a. Level 1.5 channel data

Part I contains much information relevant to the present study; that pertinent to MSG SEVIRI channels that contain reflectance information is discussed herein. VIS, NIR, and IR data at 3-km sampling distance, with HRV at 1-km sampling distance, are collected by SEVIRI on the MSG (specifically Meteosat-9) satellite as located over the equator at 0° longitude. The VIS, NIR, and IR (HRV) resolution is 3 (1) km. Over the COPS region the along-pixel resolution is near 5–7 (2–3) km due to skewing effects, while cross-pixel resolution is near the nadir value. SEVIRI possesses eight IR channels, with the central wavelengths being 3.9, 6.2, 7.3, 8.7, 9.7, 10.8, 12.0, and 13.4 μm; two NIR (0.8- and 1.6-μm channels); one VIS centered on 0.6 μm; and the HRV (Schmetz et al. 2002; EUMETSAT 2007). All data used for this project were collected between 0815 and 1900 UTC,¹ daylight hours over the COPS region in summer, early June–early August.

Reflectances at 0.6, 0.8, and 1.6 μm (Refl_0.6, Refl_0.8, and Refl_1.6, respectively), as well as HRV brightnesses, are normalized by solar zenith angle (A_sz) by [1/cos(A_sz)]. Here, A_sz is a function of latitude, longitude, and time of day. This normalization is done to avoid a diurnal signal in the results. Lindsey et al. (2006) and Rosenfeld et al. (2008) describe a procedure for computing the 3.9-μm reflectance, while Rosenfeld et al. (see also Setvák and Doswell 1991; Lensky and Rosenfeld 2008) present the method specific to MSG. In addition, A_sz is required for obtaining the 3.9-μm reflectance which is considered A_sz normalized.

With cloud transmissivity assumed to be ∼0 for optically thick clouds, 3.9-μm-specific reflectance (Refl_3.9) is found by

View Expanded

In Eq. (1), R_total is the total radiance, R_therm is the CO₂-corrected thermal component of the radiance, and S is the top-of–the-atmosphere solar constant, for the 3.9-μm channel. The units for R_total, R_therm, and S are [mW m⁻² sr⁻¹ (cm⁻¹)⁻¹] (sr = steridian). Here, Refl_3.9 is a percentage, 0–100. The value of R_therm is computed (assuming cloud emissivity is 1) as

View Expanded

where the first term {R[IR_3.9, T_B(IR_10.8)]} represents the procedure of converting the IR brightness temperature (T_B) of the 10.8-μm channel back to radiance using essentially the Plank function, given as

View Expanded

Coefficients A and B are those for SEVIRI’s 3.9-μm channel (0.9959 and 3.3592, respectively). Also in Eq. (3), C₁ and C₂ are 1.191 04 mW m⁻² sr⁻¹ (cm⁻¹)⁻⁴ and 1.438 77 K (cm⁻¹)⁻¹, respectively, and υ_c is the central wavenumber of the channel being processed. In Eq. (2), C_R3.9 accounts for the CO₂ absorption as given by

View Expanded

In Eq. (4) T_B10.8 and T_B13.4 are the 10.8- and 13.4-μm T_Bs, respectively. In Eq. (1), S is found by

View Expanded

where D is the Earth–sun distance and A_sat is the satellite zenith angle.

Data at time intervals of 15 min are processed. (Note that although 5-min “rapid scan” data were available during a portion of COPS 2007 from MSG, previous CI research utilizes 15-min time-trend fields, and hence the decision was made to follow past work.) All processing focused on the main updrafts of cumulus clouds, and hence clouds that are considered “optically thick” where the transmission from below is considered negligible, ensured by the choice of the coldest pixel. Beginning with a small 3 × 3 pixel region centered on growing cumulus clouds, the coldest IR pixel at 10.8 μm is labeled the main updraft (i.e., the coldest ∼10%, or one of nine pixels). When defining 15- and 30-min time-trend fields, this coldest pixel is monitored in three successive 15-min images. Collocation is then needed between the 3 × 3 VIS–NIR pixels that encompass a cloud (per time and per image), and the 9 × 9 HRV pixels that cover the same region. Once accomplished, the 9 × 9 HRV region is processed across all images to obtain the HRV-related fields. In addition to 10.8 μm, the 13.4-μm channel [used in Eq. (4)], is our only use of IR data in this study.

b. Derived fields

Two derived fields were created from the HRV data that describe the BV across a region of developing cumulus clouds: 1) texture estimation and 2) peak detection (PD). In the correlation analysis to follow, both derived BV fields are tested independently of other reflectance channels to identify the presence of visibly bright cumulus clouds against a darker background of ground surfaces under clear-sky conditions, more stratiform-like clouds, and/or the variability in brightness across a small area of active cumulus cloud turrets.

