492 
the four SSMyl frequencies) rises some difficulties for several reasons: on one hand, the high variability of the 
different atmospheric conditions makes it extremely cumbersome to fix absolute thresholds, and, on the other 
hand, a high degree of correlation is observed between multispectral measurements. 
Figure 2. Flow-chart of the double-clustering discriminant (DCD) algorithm for identification of r ainf all 
areas from space-bome multi-frequency microwave radiometers. 
This has suggested the extraction of the significant information by means of a Principal Component Analysis 
(PCA) of the unpolarized T B 's (Basili et al„ 1992). The transform can be expressed by the following: 
PCj = aii AT B (19) + a i2 AT b ( 22) + a i3 AT B (37) + a i4 AT b (85) (1) 
where PCj is the i-th principal component, ajj's are the transform coefficients derived from the eigenvectors of 
the covariance matrix of the simulated T B 's, and AT B represents the difference between each unpolarized T B 
and its average value. In the specific case, the first principal component accounts for the 98.9 % of the total 
variance, while the first and the second PC's cumulatively explain the 99.8 % of the variance. By analyzing 
the relationships between the first two principal components and the corresponding r ainf all rate values 
belonging to the simulated cloud-radiation dataset, we derived the thresholds for classifying precipitative 
events within the SSM/I images. Specifically, for precipitation over land the condition is PCj < 40 K and PC2 
< 12 K, while for precipitation over ocean we assume PCj < 50 K and PC 2 > -13 K. These conditions have 
been chosen imposing that the probability of surface rain rate above 3 mm/h is greater then 0.97. 
The third step of the DCD procedure finally transforms the space-borne radiometric data as in 
Eq. (1), using the initial guess of the average T B 's, determined in the second step. Then the thresholds on PCi 
and PC 2 are applied to derive the binary mask consisting of a precipitation or no-precipitation flag. 
Figure 3 shows the rainfall areas identified by means of the DCD procedure from the SSM/I 
image, shown in Figure 1. This mask has been compared to those obtainable by applying the SI and PCT 
techniques and a good agreement has been found, even though the SI and PCT thresholds have needed to be 
further adapted to the Mediterranean area. The dashed box puts in evidence the Liguria precipitation event, 
for which the retrieval results will be shown in the next Section.