Validation activity within VALERI

1
Validation activity withinVALERI

• F. Baret, M. Weiss, S. Garrigues D. Allard

2
Objectives

• Evaluation of the absolute accuracy of medium
  resolution satellite products
• LAI
• fAPAR
• fCover
• Developing a network of sites distributed across
• biomes,
• conditions and
• Locations
• Develop a methodology
• To generate high spatial resolution maps from
  ground measurements
• To compare medium resolution products to the high
  spatial resolution map derived from ground
  measurements

3
Content

• The actual network of sites
• Performing the ground measurements of LAI, fAPAR
  and fCover
• Extension of the ground measurements to provide
  a high spatial resolution map
• Comparison with actual satellite products
• Some results over CYCLOPES products
• Conclusion

4
The network of sites (1/2)
25 sites 43 campaigns
5
The network of sites (2/2)
FAO Cover Classes VALERI sites Total Validation activities sites/FAO class () FAO Classes ()
Needleleaf forest 9 26 23.2 6.4
Broadleaf forest 5 11 9.8 11.9
Mixed forest 6 31 27.7 4.9
Closed shrublands 1 1 0.9 2.0
Open shrublands 0 4 3.6 14.0
Woody savannas 0 3 2.7 7.9
Savannas 3 5 4.5 7.2
Grassland 4 10 8.9 8.6
Permanent wetlands 0 0 0.0 1.0
Croplands 11 17 15.2 10.8
Cropland natural vegetation mosaic 1 1 0.9 10.8
Barren and sparsely vegetated 3 3 2.7 14.4
TOTAL 43 112 100.0 100.0
Not too bad overall representativity but large
variations within biomes! (space time)
6
Ground measurements (1/5)

• based on gap fraction measurements
• need to be
• accurate
• distinction between green/non green elements
• Clumping
• Above/below measurements
• quick
• Allow better spatial representativity
• Performed in a range of zenith directions
• Computation of fCover and fAPAR
• Traceable (quality insurance!)

Hemispherical photos/ LAI2000 (mainly) selected
7
Ground measurements (2/5)
The CAN_EYE processing software
2-15 minutes required to process a series of up
to 20 photos
8
Ground measurements (3/5)
fCover fAPAR computation
Classification
9
Ground measurements (4/5)

• LAI Accounting for the clumping
• Gap size distribution
• Need input canopy characteristics
• Average of local estimates
• Need maximum LAI value (saturation)
• Fractal dimension
• To be evaluated

• LAI derivation from photographs
• still needs validation
• Saturation for large LAIs
• Resolution (P57)
• Clumping

Average oflocal estimates LAIf(ALA, lO(q))
10
Ground measurements (5/5)
Spatial sampling 2 scales are
distinguished
11
Extending ground measurements to the whole
siteTransfer function and collocated-cokrigeage
12
Transfer functions (1/4)
Extending the local measurements to the whole
site using a high spatial resolution image
2 types of transfer functions
Based on RT model inversion ? Need accurate
atmospheric correction ? Independency with the
method to be validated? ? Need ground
measurements synchronous with high spatial
resolution image ? Ground measurements used to
correct from biases evaluate

• Empirical Transfer functions
• ? No need for atmospheric correction ? Truly
  independent from the method to be validated
• ? No strict synchronicity between ground
  measurements and high spatial resolution image
• ? Need ESUs to represent very well the
  variability of the whole site
• ? Need accurate ground measurements

Empirical transfer functions preferred
13
Transfer function (3/4)
Representativity of the ESUs (1/2)
Classification detection of mis-represented
classes
14
Transfer function (3/4)
Representativity of the ESUs (2/2)
Convex Hull inside CH inside CH 5 outside CH 5
Statistical test
15
Transfer functions (4/4)
Selection of the best transfer function
Generally regression performs better than
LUTs Use of the robust regression to weight
outliers Include additional bare soil/water
pixels Selection of best band combination
(depends on variables) Results evaluated by cross
validation
Weights associated to each ESU for the best band
combination
16
Collocated-cokriging
To retain the spatial information associated to
the ESU measurements, collocated-cokriging is
applied. It will allow to later compute the
accuracy of aggregated maps
measurements
Radiometric information
Constraint
17
Comparison with actual satellite products
Need accurate registration between high spatial
resolution image and medium resolution satellite
products This is achieved using correlation
techniques
Correlation coefficient
Correlation coefficient
Latitude position
The aggregation process needs to account for the
actual PSF (Point Spread function) of the
product Equivalent PSF adjusted over the site
based on the high spatial resolution image (need
heterogeneous site)
18
Results over the CYCLOPES products (1/5)
Carbon Cycle and Change in Land Observational
Products from an Ensemble of Satellites
Objectives