For both BV fields, 15 additional fields are developed using the 9 × 9 pixel HRV data covering each growing cloud over the three 15-min time periods. Average, maximum, and perturbation BVs are formed for five periods within the three-image set for all 123 convective storms. These five periods are (a) the present image time, (b) the image 15 min in the past, (c) the image 30 min in the past, (d) the 15-min time trend [(a) − (b)], and (e) the 30-min time trend [(a) − (c)]. Both of these 15-min field sets are compared and one form of BV is chosen for further analysis with respect to understanding attributes of growing convective clouds.

For texture (T_var) in the HRV data, the variance-based method of Berendes et al. (2008) was followed given the small (9 × 9 pixel) region being operated on that covers a 3 × 3, 3-km VIS–NIR region. Berendes et al. (2008) used texture as one component in the convective cloud mask, cloud typing or identification algorithm used within the SATCAST system. This can be written as

View Expanded

where G is the number of gray levels in a scene (255 in this application of the HRV data²) and D_υ is the “difference vector” that represents the arithmetical difference between scaled (0–255) 1-km HRV brightness values between two points. This process effectively accumulates the absolute value of the pixel brightness differences in three directions: horizontally, vertically, and diagonally across the 9 km × 9 km HRV region.

For another means of identifying the BV, PD was determined as a difference of Gaussian (DoG) filter approach was applied to HRV data. We followed the methods outlined in Iskenderian et al. (2009) and Wolfson and Clark (2006) for the quantification of convective clouds in advance of CI. For this procedure, a two-dimensional DoG filter was run [with σ = 0; Marr and Hildreth (1980)], essentially smoothing out the brightness peaks and increasing the brightness in low-brightness regions. Areas with more uniform brightness before applying the DoG filter (e.g., stratiform clouds) remain relatively unaffected. The smoothed image was then subtracted from the original, resulting in an image with no strong cloud-edge gradients, with most of the remaining BV being that associated with “lumpy” convective cloud features. Very high brightness gradients associated with cloud edges are reduced, something that is not accomplished in the texture analysis above.

From these BV values, three fields were created in hopes of representing the utility of this derived quantity. These are the “maximum,” “perturbation,” and “average” BV. Table 1 lists the attributes of the growing convective cloud regions that these three variations of BV should describe. Perturbation BV was computed by taking the difference between the actual pixel value of BV and the average BV (i.e., BV_pert = BV_pixel − BV_ave). Maximum BV (BV_max) values are simply the largest point value of BV within the 9 × 9 pixel HRV region, whereas BV_ave is the average BV in a 9 × 9 pixel region.

After evaluation of a number of cases, the choice was made to focus on the DoG filter PD approach. Two reasons for using the DoG filter method are (a) spurious information associated with sharp edge effects, when bright clouds appear against a dark background, will be suppressed compared to analysis using the texture approach, and (b) cumulus clouds developing in otherwise clear skies, compared to those among a more continuous deck of stratocumulus (for example), will appear more similar in the PD fields; the texture analysis would tend to show these two cases with largely different textural magnitudes.

Figures 1a–d shows how the filter-derived PD is implemented, as in this example, for one time for one growing cumulus cloud event. In Fig. 1a, HRV data covering a 36 km² region with clusters of developing cumulus clouds (bright colors, highlighting the higher brightness) are shown. Data in Fig. 1a are scaled from their original 0–255 brightnesses to 0–120 for better illustration. Figure 1b shows PD (Fig. 1a) after the DoG filter has been applied, while Fig. 1c highlights the PD_pert field. In Fig. 1b, regions of convective clouds are highlighted, whereas the surrounding regions are dimmed, and the scale has been reduced by a factor of ∼2.67 (values now range from 0 to 45). In Fig. 1d, the single pixel possessing PD_max in the filtered image is highlighted (arrow in Fig. 1d), which is that used in the analysis to follow. The process shown in Figs. 1a–d is repeated for every 15-min image in the three-image sequence for all 123 cases analyzed herein.

