A convective cloud (CC) analysis is performed over the southeastern United States (SEUS) during June, July, and August 2006 and 2007, using data from the Geostationary Operational Environmental Satellite (GOES) visible and infrared sensors as processed by a satellite-based convection cloud mask and initiation algorithm. Six 5–7-day periods are analyzed between the times 1500 and 1900 UTC, representative of summertime conditions in the SEUS. The ~8.7 × 108 pixel database contains information on nonprecipitating CCs possessing various satellite-estimated attributes of cloud size, based on whether they meet set thresholds in eight infrared “interest fields.” CCs at ~1 km × 1 km pixel size in the GOES projection are evaluated in comparison with the land cover classes, elevation gradients, and normalized difference vegetation indices (NDVIs) beneath the CCs. The goals are to relate the frequency of occurrence of CCs to land surface properties, attempting to determine which of these three properties are most correlated with CCs. CCs are more likely to form over forests and dense vegetation and over higher gradients in elevation. Although forest cover classes are not the most common over the SEUS, CC occurrence increases disproportionately where steeply sloped topography and forests are coincident across large regions of the SEUS. Also, as NDVI increases, the percentage of CCs per land class also increases. Analysis of landscape heterogeneity (combining local variability in land classes, topography, and NDVI) shows that as it increases CC development is more widespread. Thus, lakes among forests and hilly topography intermingled with agricultural lands appear most conducive to high CC frequency.
To mature into cumulonimbus clouds, convective clouds (CCs) require three main ingredients: low-level moisture, instability, and a lifting mechanism, as needed to overcome the capping inversion that prevents deeper convection mixing (Johns and Doswell 1992; Doswell et al. 1996; Banacos and Schultz 2005). Rapidly developing CCs are well correlated with the future development of mature cumulonimbus clouds (Roberts and Rutledge 2003; Weckwerth and Parsons 2006). Land surface heterogeneity is known to cause convergent boundaries, developing as a result of differential heating between various land cover types, such as water versus land, forests versus croplands, pasture versus forests, and urban areas versus croplands (Hong et al. 1995; Pielke 2001). Boundaries can also develop along topographical features because of differential heating of sloped land given changing sun angle throughout a day (Wilson and Schreiber 1986; Zängl and Chico 2006) and changing vegetation with elevation. Upslope flow due to mountains or hills can also act as a lifting mechanism (Kuo and Orville 1973; Wasula et al. 2002).
The method developed by Mecikalski and Bedka (2006, hereinafter MB06) uses visible (VIS) and infrared (IR) satellite imagery to identify and examine the evolution of immature or nonprecipitating CCs, using a simple threshold technique in order to forecast deep convection [convective initiation or (CI)] 30–60 min in advance. The method isolates CCs in satellite imagery using a “convective cloud mask” (CCM; Berendes et al. 2008). In MB06, CCs are tracked within a sequence of three satellite images over 30-min intervals from a geostationary sensor [e.g., the Geostationary Operational Environmental Satellite (GOES)]. Each CC, or cluster of cumulus cloud pixels, is ranked or scored on a scale of likelihood and potential for future CI and development into a cumulonimbus cloud using pixel-based thresholding techniques operating on instantaneous and time-trended IR “interest fields” (IFs). Eight IFs are applied to the GOES pixel data in MB06. Effectively, the Berendes et al. (2008) CC identification algorithm supplied the main data used in the present study, and each field characterizes or explains a physical process dictating CC maturation (Mecikalski et al. 2008) to the extent the GOES sensor can observe these aspects of CCs.
A guiding hypothesis was that initial CC development is not random, but instead is linked to land surface features. We propose that the analysis of a large (8.7 × 108 pixels), 1-km pixel-size satellite database of CCs from a 2-yr archive of half-hourly data, over six multiday time periods during the summers of 2006 and 2007 over the southeastern United States (SEUS; Fig. 1), together with land surface data, can help identify these linkages. The six periods studied are representative of typical summertime weather conditions across the SEUS. The present study did not address CI specifically, but instead used the MB06 IR indicators interpolated to 1 km to quantify the presence of pre-CI CCs at the resolution of the GOES VIS data ~(1 km × 1 km), and evaluated their frequency of occurrence with respect to land surface heterogeneity. This evaluation related CC size by the number of IF thresholds held true per CC pixel. Statistical analysis of these IR fields highlighted areas more or less conducive to CC growth across the SEUS, related to land surface features, and henceforth allowed us to quantify which land surface factors are important for CC patterns, spatial coverage, and frequency of occurrence.
