The resulting matrices can be evaluated from the statistical indices such as the POD, FAR, CSI and EQ_PC. A high EQ_PC parameter, greater than 40%, indicates that radar and satellite datasets are in accordance, the calibration process is equilibrated with respect to the number of radar and satellite points and resulting rainy cells in matrices have a significant rain probability. The best spatial correspondence is obtained for the Spanish 3-D array (table 4.5) due to the highest POD, CSI, EQ_PC and also the lowest FAR. In second place comes the Baltic 3-D matrix (table 4.6) by taking into account
again the POD, FAR and CSI indices but with a low EQ_PC. This last factor may lend significance to the other three indices but unfortunately we are not able to quantify how much. Poor results are generated from the two 2-D arrays (table 4.3d and 4.4), where the Baltic Sea matrix should be considered with caution because it provides the worst statistical indices and the lowest EQ_PC. Correlation coefficients are around 0.4 for both 3-D matrices and around 0.3 for the 2-D arrays. These results suggest, as commented by Lovejoy and Austin (1979) and Cheng et al. (1993), that visible data in matrices can improve their accuracy for rain estimation purposes in regard to pure infrared methods. However, other negative factors may affect the rainfall estimation by taking into account visible radiances such as areas affected by shadows or extremely bright cloud zones due to direct reflections. Significant variations can be observed in the intensity of this visible radiation in the early mornings and in the late afternoons. These variations are produced in many cases by the effect of normalization of the visible pixels with cosine of the solar zenith angle in these hours of the day. In fact, noisy rainy spots with significant rain rate values during daytime are observed mostly when solar zenith angles are large. Under such conditions the normalisation process tends to over-measure visible cloudy brightness counts causing errors in the rain class extraction from the 3-D matrix. This effect can easily be avoided by limiting the use of the 3-D matrix to a smaller number of daytime hours; however this way increases the time for the employment of the 2-D matrix, which is less accurate.
One key question raised in this work was the selection of the calibration period to perform a long term calibration matrix. By selecting a considerable amount of radar and satellite data to develop the matrices over the Iberian Peninsula, statistical indices evolved into acceptable boundaries and rain classes have a tendency to converge to specific values and specific array positions. However, Baltic array classes have a tendency to drop to zero and to disperse to a broader distribution. That is presumably caused by different reasons: First, the geostationary satellite pixel is degraded due to remapping requirements in high latitudes. Second, the scarcity of convective rain cases in contrast to the stratiform over the Baltic area, even during the summer period, could be a negative factor in the calibration process. Stratiform precipitation is very difficult to be accurately detected from satellite and therefore radar and satellite rain rates do not match correctly causing a substantial drop in the probability of precipitation. A shorter calibration period seems to be important to develop the best Baltic arrays.
The present research confirms that TWV can exceed TIR in several degrees
according to observations made by Schmetz et al. (1997). These special satellite measurements in which TWV > TIR clearly corresponded to areas in our matrices where
TIR-TWV < 0. Moreover, the highest rainfall classes are obtained in rainfall arrays for the
coldest TIR and places in which TIR-TWV < 0 condition is satisfied (see tables 4.3d, 4.4, 4.8
and 4.9). This result is, also, in agreement with the work of Kurino (1997a) and shows the advantages in the use of brightness temperature differences between the 11 μm and 6.7 μm bands to detect deep convective clouds systems accompanied by heavy rainfalls.
Independent and qualitative studies not shown in this thesis report were performed for thin cirrus and stratiform clouds with precipitation over Spain. In the Baltic area the results of these studies were not definitive. The difference between TIR and TWV seems to
be adequate to eliminate most of the cold thin cirrus clouds with no precipitation because we have observed that TIR-TWV is greater than +3ºC for these kind of clouds and thus, they
are outside the rain area in matrices. However, poor results for stratiform precipitation using CRR estimations are obtained. Stratiform rain cloud tops are not as cold as convective tops and stratiform radar rainfall signal are found, also, in the positive area in matrices where TIR and TWV difference are above or close to zero and rainfall probability
is lower than the EQ_PC in most of cases (see table 4.3c). Other problems are related with radar rain pixels that are located outside the rain area as defined by the CRR matrices. In such cases the rain is probably produced by low clouds with warm cloud tops in the infrared band and/or with a TIR-TWV great enough to be missed from the rain region
of matrices.
The 3-D calibration table performed over Spain shows the greatest rainfall classes for high values of visible counts (Vc from 200 to 232 approximately). This can be explained taking into account that clouds with heavy convective cores and obviously, very thick, as studied by Vicente and Scofield (1996), are usually precipitating clouds in radar images and therefore they should be stored in the matrices as very bright clouds in the visible band. However, the 3-D matrix generated over the Baltic countries have smaller rain classes along lower visible brightness values (Vc from 168 to 204 approximately). This could be caused by two factors. Solar radiation impinges against the earth’s surface and clouds with larger solar zenith angles in these high latitudes or the direct and reflected visible radiation has to pass through a longer atmospheric path, increasing energy absorption, before the satellite sensor is reached. The first factor should be corrected by the normalisation process of the visible pixel (Binder, 1988) but the
second factor seems to be the most suitable for explaining any reduction of visible radiation over such high latitudes.