From the HRV and VIS–NIR reflectance channels, 27 separate so-called interest fields are formed. The 27 total fields are listed in Table 1: 12 reflectance and 15 PD fields.

c. COPS 2007 and case identification

The same COPS dataset used in Part I is analyzed here. The more important components of these data will be summarized. The 2007 COPS campaign (Wulfmeyer et al. 2008) was conducted over southern Germany and northeast France from 4 June to 31 August 2007 (information online at www.cops2007.de). The geographical domain was 46.5°–49.0°N, 6.5°–11.0°E. One main science goal driving COPS was to increase our understanding of CI as induced in part by orography (Wulfmeyer et al. 2008), and therefore a host of observational data were collected and archived. Two fields in addition to Meteosat-9 data were used from this archive, those being soundings (collected by the Institut für Meteorologie und Klimforschung at the Universität Karlsruhe) and dual-polarimetric radar collected in Waltenheim-sur-Zorn (48.73975°N, 7.61033°E) by the Polarization Diversity Doppler Radar (POLDIRAD) operated by the Deutsches Zentrum für Luft und Raumfahrt (DLR), Institut für Physik der Atmosphäre. Data from several other radars operating over the COPS domain (Karlsruhe, Supersite Rhine Valley, and Supersite Vosges Mountains) were used similarly.

Fifteen days during COPS possessed deep, growing, more isolated cumulus clouds occurring in conditions unobscured by higher clouds. Twenty-six of 88 COPS days were evaluated as they possessed active CI. Table 2 presents the days analyzed, the times on each day over which CI was observed, and the numbers of storms identified and processed. All events were identified by a human expert in channel 9 (10.8 μm) and HRV Meteosat-9 data. A “high quality” event is defined as clouds growing in mostly clear-sky conditions such that cumuli over a 30-min period could be observed preceding CI without “interference” from higher-level altostratus, altocumulus, or cirrus, which can lead to false identification. For a given good quality day, all 15-min imagery were gathered, and a subset domain for the COPS region was processed from the 3712 × 3712 IR and 5568 × 11 136 HRV full-disk data. This amounted to processing IR images of 131 elements by 61 lines, and HRV images of 393 × 183.

Exactly as in Part I, the goal was to capture cumuli in the mediocris (moderate) stage (just beyond the “fair weather” stage) at the initial time, over three successive 15-min images (see Fig. 1 in Part I). Therefore, by the third image in the sequence (e.g., cumulonimbus calvus), a new anvil may exist and rainfall may nearly be occurring. Data from the POLDIRAD and other radars were viewed in their native resolution as a means of confirming CI, and for this study, radar reflectivity simply had to be ≥35 dBZ to be classified as a growing cumulus–CI event. The actual dBZ value ≥35 that occurred with a given new storm was not recorded, as this was not part of the study’s goal. Sounding profiles were used to make parallax corrections to the MSG imagery so optimal radar-to-satellite comparisons could be made.

Figure 2 shows examples of the 10.8-μm T_B (Fig. 2a), Refl_0.6 (Fig. 2b), Refl_1.6 (Fig. 2c), and Refl_3.9 (Fig. 2d), as well as HRV imagery (Fig. 2e), over three times (1100, 1115, and 1130 UTC), for a growing cumulus cloud field on 8 June 2007. For Fig. 2, a 3 × 3, 3-km pixel-size region in the VIS–NIR and a 9 × 9, 1-km pixel-size region (boxes in Fig. 2) are shown to emphasize the data locations processed as convection grows. From Fig. 2a, the coldest (darkest) 10.8-μm pixel in the 3 × 3 box is highlighted with a plus sign (+), which is the location where the interest fields are developed. For the PD analysis, data in the entire 9 × 9 pixel box (Fig. 2e) are processed. The reflectance differences and time changes between the 0.6-, 1.6-, and 3.9-μm channels are substantial in Fig. 2, and this will be discussed below.

a. Reduction and redundancy

For determining criteria 1 and 2 above, the following procedures were followed. A 27 × 27 correlation matrix was analyzed. Cases where field-to-field cross correlations were >0.70 were highlighted. High cross correlations between three or more fields were also identified. The result was the removal of several fields that contained redundant information. At this point, the dataset was split into 15 PD and 12 reflectance fields, considering that these explain different attributes of the convective scene.