Another hypothesis for this study was that CC development and evolution are more likely over regions of increasing land surface heterogeneity because these are the regions most likely to induce mesoscale circulations that support updrafts for CC development. Therefore, this study intends to assess the individual role that specific land surface features play in CC development, and then relate CCs to land surface heterogeneity. Land surface heterogeneity characteristics (e.g., forests and lakes) for this study were examined in terms of land cover classification, elevation gradients, and normalized difference vegetation index (NDVI). Land with no heterogeneity was defined as having no slope and no variations in vegetation or land cover character. The goal will be to determine if correlations exist between lands with various types of heterogeneity (e.g., steep slopes) and CC frequency of occurrence; high correlations will confirm our hypotheses, while weak to no correlations support the null hypothesis.
Section 2 provides a background by examining previous research, and the MB06 algorithm that supplies the main dataset for this effort. Section 3 discusses the process of acquiring and analyzing these data. Section 4 presents the CC analysis and the assessment of land surface heterogeneity’s role in CC occurrence, with section 5 discussing the main conclusions of the research.
a. Previous research
Relevant to this study, we reviewed previous research related to CCs, their frequency of occurrence, and land surface characteristics. Klitch et al. (1985) developed a CC climatology over the Rocky Mountains and high Great Plains based upon an analysis of GOES VIS and IR satellite imagery. The CCs were limited to those with cloud-top temperatures ≤−30°C in IR imagery and clouds that were sufficiently bright in VIS imagery. Their criteria were mostly for mature CCs (near the cumulonimbus stage), and their main finding was a positive relationship between increasing elevation gradients and CC formation related to sun-relative slope angle.
Rabin and Martin (1996) also used satellite imagery to monitor CC coverage, but did so over the central United States. Hourly GOES VIS and IR satellite images isolated shallow CCs by way of albedo (20%–100%) and temperature (285–310 K) thresholds. The purpose of their study was to determine the frequency of CCs in relation to land surfaces to aid in climatological analysis. Data from July 1987 and 1988 were analyzed together with land use, elevation, and NDVI datasets in Rabin and Martin (1996). It was found that the frequency of CCs is higher with higher elevation and lower NDVI. CC development was enhanced through differential heating along sloped topography. Lower NDVI values were associated with higher values of surface temperatures and sensible heat flux and decreased values of latent heat flux, which led to an increase in the frequency of nonprecipitating CCs during relatively dry atmospheric conditions.
Kuo and Orville (1973) and Wasula et al. (2002) provide useful comparison studies because they consider similar relationships between land surface features and CC development, yet differ in that they analyzed only CI events, using severe storm reports, radar, lightning, or various other methods and combinations. Kuo and Orville (1973) developed a radar-based climatology for the Black Hills area in South Dakota, and data were collected when cloud seeding was occurring within the area. Results showed the Black Hills enhances CCs and CI events by differential heating near the sloped topography, and topographic channeling of airflow and enhanced convergence and/or lift. Similarly, Wasula et al. (2002) formed a climatology based on severe weather reports of wind, hail, and tornadoes in eastern New York and western New England. They verified the occurrence of thunderstorm development with an analysis of cloud-to-ground lightning. Their climatology showed that terrain does indeed play a role in influencing CI by funneling airflow, again causing convergence, similar to Kuo and Orville (1973). Although both Kuo and Orville (1973) and Wasula et al. (2002) focus more on post-CI associations, their studies are relevant, because they show that land surface features influence CCs and convective activity.
