by Philip FrostMain Authors: Else Swinnen & Karen Wentzel

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

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

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

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.

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

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

!!

!!

+

"=

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

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

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)

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

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

RMSEorig - RMSESRF

Lower RMSE

Higher RMSE

S

Tabs

Trel

Iabs

Irel

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

VGT time series 1985 – 2005SRF corrections not applied

Acknowledgement

• GLOVEG-project: scientific support of the VEGETATIONinstrument

integration of sensors applied on south african ecosystems (isafe) · south african ecosystems...

Documents