International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part BI. Istanbul 2004
2. “LAND SURFACE” LEVEL 3 ALGORITHM
The first algorithms of the “Land Surface" processing line,
applied to the ADEOS-1/POLDER-1 data, took advantage of
the POLDER directionality (Leroy et al., 1997). The Level 3
products are generated over a synthesis period of 30 days with a
sliding window to get a temporal resolution of 10 days.
Advanced algorithms have been developed to be applied to
ADEOS-2/POLDER-2 data. In addition to the Leaf Area Index
(LAI) retrieval, the major improvements of the algorithm are:
1) a multi-temporal filtering module that eliminates the
observations contaminated by residual clouds and/or
aerosols.
2) the application of temporal weighting favouring the
data collected in the middle of the synthesis period.
Thus, the smoothing of the temporal variations of the
biophysical parameters due to the monthly synthesis
decreases;
3) the calculation of an error associated with each
parameter. This value depends on the noise on the
input data (i.e. the measured bi-directional spectral
reflectances) and on the retrieval algorithm.
The “Land Surface” Level 3 algorithm relies on 3 major steps as
shown on Figure 1.
LEVEL 2
Surface Directional Reflectances
Filtering
module
Vegetation model
inversion by a
neural network
BRDF model
inversion
LEVEL 3
« Directional Parameter »
C nroduct
LEVEL 3 z^
« Albedo and Vegetation »
product
HS XL,
Figure 1. Diagram of the "Land Surface" Level 3 algorithm
2.1 The filtering module
The bi-directional spectral reflectances are the inputs of the
*Land Surface" Level 3 processing line. Their quality controls
the relevance of biophysical parameters. In order to complete
the Level 2 cloud masking and eliminate the disturbed data, a
multi-temporal filtering module has been implemented. It
determines the type of surface (ground, snow or mixed),
identifies the temporal inconstancies of the measured
reflectances over the synthesis period of 30 days, and filters the
observations contaminated by residual clouds and/or aerosols.
This latter point is based on the fitting of the directional model
of Walthall et al., (1985) on the reflectances acquired at 443nm
under angular configurations close to the perpendicular plane.
The advanced algorithm has been tested on ADEOS-
1/POLDER-1 data (Lacaze et al., 2003). The inconsistent spatial
variability over cloudy areas is clearly reduced, and the bio-
geophysical parameters appear more homogeneous.
2.2 The linear inversion of a BRDF model
A new linear semi-empirical BRDF model proposed by
Maignan et al. (2004) has been implemented in the Level 3
processing line to normalize the bi-directional POLDER-2
measurements. This model combines the reciprocal geometric
kernel of “Li_sparse” (Lucht et al., 2000) with the volumic
kernel of "Ross thick" (Roujean et al., 1992). The innovation is
the merging of the *Ross thick" kernel with a hotspot module
(Bréon et al., 2002) which allows to reproduce more accurately
the hotspot phenomenon well described on the POLDER
BRDFs. A Gaussian temporal weighting is applied to measured
reflectances to enhance the representation of the center of the
synthesis period. The resulting spectral directional coefficients
are:
e a nadir-zenith reflectance, kO
e a roughness indicator, kl
e a volume scattering indicator, k2.
However, because of the enhanced correlation between the
reciprocal “Li-sparse” kernel and the new "Ross thick hotspot"
kernel, the individual meaning of the directional coefficients as
surface indicators should be cautious. Their optimal use is as a
set of coefficients to accurately simulate the BRDF. So, they are
used for computing the spectral Directional Hemispherical
Reflectances (DHR) for the median sun angle of the synthesis
period. Then, the NDVI, corrected for the directional effects, is
derived from DHR670nm and DHRg6snm-
The quality of the inversion remains dependent on the angular
distribution of acquisitions in the directional hemisphere. It is
estimated through the coefficient of determination R? and the
root mean square error (rmse) between the measured and
simulated reflectances. R? and rmse are provided in the Data
Quality Index (DQX).
2.3 The inversion of a vegetation model by a neural network
The Leaf Area Index (LAI) is defined as half the total
intercepting green foliage area per unit ground surface area
(Chen et Black, 1992). The ADEOS-2/POLDER-2 algorithm
computes the LAI using a neural network, which inverts the
radiative transfer model of Kuusk (1995) considering the
vegetation as a turbid medium of leaves with spherical
orientation. This model simulates the simple scattering in the
canopy (in particular the hot spot phenomenon quantified by the
parameter 1*, ratio of the leaf size to the canopy height)
following the Nilson and Kuusk (1989) approach, and the
multiple scattering according the SAIL model (Verhoef, 1984).
Furthermore, the leaf optical properties are described by the
PROSPECT model (Jacquemoud et al, 1996) whereas the
spectral and angular properties of soil are reproduced by the
coupling of the functions of Price (1990) and the directional
model of Walthall et al., (1985), respectively.
Internati
For the :
the PRC
chloropl
concentr
equivale
inside a
first fun
consider
network
80u g/en
paramet
0.1 and :
The net
reflectan
and the
estimate
algorithi
retrievec
characte
Vegetati
covered
Le
trig ey Wood nd
ü sb
Us
0.7
uerit
NDVI
us|-i-
us
m
Le)
E
er
-
sn?
vo wd
Mi
DHR, 670 nm
e
à:
©
u
ar
|
Figur
The Lev
the con