High correlation (>0.98) was found between all 0.6- and 0.8-μm fields, and both 15- and 30-min time trends (e.g., 30-min trends in 0.6 and 0.8 μm). Therefore, the decision was made to not consider all 0.8-μm-related fields, which resulted in removing 3 of the 12 reflectance fields. Use of the 0.6- versus 0.8-μm data was chosen because the latter are more proven in previous studies as a means of estimating the optical depth in clouds (Nakajima and King 1990). The remaining nine interest fields were then subject to principal component analysis (PCA; Johnson and Wichern 2002) as described below.

Very high correlations (>0.95) were also found between all three types of PD (PD_pert, PD_ave, and PD_max, in a 9 × 9 HRV box), as well as between the two time trends for all three BV fields. After consideration of the differences and attributes of PD_max and PD_pert, it was determined that PD_ave would be most valuable. This is supported by Wolfson and Clark (2006), who use PD_ave, and hence this is a more proven field. Using only PD_ave resulted in the loss of 10 fields from the original 15 HRV-related fields. Table 3 presents the remaining, less correlated fields.

b. PCA

Given nine reflectance and five BV interest fields, PCA was subsequently performed on each separately. The goals of using PCA are to help further quantify field redundancy and to establish a level of field uniqueness for describing the given outcome or event (substantial cumulus cloud development prior to CI). Given that the units of all interest fields were either reflectivity, the change in reflectivity, PD, or the change in PD, data normalization was done in this PCA using the correlation matrix [as opposed to the covariance matrix; Wilks (2006)]. Without the normalization the correlation matrix provides, data with larger magnitudes (e.g., K versus reflectances from 0 to 1) will artificially contribute more to the explained variance.

Table 4 presents the PCA results for reflectance, while Table 5 presents the PD fields. Rankings are determined from the PCA, specifically the size of the eigenvalues in principal component 1 and, if needed, component 2. The rankings may be used as a means of weighting each interest field in terms of importance when describing a given attribute or process, per MSG pixel, or per cumulus cloud object. Later, suggestions are provided on how to utilize these fields in 0–1-h CI nowcasting applications. From Table 4, the 30-min trend in Refl_3.9 is the most important field when describing convective cloud development. This is followed by the 30-min trend in Refl_1.6, the 15-min trend in Refl_3.9, the 15-min trend in Refl_1.6, and the instantaneous Refl_3.9 and Refl_1.6. From Table 5, PD_ave 15 and 30 min prior to the latest image time are the two most important fields, while the present-time PD_ave values are third in importance.

As a means of verifying that the results in Tables 4 and 5 are robust, the correlation and PCA analyses were repeated using a dataset in which a random 20% of the storms were removed. The PD fields were unchanged from those in Table 5, and therefore Table 6 shows the reflectance fields only when 100 events were analyzed. Little overall change is noted. Fields that ranked in the top three for the whole dataset remained of the same importance. The rankings for the 15-min trend in Refl_1.6 and Refl_3.9 switched, and instantaneous Refl_1.6 replaced the Refl_0.6, as shown in Table 4. Given the similarities between Tables 4 and 6, the top five reflectance fields are shown to be robust.

c. Physical attribute relationships

Table 1 briefly summarizes the physical interpretation of a given interest field with respect to growing cumulus clouds. Specifically, however, what does each interest field provide that is unique to describing growing cumulus clouds prior to CI? Other questions from the results in Tables 4 –6 include the following: 1) What do the Refl_0.6, Refl_1.6, and Refl_3.9 provide in such an analysis about cloud-top properties? 2) What do PD_ave and PD_pert explain with respect to growing convective clouds? 3) How can these results be applied when monitoring active convective clouds in HRV, VIS, and NIR data prior to CI?

