480 
So for such applications, the use of the SIA requires prior efforts to develop tractable models with few 
input parameters, whose statistical behavior is well known (in terms of spatial and temporal 
variations). 
With respect to minimization problems in SIA, the adjoint technique has proved to be a very efficient 
tool in the minimization process of cost functions. It has been associated to the variational approach for 
assimilating remote sensing measurements in meteorology (Thepaut and Moll, 1990). Variational 
methods are used in data assimilation schemes for numerical weather prediction. As for the SIA, they 
consists in minimizing a cost function which measures the distance between the analysed atmospheric 
state and a set of available function. The adjoint technique is implemented to compute first and second 
derivative of the cost function without requiring additive simplifying assumptions about the forward 
model. These terms can be used both to evaluate the number of independent pieces of information 
contained in the observed data, and to minimize cost function dedicated to inversion problem. 
An alternative approach to the SIA is to use an artificial neural network (ANN). First an appropriate set 
of input-output data is generated, using the forward model. Then a copy of the forward model is made 
by training the ANN on the set of data. The ANN is hence able to capture very complex and non-linear 
behavior within its self-organizing connections (Garett et al., 1990). When forward model simulations 
require a large amount of computer time, an advantage of the ANN technique is that once the ANN is 
trained, the inverse problem of finding the surface parameters from observed measurements can be 
accomplished quickly (Ishimaru et al., 1992). Furthermore, several ANN inversion methods are 
available. For instance, the constrained iterative inversion exploits relationship between neighbouring 
microwave measurements (Zurk et al., 1992). It also gives the ability to incorporate additional 
information (a priori information, ground data, etc.) into the iterative inversion process (Davis et al, 
1993). It can therefore be very efficient for mapping problem. 
3. APPLICATIONS 
Several studies rely on passive microwave measurements over vegetation cover, to retrieve land 
surface parameters, mainly surface temperature, volumetric soil moisture (m 3 /m 3 ) and the vegetation 
water content (kg/m 2 ). A brief survey of some applications are presented in the following sections. 
They will be classified according to the forward modeling approach which is used: statistical or 
functional. 
3.1. Statistical approaches 
Land-surface type classification using microwave measurements rely most often on statistical analysis. 
Based on observations of SSM/I data and brightness temperature thresholds, several rules were 
developped to classify surface types: dense vegetation, standing water, agricultural fields, dry and 
moist bare soil, etc. (Neale et al., 1990). Similarly, Wegmiiller (1993) and Paloscia and Pampaloni 
(1992), used combination of microwave parameters, as spectral and polarization differences, to 
perform a clear separation of several crop types: wheat, corn, sugar-beet, potatoes, grass, etc (figure 1). 
Regression analysis can also be used to retrieve land surface parameters. As mentioned above, and 
assuming linear relationships, microwave soil brightness temperatures and backscattering coefficients 
can be used to retrieve soil volumetric moisture (Prévôt et al., 1993), or precipitation index (Teng et 
al., 1993). McFarland et al. (1990) retrieved land surface temperature using linear combinations of the 
four spectral channels of SSM/I, with an accuracy in the 2 - 2.6K range. 
Numerous studies rely on microwave indices (especially the polarization difference PD=T Bv -T Bh , 
difference between Vertical and Horizontal Brightness temperature at a given frequency, or the 
polarization ratio PR (PR=(T Bv -T Bh )/(T Bv +T Bh ))), which are good indicator of vegetation biomass. The