484 
In this frequency range, the x.co-model, which do not take into account P, could still be used provided 
that vegetation elements have a very strong forward-scattering pattern (Joseph et al, 1976). If this 
condition is not fulfilled, the physical meaning of retrieved values of x and co is difficult to establish. 
For instance Pampaloni and Paloscia (1986) and Van de Griend and Owe (1993) neglecting P at 
37GHz, retrieved very low values of the single scattering albedo œ in comparison with measured 
values obtained by Matzler (1989). 
To calculate P, the already-mentioned discrete and continuous modeling of the vegetation medium can 
be used. In the continuous approach, instead of x, to and b the vegetation is characterized by the 
correlation lengths l z and l p , the volumetric density fracV (m 3 /m 3 ) and the gravimetric moisture M g 
(kg/kg) of the vegetation material. The basic principle of inversion is similar as for the x.co-model. 
Prior calibration of l z and l p are performed using ground measurements. Then assuming that the 
correlation lengths can be considered as constant during the inversion process (Calvet et al, 1993; 
Wigneron et al, 1993a), geophysical parameters of interest such as soil moisture, vegetation density 
and surface temperature are retrieved (figure (3)). 
Very few studies based on the discrete approach deal with retrievals. The problem is complex due to 
the numerous input parameters of the model that should be retrieved. Recently, inversion of the active 
model MIMICS has been carried out, using neural networks (Pierce et al, 1992). Instead of 
performing direct inversion of discrete models, composite approaches can also be used. Prévôt et al. 
(1992) adjusted a parametric model, dedicated to retrieve vegetation and soil parameters, using 
MIMICS simulations. Wigneron et al (1994), developed a composite discrete-continuous approach. In 
the continuous approach, if the modeling of scattering (scattering coefficient k s and phase matrix P) is 
satisfactory modeling of absorption (absorption coefficient kj is not accurate and is not expressed as a 
function of the correlation lengths (l z , l p ). As a consequence, accurate calibration of the correlation 
lengths is difficult, especially a low frequencies (Wigneron, 1993a). Consequently, it was proposed to; 
(1) express k a as a function of to and k s (l z , l p ) 
(2) use a discrete approach and relevant a priori information to calculate the scattering albedo (0 (figure 
(4)). 
This composite model was implemented to retrieve both soil moisture and vegetation density using 
ground-based microwave measurements over a wheat field (Biard, 1993). 
4. DISCUSSION AND CONCLUSION 
At the present time, no algorithm is available to retrieve geophysical parameters from passive 
microwave spacebome measurements over vegetation-covered areas. The main reason is that no sensor 
provided the scientific community with relevant set of measurements, that could be used to train 
inversion algorithms. Relevant sensors should operate at low frequencies (in the 1-10 GHz range) with 
sufficient spatial resolution to be able to scan both soil moisture under vegetation cover and the 
vegetation layer itself. For instance, the lowest frequency of the passive microwave sensor SMM/I is 
19GHz and the ground resolution of the order of 50 km. Considering this, most of the inversion 
algorithms have been tested using measurements performed by ground-based or airborne sensors. So 
the algorithms could not take advantage of a priori information, that are provided by long-term 
temporal evolution of spatial measurements. Furthermore, scaling problems that are very accute if we 
consider the poor spatial resolution of the passive sensors could not be adressed. 
Difficulties to obtain accuracies needed by the end users are due to the large number of parameters 
influencing the microwave emission behavior of land surfaces (Pulliainen et al, 1992). To overcome 
these limitations, microwave modeling must take into account with more considerations the inversion 
problem. Statistical approach are tractable means to perform retrievals of land surface parameters 
when the detailed form for the process which is involved is unknown or very complex. Hence, they are