In general, the differences between the 0.6-, 1.6-, and 3.9-μm channel data observations are related to (a) scattering properties of ice versus liquid water hydrometeors, (b) changes in optical depth, and (c) the increase in hydrometeor size as clouds grow and persist. Reflectance changes that correlate to the transition to mainly ice hydrometeors, and to particle size (effective radius) increases that accompany cloud-top glaciation, are observed in the 1.6- and 3.9-μm data. Optical depth changes can be estimated from changes in the 0.6-μm channel (Baum et al. 2000a,b; Greenwald and Christopher 2000), yet for the “optically thick” convective clouds being analyzed here, less sensitivity is expected (as seen in Fig. 1b, cf. with Figs. 1c and 1d). Greenwald and Christopher (2000) show that changes in Refl_1.6 are associated with changes from water to ice hydrometeors. Nakajima and King (1990) show that both the Refl_1.6 and Refl_3.9 (and associated time changes) are correlated with the effective radii of particles for both liquid and water hydrometeors. Therefore, as a cumulus cloud’s main updraft region partially or completely glaciates, Refl_1.6 and Refl_3.9 will drop as ice particles form and their sizes increase significantly. The reflectance (-drop) signature from the 1.6-μm channel is found to be more immediate and larger initially when compared with the Refl_3.9 as glaciation first occurs, yet as clouds deepen and larger particles form (into the ≥10-μm range), Refl_3.9 is seen to fall more quickly (Nakajima and King 1990).

For Refl_3.9 signatures, cloud-to-cloud variability has been correlated with updraft velocity and storm intensity, and hence the time particles have to grow prior to freezing (Levizzani and Setvák 1996; Lensky and Rosenfeld 2008; Rosenfeld et al. 2008). Additionally, Refl_3.9 will vary with cloud condensation nuclei (or aerosols) concentrations and cloud-base temperature (via changes in particle size). As documented then, 3.9-μm (and likely 1.6 μm) information will possess more variability between cumulus clouds as local storm environments influence instability, updraft strength, and cloud microphysics. Updraft strength can subsequently be inferred in time trends in Refl_1.6 and Refl_3.9, in part due to decreases in particle sizes at cloud top, as shown in Setvák et al. (2003).

As a means of directly comparing the Refl_0.6, Refl_1.6, and Refl_3.9, Figs. 3a–d show scatterplots of these data at the third time in the sequence of images, only for the coldest pixel in the 3 × 3 pixel area per convective cloud. In Figs. 3a–c, colors correspond to the 10.8-μm T_B, which allows us to understand these relationships as a function of cloud depth, cloud-top temperature, and the likely presence of ice. For Fig. 3a, the relationship between 0.6 and 1.6 μm is shown, which is contrasted with Fig. 3b. In Fig. 3a, in general, as cloud tops cool, 1.6-μm reflectance decreases, while the Refl_0.6 remains highly variable. The scatter is somewhat larger for warmer clouds as effective radii of particles remain small. The wide variance in Refl_0.6 suggests that, with large optical depths for all clouds studied (≫1), the roughnesses of cloud tops remain similar as they grow. (The 0.6-μm reflectance is known to be more or less independent of particle size.) Also, as the amount of cumulus cloud occupying a 3-km pixel, or its horizontal extent, increases, so will Refl_0.6 simply because the amount of cloud in the scene increases.

In Figs. 3b and 3d, low (<3%) Refl_3.9 are also associated with colder, glaciated cloud tops and the presence of larger ice particles [to as large as 30–70 μm; Levizzani and Setvák (1996)], while again the 0.6-μm values do not vary as a function of cloud-top T_B. Because of the positive correlation between glaciation and optical depth (Baum et al. 2000a), the 1.6- and 0.6-μm channels are more linearly related, with increasing optical depth (Refl_0.6) correlating with increased ice hydrometeors (Refl_1.6), with lesser correlations with warmer clouds.