Aside from CC and CI climatologies, several studies addressed land surface heterogeneities effects on CC growth. Hong et al. (1995) and Pielke (2001) both showed that land-surface properties influence CC growth. Hong et al. (1995) used a numerical model with variable land surface and atmospheric conditions to demonstrate the effect of land surface on CC growth. They showed that vegetation and resultant latent heat flux can trigger CCs for relatively moist atmospheric conditions, but the mechanisms for CC development are sensitive to the prevailing synoptic-scale moisture and stability fields. Pielke (2001) reviewed other published work to show an association between land surface and CC growth. Higher leaf area index was shown to increase the latent heat flux, which in turn increased the convective available potential energy, which enhanced the potential for CCs and deep convection (Pielke et al. 1997). However, other studies (Rabin et al. 1990; Wetzel et al. 1996) showed a positive association between lesser vegetation, higher sensible heat, and enhanced CCs, similar to Rabin and Martin (1996). Overall, areas that enhance low-level convergence and/or lift are favorable to enhancing nonprecipitating CCs.
All of the above studies indicate that CC growth is related to land surface properties, yet uncertainty prevails as to which property (land cover type, vegetation, topography) dominates CC development. The uncertainty and differences in the previous studies are likely due to different atmospheric parameters and synoptic conditions that ultimately determine convective activity. None of the previous studies are satellite-based analyses performed over a large geographical region, in this case, over the SEUS.
b. CC from satellite
The present study exploited archived GOES fields as used in MB06. GOES IR and VIS satellite imagery is useful as it monitors CCs before they mature into a thunderstorm (i.e., they can be tracked as they grow; Purdom 1976). The MB06 algorithm is based on the premise that an immature or nonprecipitating CC, showing signs of increasing development, will continue to evolve into a deep CC. Because the current study does not address CI, but instead the potential for CCs development and growth, possibly leading to mature CCs and CI, no attempt is made to verify whether the CCs in the database did, in fact produce rainfall with reflectivities ≥35 dBZ (hence, went through the CI process). The rainfall echo intensity of ≥35 dBZ is typically used as the definition of CI (Weckwerth and Parsons 2006).
A CCM is used to isolate CCs in satellite imagery. Five types of clouds are classified in the CCM (Berendes et al. 2008): low and midlevel cumulus (“fair weather” or cumulus humilis), moderate towering cumulus (cumulus mediocris), large towering and deeper cumulus (cumulus congestus), and thick cirrus ice clouds–cumulonimbus tops. IR IFs are formed for all cloud types except thick cirrus ice clouds–cumulonimbus. Mesoscale atmospheric motion vectors (MAMVs; Bedka and Mecikalski 2005) are used to track the CCs in subsequent 15-min images. Some spatial offset from advection does occur between the source region for warm thermals and where CCs may first form. Given that six 5–7-day periods were examined, with 106 pixels per period, and that “tranquil” low-cloud-top-level wind conditions were examined [i.e., CC motions <(8–9) m s−1], we propose that this analysis will capture the main relationships between land cover variability and CC frequency. CCs possessing higher numbers of IFs (so-called per-pixel scores) are assumed to be large and active.
MB06 relates the 8 IR IFs (composed of combinations of 6.5-, 10.7-, and 13.3-μm brightness temperatures, TB) to physical aspects of growing CCs. The fields are IR cloud-top TB (indicating cloud-top glaciation), IR multispectral channel differences (for estimating cloud depth relative to the tropopause or equilibrium level), and IR cloud-top TB–multispectral temporal trend assessments (for estimating updraft intensity). These IFs, along with their resolutions and critical thresholds, are listed in Table 1. Each IF equals “1 point” (up to 8) in the per-GOES pixel thresholding (or scoring) approach to nowcasting CI as proposed by MB06. The sum of the IFs that reach a critical, threshold value equals the score for a given CC. Since we were not evaluating CI as in MB06, we chose to rename the “CI score” per pixel to “CC score,” for clarity. CCs that meet ≥6 IF thresholds are rapidly growing, immature, and nonprecipitating, but likely to mature into a cumulonimbus cloud within the next 30–60 min (MB06). CCs with scores from 3 to 5 are generally tall or “towering” cumulus clouds, whereas CCs with scores ≤ 2 are “fair weather” and short lived (~15 min).
The analysis herein quantifies CCs as a function of score toward estimating clouds in varying stages of development. The presumption is that CCs with higher scores possess a stronger, more persistent updraft based on interpretation of the IFs relative to those with lower scores, and are realizing greater instability and/or are connected to a stronger low-level convergence circulation. We then link land surface heterogeneity to CC updrafts, helping to address the hypotheses above.
