GPP 
NPP 
1000 
623 
30 
20 
2i(°C) 
£?(°Q 
a 
mean 
602.6 
3202.3 
0.479 
std dev 
23 
51 
0.019 
Table 2 : retrieved parameters, mean and standard deviation, after assimilation of noisy observations, for 50 
experiments. GPP and NPP values are in g C m-2 year-1. 
The biases on parameters are lower than 4% as are the biases on resulting GPP (1000 g C m-2 year -1 instead 
of 1010 for reference) and NPP (623 g C m-2 year-1 instead of 625) The standard deviations of the 
parameters and annual carbon fluxes are satisfactory, and the CC>2 fluxes are greatly improved (in amplitude 
and phasing) compared to the fluxes computed with a priori parameters chosen for the first guess. Theses 
results show that the assimilation scheme is stable, and also that the simplex algorithm statistically finds the 
best agreement between model and 'observations'. It confirms also that a temporal knowledge-based link 
between satellite observations at different dates can yield to accurate LAI retrieval even when high LAI values 
and noisy measurements are considered. Non linear systems, as this vegetation model, can easily generate 
local minima in the cost function and then prevent efficient assimilation of the observations. We therefore 
investigated the sensitivity of the adjustment to the first simplex of the iterative minimization. The first guess 
parameters were the 36 combinations of : 166 / 333 / 500 / 666 °C for Zj, 2500 / 3333 / 4166 °C for £2 ^ 
0.35 / 0.60 / 0.85 for a. These guesses correspond to severe a priori errors. The results of adjustment to a 
unique noisy 'observations' set show that in 31 cases, a satisfactory minimum was found, but the simplex 
algorithm was trapped in a local minimum in 5 cases ; This can be avoided by different methods: systematic 
minimization with a set of different first guesses, random-search algorithms, but it also highlights the irregular 
behavior of this kind of model. 
These different results show that control by remote sensing data seems to be possible, for this 
vegetation model, which is designed for global scale satellite assimilation. Moreover, assimilation can improve 
both amplitude and phasing of terrestrial carbon fluxes. However, two important questions were not addressed 
in this experiment: the ability of the model to simulate real vegetation, and the accuracy of reflectances 
modelling from vegetation model outputs (in particular, perturbations of satellite measurements are not 
expected to be gaussian and to have a zero mean). 
4 - DISCUSSION - CONCLUSION 
The theoretical considerations and results of a synthetic study suggest that controlling a prognostic vegetation 
model by satellite data can improve the accuracy of the model outputs. This is a new approach in the field of 
vegetation functioning studies, and some remarks can be made to point out particular questions. The method 
that we propose is a 'model to satellite' approach (Hollingsworth 1990), insofar as it uses direct modelling from 
the vegetation model to the remote sensor. One can try to retrieve, in a first step, some biophysical variable, 
weekly LAI or APAR for instance and to find agreement between the vegetation model course and these 
retrieved variables (a step towards 'satellite to model' approach). In theory, the first method should lead to 
better results, avoiding errors in LAI inversion. In fact, inversion of physically based reflectance models has 
given interesting preliminary results (mainly over crop canopies), but application to natural vegetation and 
coarse resolution data raises additional difficulties. The choice between these different strategies which can be 
expressed as ' where to perform the inversion ' depends essentially on the errors existing in the observations, 
in the reflectance model and in the vegetation model. Clearly, further works should be devoted to the analysis 
of these errors. 
For this assimilation experiment, we perform minimization of the difference between model and 'observation' 
with a least square criterion, which is suited for gaussian perturbations. However, the data set may contain 
outliers, for instance residual clouds may cause non realistic high reflectances. Different cost functions can be 
used (even non symmetric functions). Moreover, it is possible to assign different a priori weights to the 
different observations: for example, some view and illumination angles lead to large error in modelling 
atmospheric perturbations (related to long optical path) and surface bi-directional reflectance. These 
geometrical configurations also causes large pixel deformations and contains confusing information. A priori 
knowledge about these errors can be taken into account when estimating, in the misfit function, the specific 
contribution of a couple of observed-modelled reflectances. The statistical distribution of noise observation 
and reflectance model error can be assessed by simulation studies based coupled surface-atmosphere