478 
correlated to surface temperature as infered from thermal infrared measurements, especially at high 
frequency (above 10 GHz). Passive microwave measurements could hence provide a unique tool to 
assess surface temperature, when dense cloud cover mask surface features to thermal infrared sensors. 
As described above, many parameters drive the microwave emission of land surfaces. In order to 
retrieve key parameters related to the global energy and water balance (such as soil moisture content, 
surface temperature and hydric vegetation characteristics) from passive or active microwave 
measurements, various methods have been developed. They are based on either statistical techniques 
(most often regression analysis), or on forward model inversion. In this paper, a review of the different 
approaches of the retrieval problem is made. The review focuses on vegetation and soil monitoring 
using passive microwave measurements. It must be noted, however, that such techniques are widely 
used by the remote sensing community. 
2. RETRIEVAL METHODOLOGY 
Two types of approach can be used to retrieve land surface parameters from spectral measurements. 
Most of the time, a forward model is created (input = land surface parameters; output = remote sensing 
measurements) and manipulated to yield an inverse model. Another approach is to create an explicit 
inverse of the physical process (input = remote sensing measurements; output = land surface 
parameters). This second approach often suffers from many-to-one problems while forward models 
can accurately model causal relationhips (Hwang et al, 1992; Ishimaru et al., 1992). The retrieval 
methodology requires two steps: selecting first a forward model and then a method of inversion, which 
are both relevant to the specific retrieval problem. 
2.1 Forward models 
Following Hwang et al. (1992), we classify forward models in four main categories: parametric and 
non-parametric data driven models, implicit functions and explicit functions. Data driven models can 
efficiently retrieve information, when no simple approach can be found to describe complex 
relationships between input and output data. More specifically, non-parametric data driven models 
(NPDD-models) do not require a priori knowledge of the functional form for the process which is 
modeled. Such models are used as black box, which fit any sets of input-output data and are easy to 
inverse. Many techniques can be used to build NPDD-models: linear regression analysis, principal 
component regression, neural networks, etc. 
Parametric data driven models (PDD-models) can be derived from empirical, semi-empirical or 
physical considerations. Prior to the inversion process, the parameters must be adjusted to minimize 
the squared error between the data and the actual output of the model. This step requires a good 
characterization of the land surface, using ground measurements. As an illustration of semi-empirical 
approaches, Wang and Choudhury (1981) take into account soil roughness, using a two-parameter 
modeling: a roughness parameter h r and a polarization mixing parameter Q (h.Q-model). With regard 
to vegetation modeling, in the low-frequency region, multiple scattering effects can be neglected. So 
vegetation emissivity can be modeled using a simple radiative transfer approach based on two 
parameters (Mo et al., 1982; Kerr and Njoku, 1990): the single scattering albedo co and the canopy 
opacity x (x.co-model). As frequency increases, multiple scattering effects are less and less negligible 
and can be taken into account by the phase matrix P. The continuous medium concept is a physicaly- 
based approach to calculate P, using two correlation lengths l z and l p which parameterize permittivity 
fluctuations inside the canopy layer (Tsang and Kong, 1980). As l z and l p are synthetic parameters that 
cannot be easily related to vegetation characteristics, they must be calibrated, prior to the inversion 
process (Calvet et al.. 1993; Wigneron et al, 1993a).