In the CC dataset with respect to physical processes (Table 1), IFs 1 and 7 indicate tall CCs that have likely broken through a capping inversion. Fields 2 and 8 indicate the presence of a wide updraft (because the 13.3-μm channel is 8-km pixel size on GOES-12). Fields 3 and especially 4 suggest that cloud-top glaciation is likely to occur for the CCs, whereas IFs 5 and 6 indicate a growing CC updraft and can be used to estimate updraft strength (Adler and Fenn 1979; Roberts and Rutledge 2003; MB06).
3. Data acquisition and analysis
a. CC data
For this study, the per-GOES pixel scored CCs over the SEUS were averaged over six 5–7-day periods in summers 2006 and 2007, between 1500 and 1900 UTC for each day every 30 min. The 5–7-day periods were chosen because monthly averaging often blurs CC–land surface relationships, and 1–3-day time scales often possess too much synoptic interference and relatively small CC numbers for analysis. We avoided times when organized convergent boundaries on meso-β scales or larger (≥25 km; e.g., fronts and outflow boundaries) were present, that cause CC formation, implying that local, land surface–related factors are not primarily responsible for CC development (Carleton et al. 2008). Each 5–7-day period required the analysis of 2528 × 1061 sized images at 1-km pixel size, for each 30 min spanning 1500–1900 UTC. Therefore, ~24.1 × 106 pixels were evaluated for each day of the 5–7 days. Using an average of six days per period, for six periods, this equated to ~8.7 × 108 pixels being analyzed, with just over 49% of these pixels containing CCs with scores ≥ 1.
Time period selection was based on consecutive days within June, July, and August 2006–07 that had the least synoptic-scale influences, with relatively calm or tranquil conditions. The “synoptically calm” criteria of Brown and Arnold (1998) were used when selecting time periods, which are (i) no synoptic near-surface “forcing mechanisms” identifiable within 500 km (i.e., no fronts, little to no quasi-geostrophic influence), (ii) 500-hPa wind speeds within the region < 7.5 m s−1, and (iii) surface wind speeds < 5 m s−1. A review of surface and 500-hPa analyses was performed to determine whether conditions (i)–(iii) indeed existed. Widespread high boundary layer humidity prevailed on all days processed (surface dewpoint temperatures Td generally ≥17°C), with variations in Td on the mesoscale of ≤5°C.
As noted above, because of the focus on weak wind conditions, the influence of advection on the results presented is expected to be small. Specifically, for horizontal speeds of 0–8 m s−1 at cloud level, the horizontal advection between the observed CCs and the surface features that contribute to their growth will be on the order of 0–2 GOES 4-km pixels. Given the large number of CCs analyzed, in light of small advection distances, we did not anticipate this factor to significantly influence or bias the results.
Six time periods met our requirements: 5–10 June 2006, 8–14 July 2006, 29 July–3 August 2006, 24–28 June 2007, 15–19 July 2007, and 3–9 August 2007. The 1500–1900 UTC time of day (1000–1400 LST, depending on time zone) was chosen to capture the first development of CCs on a given day, caused by differential surface heating, mesoscale thermal circulations formed along sloping terrain and contrasting surface covers, as well as upslope flow prior to the time of maximum solar radiation. Using times after 1900 UTC resulted in the inclusion of CCs forced by convergence from convective outflows, and obscuration due to anvil cirrus (which prevents the MB06 algorithm from working). Times earlier than 1500 UTC were not considered because of the relative lack of CCs. For each time period, average CC scores per pixel were computed as the sum of the scores divided by the sum of the pixels.