Figure 3c illustrates the relationship between Refl_1.6 and Refl_3.9, which shows that as ice hydrometeors form, Refl_3.9 drops significantly, being mostly below 3% when cloud-top T_Bs are below 235 K. The Refl_1.6 decreases to mainly <25% for clouds below 240 K, however, not exclusively. The lack of a 1:1 relationship in Fig. 3c is caused by differences in the indices of refraction at both wavelengths, which vary between ice and water to also make the single-scattering albedos different. There is some sensitivity of Refl_1.6 and Refl_3.9 to particle sizes, yet we expect this to be small. From Figs. 3a–d we conclude that use of both the 1.6- and 3.9-μm channels (the top four fields in Table 4) provides us with unique tests for glaciation and associated particle size changes, which subsequently can be related to the CI process when ice is important in precipitation formation. It is also clear that reflectance from the 0.6-μm channel provides less definitive information about cloud-top character, hence it being least important (Table 4), yet it can be used to confirm that a cloud is optically thick and hence that use of Refl_3.9 is acceptable.

From Table 5, the three top PD fields show that the presence of active up- and downdrafts within isolated or clustered cumulus clouds are important indicators of rapidly growing cumulus. High PD 15 and 30 min prior to the current time, and at the current time, imply active convective updrafts, and exceptionally lumpy clouds (as opposed to more smooth-topped clouds that accompany stratiform clouds). The perturbation values at 15 min prior to, and at, the current image time compose the remaining fields. These latter two fields help describe the presence of either locally tall and isolated turrets, or simply the existence of one main towering cumulus cloud updraft.

Table 7 lists the mean and standard deviation (SD) values of the retained interest fields. Note that the SD values are often much larger than the mean values, sometimes by a factor of 2–3. This implies a wide range of variability in all reflectance and PD fields across this population of growing cumulus clouds, quite likely related to varying convective environments on the COPS days studied.

Next, we examine the subset of clouds that were relatively warm at the third image time, as was done in Part I. Growing cumulus cloud events in which the cloud tops remained ≥240 K were analyzed separately, with the results presented in Table 8. Several interesting conclusions can be drawn when comparing Tables 7 and 8. First, whereas the Refl_0.6 field is of minimal importance in describing clouds over the entire dataset, for warmer-topped clouds it becomes an important and unique field. For cumuli either growing and nonglaciated, or just beginning to glaciate, the two time trends in Refl_0.6 are positive, in contrast to negative reflectance trends seen at 1.6 and 3.9 μm. The high brightness (35%) seen in Table 7 for the instantaneous Refl_0.6 seems to be consistent with the upward trend in this field seen in the top two fields in Table 8. This implies that increasingly lumpy, visibly bright clouds are associated with active convection, and (as noted above) that Refl_0.6 will increase as clouds occupy more of a single 3-km pixel as they increase in horizontal extent. The 1.6-μm trends in reflectance remain negative (as in Table 7), yet again they fall behind the 0.6-μm fields in terms of importance. Also in Table 8, the top three PD fields remain nearly the same as that seen for the entire dataset, with PD_pert 30 min previous to the latest image in the sequence replacing the present-time PD_pert, as seen in Table 7.

d. Application of results

Methods for using the information provided in Tables 7 and 8 are proposed in an analogous way to that presented in Part I. When using MSG’s VIS–NIR data alone in nowcasting growing cumulus clouds toward CI, weighting each field along the lines of the rankings in Tables 4, 5, 7, and 8 while assuming normalization between 0 and 1 is one possible approach of threshold-based tests. Another method would be to score a cumulus cloud object from 0 to 11 for cloud-top glaciation, and then combine this with information determined from the IR interest fields (Part I). For estimating CI likelihood for individual convective clouds (per pixel, or per cluster of pixels), use of the mean values to define a range per interest field that accounts for nonlinearities between HRV and VIS–NIR observations and physical processes (i.e., glaciation, hydrometeor size variability) would be a practical application. This was the method used when implementing the algorithm as described in Mecikalski and Bedka (2006) and Siewert et al. (2010).

To help estimate a range of values per interest field, one may add or subtract (whichever is appropriate) ½ to 1 SD from the mean value, per field. For example, a critical value for the 30-min Refl_3.9 trend might be −0.026 ± 0.041, or −0.014 to −0.067 (or a 30-min reflectance decrease of ∼1.4–6.7%), or defining the “critical” value for the 30-min Refl_1.6 trend to be between −0.021 and −0.117 (from −2.1% to −11.7% per 30 min; −0.048 ± 0.069). Use of null events (in which clouds failed to CI or continue developing) should be done to help set optimal threshold values.

