491 
measurements to the spatial resolution and sampling density of the high resolution 85-GHz channel, balancing 
the tradeoff between resolution and noise. 
Figure 1. Vertically-polarized Tb's at 85.5 GHz measured by SSM/I on September 27, 1992 at 16:15 IJTC 
dining the DMSP FI 1 satellite ascending pass over the Italian peninsula. 
3.2. Identification of rainfall areas from satellite data 
The r ainf all identification must be taken into consideration when performing precipitation retrieval since 
different kind of backgrounds may have radiometric microwave signatures very similar to that one of rainfall 
events. A double-clustering disc riminan t (DCD) technique has been developed in order to identify the rainfall 
area from satellite radiometric images and its flow-chart is shown in Figure 2. The DCD procedure is based on 
the simulated cloud-radiation dataset, and tries to overcome some of the limitations connected to current 
discriminating algorithms, specifically the Scattering Index (SI) method (Grody, 1991) and the Polarization- 
Corrected-Temperature (PCT) technique (Spencer et al„ 1989). 
The first step of the DCD algorithm is identifies the type of surface being observed by the 
space-borne radiometer, since precipitation appears as a cold microwave source over land, while the opposite 
occurs over sea. To this aim, we have used the brightness temperature data of the lower SSM/I frequency, 
(i.e., the 19 GHz ch ann el) In fact, a simulation analysis through the radiative transfer model showed that, 
even for relative high rainfall rate values (about 20 mm/h), measurements made at this frequency were still 
able to identify the surface type. Thus, an unsupervised classification method is applied to the 19 GHz data, 
based on a nearest-centroid clustering analysis that groups the image pixels into two classes (Richards, 1986). 
The association of each cluster to a land or a sea background is based on the value of its centroid: if it results 
greater than 200 K , the cluster contains land pixels, otherwise the opposite. The threshold value has been 
fixed taking into account possible misleading coverages, such as snow over land and clouds over sea. 
The second step regards the computation of a first guess of the average Tb's within a 
precipitation area that will be used in the third and last step of the procedure. For this purpose, a method 
similar to the previous step is used, but considering the 85 GHz c hanne l First, a clustering algorithm is 
applied to pixels belonging to an homogeneous background and then thresholds are used to preliminarily 
identify the clusters of precipitation pixels. The thresholds have been fixed to 200 K for land background and 
to 180 K for sea. Therefore, the pre liminar y evaluation (first guess) of the average Tb's can be computed for 
both the background classes. 
Before describing the third step that identifies the actual precipitation areas within the two 
classes of background, it must be mentioned that the use of the four-dimensional T B vector (corresponding to