The distribution of scores for CCs, from 0 to 8, was found to resemble exponential (or near exponential), with decreasing CC numbers as scores increase. The Kolmogorov–Smirnov test was employed as a way of verifying exponential distributions (Lilliefors 1969). Ocean areas were processed and used to define the “null” hypothesis. The null hypothesis was that the land surface features have no effect on where CCs develop or organize. Over-ocean pixels therefore help estimate the relative importance of various land surface features with respect to CC occurrence. Specifically, if the CC distributions over land and ocean were similar, then one can conclude that NDVI, topography gradients, and land surface cover types do not influence how CCs are distributed. If there are significant differences between land and ocean, then over-land CC distributions are nonrandom. This reasoning assumes that CC development over oceans is more random, while that over land areas is nonrandom because of land surface features, during synoptically calm conditions. Exceptions occur when sea surface temperature (SST) gradients influence CC development, yet by June and especially August, SST gradients will be small over the Gulf of Mexico.
b. MODIS land cover
Land cover classification fields from the National Aeronautics and Space Administration’s (NASA) Earth Observing System (EOS) Moderate Resolution Imaging Spectroradiometer (MODIS) were used. The MODIS/Terra land cover is 1-km pixel size in sinusoidal projection, using 12 months of data for the yearly product (Strahler et al. 1999); the 2004 MOD12Q1 land cover data were used. The data were obtained from the EOS Data Gateway of the U.S. Geological Survey (USGS) Earth Resources Observation and Science Center.
The land surface scheme of the International Geosphere–Biosphere Programme (IGBP) was chosen for its in-depth classification of 17 classes of land cover (Table 2). The data were mosaicked together and reprojected geographically using the MODIS Reprojection Tool (MRT) to match the spatial domain of the GOES CC data.
Topography data were collected from the USGS Center for Earth Resources Observation and Science. We used data from the Global 30 arc second elevation dataset (GTOPO30), which is a global digital elevation model with a horizontal grid spacing of 30 arc seconds, or ~1 km (U.S. Geological Survey 2006). Elevation is in meters above mean sea level. Ocean areas are identified within the dataset, allowing for a distinction between land and water. The topography data were used to calculate the elevation gradients for the SEUS to assess sloping terrain in regard to CCs.
d. MODIS NDVI
NDVI data were collected from NASA’s EOS MODIS, and we used the MOD13A2 product (1-km pixel size), which provides 16-day global coverage in sinusoidal projection (Huete et al. 1999). Two 16-day periods were chosen in 2006 and 2007 to coincide with the GOES datasets: 28 July–12 August for both 2006 and 2007. Like the MODIS land cover data, the NDVI data were mosaicked and reprojected using the MRT for each 16-day period. NDVI data are used to quantify the density of green vegetation on a scale of 0–1. Higher NDVI values indicate a large density of green leaves such as in the case of forests; however, lower NDVI values indicate sparse vegetation, such as in urban landscapes. An NDVI value of 0 indicates no vegetation.
e. Sources of error
Several sources of error can influence the research results. One source involves the improper identification of CCs, which relies on the CCM, and certainly the influence of including non-CCs in the analysis will lead to systematic errors. For example, if a thin nonconvective cloud (e.g., thin cirrus) overlies a warmer surface (ground or low stratus), the CCM may mislabel it as a CC, and it will be included in this analysis. This source of error will occur more often for low-score pixels (<2) as higher-score pixels, exhibiting multiple IR thresholds (Table 1), represent active CCs. Higher-scored pixels possess IR signals that are not possible from nonconvective cloud types. As the Berendes et al. (2008) CCM method classifies CCs based on statistical thresholds, outside of thin cirrus, clouds would be classified as “cirrus/ice” or “unknown.” Hence, we are confident that the CC database is not significantly impacted by misclassified clouds.
Mecikalski et al. (2008) quantify the errors that can result from improper tracking of CCs in the MB06 algorithm, and the subsequent poor threshold estimation. These tracking errors are unbiased. However, use of data containing the most serious tracking-related errors were avoided because low-wind weather periods were chosen for this study in which cumulus clouds moved slowly [generally <(8–9) m s−1] and/or vertical wind direction shears were small (<20° km−1). Therefore, tracking errors were minimal.
Synoptic conditions would ideally be calm or tranquil, like those defined by Brown and Arnold (1998), throughout the entire region of analysis. However, because of the large coverage area and multiconsecutive day periods selected, there are likely times when over small portions of the domain (~5%) this was not the case, where meso-β-scale convergent boundaries caused CC development in some isolated locations, regardless of land cover and topography influences. Variation in the synoptic-scale environment between time periods can also be attributed to some of the differences (e.g., variations in mean surface Td’s, depth of moisture across the SEUS). These sources of error are likely to mute the CC–land surface relationships as CCs would occur regardless of land surface influences in the presence of convergent boundaries.
f. Data analysis
Collocation of all three datasets was performed over the area approximately 25°–37°N, 79°–97°W, with the MODIS and topography mapped to the GOES projection and CC dataset. Many results are presented in histogram form to show the distribution of various datasets in relation to each other with the following scale:
with x being the number of occurrences.