Convective environments [or convective regimes; Boccippio et al. (2005); Li et al. (2006)] vary considerably in space and time; therefore, being aware that growing cumulus clouds will exhibit different cloud-top signatures in reflectance and BV is important. This is caused by 1) variation in aerosols, which influence the scattering and attenuation of visible light returning to the Meteosat-9 sensor; 2) the local freezing altitude, which impacts cloud-top glaciation rates and hence the reflective properties of newly glaciated clouds; and 3) instability, the amount of convective available potential energy (CAPE), and its vertical profile, which influences updraft velocity directly and, indirectly, the spectral signature of glaciation via cloud-top microphysics (Setvák and Doswell 1991; Rosenfeld et al. 2008; Lensky and Rosenfeld 2008). High instability (i.e., high CAPE) and vertical variations in stability should enhance BV as distinct, vigorous cloud turrets will be present within a scene.

More sophisticated procedures for using HRV and VIS–NIR data would involve the creation of histograms relating interest field values and environmental factors (e.g., vertical temperature and/or moisture profiles), developing empirical relationships, and subsequently determining the likelihood of CI. This method is similar to an ensemble approach and would require the analysis of a substantial number of events such that a robust statistical model or empirical function could be formulated. Depending on whether or not the “rate of CI” is important, critical thresholds may be increased or decreased accordingly based on relationships to known aspects of the convective environment in which the cumulus clouds are embedded, perhaps allowing for different weights to be placed on fields as a function of the atmospheric characteristics. Specifying a variable range in interest field-specific “critical” thresholds as a function of environmental conditions is another way of accounting for these factors. Environmental information can come from regional numerical weather prediction models (e.g., the European Centre for Medium-Range Weather Forecasts, ECMWF), via passive microwave remote sensing retrievals (which possess a much larger field of view), or also from IR with ∼(3–6)-h sampling times (e.g., the Atmospheric Infrared Sounder, AIRS).

PD fields can identify rapidly growing updrafts (a lumpy appearance in the imagery), allowing a user to discern between convective and nonconvective clouds in an unstable environment. In cases of nonconvective, stratocumulus, or cumuliform clouds, when weaker updrafts are providing the main cloud-growth signals, low PD values will be obtained. Low PD situations can then be evaluated more carefully, leading to a reduction in the false detection of new CI when other VIS, NIR, and IR information is used.

e. Error sources

Because of the relatively large sample size (123 events), most of the main statistical signals should be well represented. Yet, errors do exist. The main sources of error in this study are summarized as follows. First, as in Part I, since the events used in this study were determined subjectively, it is likely that several of the 123 CI cases are not ideally oriented in time and space relative to the Meteosat-9 data used. Specifically, it is possible that at the initial time (of a sequence of three, 15-min images) the cumulus clouds may have already reached the “towering” cumulus stage, and by the third image in the sequence, rainfall (a ≥35-dBZ echo) may have already occurred. In these cases, the physical processes of growing cumuli related to the Meteosat-9 data would still be captured, yet several of the trend fields may be erroneous as the cloud would have possessed an anvil. Second, the subjective determination of cumulus cloud types in a three-image sequence is sometimes incorrect, and nonconvective clouds would be analyzed instead to define the interest fields. This is caused by challenges involved in identifying cumulus in 1-km HRV data and 3-km pixels.

Other errors are associated with view angle influences over the COPS domain. With the southern edge of this domain at 46.5°, high view angles exacerbate the interpretation of cloud types from MSG data, which can lead to some false determination of the size of the pre-CI convective clouds as a significant portion of the sides of cumulus clouds are viewed (dulling the “cloud top” signal). A third source of error is the likelihood that higher-level cirrus clouds obscured lower cumuli in some of the images used per case, which will weaken the observed VIS–NIR cloud-top signatures with respect to solar reflectance. Also, error caused by the inadequate ground-based radar observations of a given new thunderstorm development over the COPS domain will influence the precise time determination of CI (and hence the proper image set used per event to define the interest fields). Finally, it is very possible that some CI events never achieved >35-dBZ echo intensity, because of dry environments or poor radar observations. Errors caused by misregistration between the HRV and VIS–NIR fields are not expected to be a source of error (see Schmetz et al. 2002).