Diapositiva 1

1
Barrax validation agricultural site Leassons
learnt during SPARC campaigns B. Martínez, F.
Camacho-de Coca, F.J. García-Haro, A. Verger UIT-
Universitat de Valencia F. Baret, M. Weiss INRA
- Avignon
2
CONTENTS

• OBJECTIVE
• TEST SITE AND FIELD CAMPAIGN
• TEST SITE CHARACTERIZATION
• - In-situ measurements
• - Sampling strategies
• TRANSFER FUNCTION Spatial extension of in-situ
  measurements
• A CASE STUDY WITH LOW RESOLUTION PRODUCTS

3
OBJECTIVE
To derive accurate high-resolution maps from
in-situ measurements for the validation of
(SEVIRI) coarse satellite vegetation products
VALERI methodology
Up-scaling
4
CONTENTS

• OBJECTIVE
• TEST SITE AND FIELD CAMPAIGN
• TEST SITE CHARACTERIZATION
• - In-situ measurements
• - Sampling strategies
• TRANSFER FUNCTION Spatial extension in-situ
  measurements
• A CASE STUDY WITH LOW RESOLUTION PRODUCTS

5
Test Site Barrax (Albacete)
Two different areas selected
Direct Validation The Barrax agricultural site
of 5?5 km2 (Las Tiesas) selected for ground
measurements acquisitions. All facilities are
available. Indirect Validation A larger area
of 50?50 km2 is selected for inter-comparison and
validation of SEVIRI products. Very flat area.
Crops and natural vegetation.
6
Field Campaign SPARC experiments
The SPARC campaigns are a combination of
different initiatives (ESA, CNES, EU, EUMETSAT)
but with the common interest of in-situ
characterisation simultaneously to airborne and
multi-sensors data acquisitions mainly focused on
algorithm and product validation. SPARC03 ?
from 13th to 14th of July 2003 SPARC04 ? from
13th to 17th of July 2004 Our participation was
funded by LSA SAF (EUMETSAT) !
In situ measurements Available Imagery
Gap Fraction (LAI,FVC, FAPAR) Chlorophyll Radiometry Temperature Emisivity Atmospheric profiles ROSIS (1m) HyMAP (5m) AHS (2.5m) SPOT/HRV (20m) Landsat/TM (30m) CHRIS/PROBA (34m) MERIS/Envisat (300m-1km) SEVIRI/Meteosat-8 (3 km)
7

• OBJECTIVE
• TEST SITE AND FIELD CAMPAIGN
• TEST SITE CHARACTERIZATION
• - In-situ measurements
• - Sampling strategies
• TRANSFER FUNCTION Spatial extension in-situ
  measurements
• A CASE STUDY WITH LOW RESOLUTION PRODUCTS

8
TEST SITE CHARACTERIZATION Sampling Strategy
Hemispherical camera Sampling design ? VALERI
methodology. 12 Photographs per ESU GPS was
recorded at the center of the ESU
LICOR LAI2000 Average of 3 replications 24
measurements per ESU The replications were
distributed randomly within the ESU
9
TEST SITE CHARACTERIZATION In-situ measurements
10
TEST SITE CHARACTERIZATION Results
1. Retrieved biophysical parameters
SPARC03
SPARC04
LICOR
DHP
11
TEST SITE CHARACTERIZATION Results
2. Comparison of LICOR and DHP LAI estimates
Relative Error between LICOR and DHP mean values
per fields.
LAI
SPARC 2003
LAI DHP-LICOR Relative error typically lt25
FVC DHP-LICOR Relative error lt15 for all cases
12
TEST SITE CHARACTERIZATION Results
3. Dependent on the camera position
During the SPARC04 field campaing, all the
photographs were taken UPWARD LOOKING when it was
possible. Special attention was paid in comparing
the results when the camera was in downward and
upward looking for some crops.
Sunflower
Corn
Er(LAI_DW) 37
Er(LAI_DW) 57
The estimated LAI can be twice in downward
looking position !!!
13
TEST SITE CHARACTERIZATION Results
Attention was paid during SPARC04 in measuring
simultaneously with LICOR and DHP. In addition,
an intercomparison of different LICORs was done
HP vs LICOR
LICOR vs LICOR
FIELD Er()
CORN (C1) 24
GARLIC (G1) 48
POTATO (P) 18
SUGARBEET (SB) 48
Good agreement between DHP (upward looking) and
LICOR estimates. The largest discrepancies found
for dens cover (Sugar Beet) are similar to that
shown by different LICOR instruments (around
50).
14

• OBJECTIVE
• TEST SITE AND FIELD CAMPAIGN
• TEST SITE CHARACTERIZATION
• - In-situ measurements
• - Sampling strategies
• TRANSFER FUNCTION Spatial extension in-situ
  measurements
• A CASE STUDY WITH LOW RESOLUTION PRODUCTS