Land cover classes that occupied <1% of the entire domain were omitted from examination over the SEUS. These classes were 3 [deciduous needleleaf forest (0.03%)], 6 [closed shrublands (0.12%)], 7 [open shrublands (0.13%)], 11 [permanent wetlands (0.63%)], and 16 [barren or sparsely vegetated (0.09%; Table 2)]. Other classes used in this analysis occupied much higher percentages, making the <1% delineation appropriate.
The total pixels in each class were determined from the output of the histograms, and the total number of pixels in each land class was divided by the total number of pixels in the domain (multiplied by 100), to determine the percentage of pixels for each land cover class. This percentage was computed to separate land cover class frequency from CC frequency.
To determine each land classes’ CC potential, the CC percentage (CC%) was calculated. Averaged scores were totaled for each class, and CC% was computed using
where pixels with averaged per-pixel CC scores > 0, as calculated by Eq. (1), are divided by the total number of pixels in the land cover class (multiplied by 100). For example for all time periods added together, class 1 (evergreen needleleaf forest) had 315 510 pixels. Of those pixels, 217 692 pixels had averaged CC scores > 0. The CC% for class 1 (evergreen needleleaf forest) would be 69%. This process was performed individually for all 5–7-day periods. Then, Eq. (2) was used again to determine the average CC% for 2006 and 2007 as a whole for the SEUS.
Elevation gradients, rather than elevation alone, were used in this study, as guided by previous CC-related research. Elevation gradients were calculated using the quadratic method of Zevenbergen and Thorne (1987) to determine the change in elevation with respect to each 1-km pixel. The method calculates the rate of change in elevation in both the north–south and the east–west directions. Correlating CC formation and frequency to terrain slope was the main goal. The variability in terrain across a large section of the SEUS is small (<±200 m over several kilometers), with the obvious exception being in portions of the Appalachian and Ozark Mountains.
Each land surface dataset was analyzed individually in comparison with the CC data. After the individual analysis, a measure of land surface variability (LSV) was created, combining all three land datasets. This was done to assess whether the combination of land cover, topography gradients, and NDVI across a region is more important in generating CCs versus any individual component. For LSV, land cover classes were assigned appropriate average vegetation heights (Table 2) so they could be assessed quantitatively in regard to the other datasets. LSV was computed by taking the standard deviation (SD) of each land surface dataset (pixel by pixel), and dividing it by the maximum SD of each corresponding dataset. Hence, LSV is given as
LSV has a range of 0–3. A 0 indicates no gradients in any fields, while 3 is the maximum possible found across the domain. For the time periods examined, the only variability in LSV per location was caused by the NDVI dataset, which varied by <5% per year. We used the same land cover and elevation gradient datasets for both years.
a. General patterns
General patterns of CCs over the SEUS were examined, and then compared to land cover, elevation gradients, and NDVIs. Figures 1a–d show all fields for 8–14 July 2006 averaged between 1500 and 1900 UTC, inclusive. Figure 1a shows averaged per-pixel scores, Fig. 1b the land cover classes, Fig. 1c elevation gradients, and Fig. 1d the NDVI for the period 28 July–12 August 2006. Since the results for all time periods were generally similar, the plots in Fig. 1 and the following histograms (e.g., Fig. 2a) focus on 8–14 July 2006 as a means of limiting the presentation, with the accompanying graphs and tables presenting results for all six periods. One exception to the 8–14 July 2006 case was CCs over the ocean. The 8–14 July 2006 had the highest ocean CC% in relation to the other time periods. Therefore, this time period was not representative of the ocean domain for all time periods. The average CC% for the other time periods was ~24% in comparison to 41% for 8–14 July 2006.