• Demonstrate the capacity of producing
  operationally consistent global fields of
  biophysical variables over long and continuous
  time series
• Use the products within two applications related
  to Climate change issues (GMES)
• Detection and categorization of land cover change
• Introduction within Global carbon cycle models

The products

• Sensors
• AVHRR
• VEGETATION
• POLDER
• MERIS
• (MSG)

• Resolution 1km - 8km
• Temporal sampling10 days
• Coverage Global
• Duration 1997-2003

• Biophysical variables
• Albedo
• fAPAR
• fCover
• LAI

19
Results over the CYCLOPES products (2/5)
Version 1 (May 2004)
Using already existing algorithms to meet the
deadline (To12)!! Global 8km for 2002-2003
Africa-Europe for 1km 2002-2003 (VGT)
VI Roujean/Lacaze
Aerosol Climatology
Original threshold
20
Results over the CYCLOPES products (3/5)
8 km Global 2002 2003
H0V0
H1V0
H2V0
H3V0
H4V0
H5V0
H0V1
H1V1
H2V1
H3V1
H4V1
H5V1
1 km Europe Africa 2002 2003
1 year global requires about 1 month processing !!
21
Results over the CYCLOPES products (4/5)
Validation over ground measurements based on
VALERI
22
Results over the CYCLOPES products (5/5)
Temporal consistency evaluation over CYTTARES
CYTTARES a network of sites representative of
all biomes and conditions 420 sites
(50x50km²)-(11x11km²)
Very good temporal consistency although gaps
(clouds) are observed To be compared with MODIS
(still in progress)
23
CONCLUSION

• A methodology has been developed and refined to
  generate high spatial resolution maps from ground
  measurements and a high spatial resolution image.
  
• 3 days preparation (1 people)
• 2-4 days needed to sample a site (2 to 4 people)
  
• 15 days to process a site and generate the
  reports (1 people).
• Limiting factors
• Accuracy of ground measurements. Still to
  evaluate
• Accounting for understory
• Registration/equivalent PSF
• How to extend above 3x3km² sites?
• Overall accuracy of aggregated biophysical
  variables
• Importance of the cooperation between validation
  activities.
• VALERI results available at avignon.inra.fr/valeri
  but slowly processing the archive (lack of
  resources!)
• Intercomparison of products very rich process. To
  be achieved over the same data base CYTTARES?
• Extension of the validation activity to
• Calibration of algorithms? Need lot of sites!!
• Validation of higher spatial resolution sensors
  (future)

24
Acknowledgements

• Many thanks to all the people who participated
  and supported VALERI, including
• Marc Leroy, Hervé Jeanjean, R. Fernandes, R.
  Myneni, J. Privette, J. Morisette, H. Bohbot, R.
  Bosseno, G. Dedieu, P. Sternberg, C. Di Bella, B.
  Duchemin, M. Espana, V. Gond, X.F. Gu, D. Guyon,
  C. Lelong, P. Maisongrande, E. Mougin, T. Nilson,
  F. Veroustraete, R. Vintilla, R. Silberstein, J.
  Wallace, N. Bruguier, R. Olivier, J.F. Hanocq, O.
  Marloie, B. Combal, O. Bajoles, N. Rochdi, C.
  Lazar, L. Prévot, L. de Beaufort, R. Cardenas,
  J.L. Roujan, G. Smith, S. Rambal, Freddy Loza de
  la Cruz, Rufian Villca Carex, Justina Mollo de
  Villca. P. Hiernaux, L. Jarlan, F. Lavenu, G.
  Marty, P. Mazzega, Y. Tracol, B. Ruelle, S.
  Weber, O. Ngwété et R. Santé, M. Canci, Noami
  Kerp, G. Hodgson, I. Colquhoun and all those
  who are cited!!
• CNES for its constant support
• ESA, NASA, BU for providing images