15
TRANSFER FUNCTION Spatial Extension to High
resolution
Transfer Function Weighted Multiple linear
regressions were computed with all possible SPOT
bands combinations
The band combination was selected based on the
lowest RMSEW (Weighted Root Mean Square Error),
RCROSS (Cross Validation RMSE) and weights null,
following the methodology proposed by
Weiss,(2004)
16
TRANSFER FUNCTION Spatial Extension to High
resolution
FVC
MEAN 0.38
STD 0.28
FAPAR
MEAN 0.46
STD 0.33
HP LICOR
MEAN 1.91 1.71
STD 1.48 1.62
Correlate 0.99
Bias 0.204
RMS 0.263
17
Influence of Sampling Spatial Strategy
Transfer Function ? FLAG IMAGE Computation of
convex hull over the collocated radiance values
with the in situ measurements Smallest convex
region that
contains the data set
CONVEX HULL for DHP
White?Interpolated Black? Extrapolated Blue?
Considering a relative error of the 5
Different Samplings Designs
18

• OBJECTIVE
• TEST SITE AND FIELD CAMPAIGN
• TEST SITE CHARACTERIZATION
• - In-situ measurements
• - Sampling strategies
• TRANSFER FUNCTION Spatial extension in-situ
  measurements
• A CASE STUDY

19
A CASE STUDY Comparison with large scale
products (1 km res)
Barrax test site (5x5 km)
PRODUCTS (10-days composition) VGT_LandSAF_v1?
SEVIRI Land SAF algorithm (UV) on VGT k0 data
VGT_CYCLOPES_v1 MODIS/TERRA
20
A CASE STUDY Comparison with large scale products
Barrax large area (50x50 km2)
Leaf Area Index (LAI)
MODIS
VGT_LandSAF_V1
VGT_CYCLOPES_V1
In-situ_3km
Good agreement (RMSlt0.25, r2gt0.85) for both
algorithms LSA SAF_V1 and CYCLOPES_V1 on VGT data
21
A CASE STUDY Comparison with large scale products
Barrax large area (50x50 km2)
Fraction of Vegetation Cover (FVC)
In-situ_3km
VGT_CYCLOPES_V1
VGT_LandSAF_V1
Good agreement (RMSlt0.09,r2gt0.7) for the LSA
SAF_v1 algorithm on VGT data
22
A CASE STUDY Comparison with large scale products
Fraction of Absorbed PAR (FAPAR)
In-situ_3km
MODIS
VGT_CYCLOPES_v1
Good agreement (RMSlt0.08, r2gt0.8) for MODIS
product is found
23
CONCLUSIONS
?Large amount of ground and airborne data has
been collected in Barrax from 1998 (DAISEX) up to
now (SPARC, DEMETER), and two new ESA field
campaigns are planned for 2005 (June and July).
?DHP is used for LAI, FVC and FAPAR estimates in
addition to LICOR. ?The pre-processing with
CANEYE has been found quite independent of the
operator (UV vs INRA).
?However, DHP estimates are dependent on the
camera position. Downward looking overestimates
the LAI up to 50 regarding Upward looking. ?
Upward looking shows better consistency with
LICOR estimates, with errors not larger than
those found between LICORs estimates.

• Concerning the spatial extension of in-situ
  measurements, the best results have been obtained
  using a transfer function derived from a Weighted
  Multiple linear regression.
• Besides different sampling strategies were
  performed, the derived high-resolution LAI maps
  (LICOR and CAMERA) show relatively small
  differences (RMSElt0.3).
• ? Around 50 units covering the different crops
  seems to be enough for obtaining a good transfer
  function for the Barrax test site.

24
CONCLUSIONS
?Comparison of In-Situ degradated maps with
different products at 1km and 3km resolution over
the small and large area shows that the UV LSA
SAF algorithm on VEGETATION data shows the best
correlation for FVC and LAI with an RMS lt
VGT_CYCL-V1lt MODIS. ?MODIS FAPAR product shows a
good agreement (RMS0.08,r20.8). ?CYCLOPES
products overestimate FVC and LAI, whilst
underestimate FAPAR.
?Open issues - Derive high-resolution maps for
SPARC04 - Assessment of the TF outside the study
area (5x5km) - Evaluate different upscaling
methods from 20m to 3km
?Contact us for -LAI/FVC/FAPAR data and maps?
Fernando.Camacho_at_uv.es -SPARC database and new
ESA activities? Jose.Moreno_at_uv.es
25
Thank you for your attention !!
26
Influence of Sampling Spatial Strategy
Transfer Function ? FLAG IMAGE Computation of
convex hull over the collocated radiance values
with the in situ measurements Smallest convex
region that contains the data set
BIDIMENSIONAL COVEX HULL FOR THE BANDS
COMBINATION (GREEN, RED, NIR)
RED-GREEN
NIR-GREEN
27
AGGREGATION FOR VALIDATE LOW RESOLUTION
fAPAR HP (0-1)
LAI LICOR (0-5)
FVC HP (0-1)