In Fig. 1a, averaging scores reduces per-pixel numbers to generally between 0 and 4 since no single pixel possesses continuous CC cover over any 6-day period. CCs occur most frequently over land as opposed to water. This is seen as most of the SEUS is covered by blue shading, whereas the ocean is primarily white with CC scores = 0, that is, no CCs. In Fig. 1, occurrences of pixels with scores ≥ 5.0 (i.e., tall cumulus clouds) are indicated by squares to exemplify the distribution of large, growing cumuli since they are infrequent compared to lower-scored CCs. The color of the squares is not indicative of a specific CC value or score, but instead was chosen for clarity against the other plotted data.
From Fig. 1, several clusters of higher-scored CCs are noticeable, which points to mesoscale variability in the CC field and the more likely locations where updrafts were prominent and persistent. Consistent patterns of CCs over the Appalachian Mountains (and to a lesser extent along the ocean coasts) were noted in all periods (Fig. 1a), which suggests elevation alone does play a role in CC occurrence; however, CC frequency of occurrence is actually lower over the highest terrain of the Appalachians, which is consistent with previous studies highlighting elevation gradients as a probable more important factor regulating CC development versus elevation alone (Kuo and Orville 1973; Klitch et al. 1985; Rabin and Martin 1996).
b. Land cover relationships
Figure 1b shows CCs with scores ≥ 5.0 plotted onto land cover. Since it is difficult to distinguish many land cover classes in relation to CCs in Fig. 1b, CC correlation to land cover was examined using the frequency histograms for 8–14 July 2006 (Fig. 2a) to show the distribution of averaged CC scores for each land cover class. For all time periods and land classes, the highest frequency was for CCs with per-pixel scores < 4.0 (higher scores were least frequent). For each time period only a few classes had averaged CC scores ≥ 7.0 (large, persistent cumulus clouds nearing the cumulonimbus stage). These classes varied for each time period. For all time periods, the classes that always had CC scores ≥ 7.0 were 0 (water), 1 and 2 (evergreen broad and needleleaf forests, respectively), 4 (deciduous broadleaf forest), 5 (mixed forests), 8 (woody savannas), 9 (savannas), 10 (grasslands), 12 (croplands), and 14 (cropland–natural vegetation mosaic; see Table 2).
To distinguish the most prolific land classes from those that were consistently associated with CC occurrence, the CC% of each class was determined using the values from Fig. 2a and Eq. (2). Table 3 lists CC% rankings by land cover class for each period, with the average for both years together shown as well. Some consistencies do emerge from Table 3: For 8–14 July 2006, the top classes were 2 (evergreen broadleaf forest), 9 (savannas), 4 (deciduous broadleaf forest), 8 (woody savannas), and 5 (mixed forests). Three of these top five classes were forests. The other forest class, 1 (evergreen needleleaf forest), had a moderate ranking. The other two high-ranking classes, 8 and 9, were both savannas. The lowest ranked classes were 13 (urban), 12 (croplands), and 0 (water). After averaging all periods together, the top classes were 2 (evergreen broadleaf forest), 13 (urban), 1 (evergreen needleleaf forest), 5 (mixed forests), and 4 (deciduous broadleaf forest). This can be seen in Table 3 and Fig. 2b where land cover classes are plotted with respect to the average CC% for both years as blue diamonds. Four of the top five classes were forests. The classes with the least potential for CC development were 0 (water), 10 (grasslands), and 12 (croplands), which was consistent for most time periods.
Figure 3, a plot of land cover classes versus elevation gradients over the SEUS, shows that some classes are more prevalent over higher gradients. The most prevalent classes over land with higher topography gradients were forests [classes 4 (deciduous broadleaf forest) and 5 (mixed forests)]. Steep slopes tend to be forested in the SEUS, where urban development and especially agriculture are difficult. Yet, the question remains as to what is more important in CC distribution—land cover or elevation gradient—or whether they are interconnected because of land use–topography relationships. More will be said about this in section 4c.
Several reasons for the higher frequency of CCs over forests are possible, including forests’ ability to process moisture and release more latent heat, as related to their low albedo, compared to other land cover classes such as grasslands (Lawton et al. 2001). These influences result in slightly increased low-level moisture and hence greater low-level instability over forests, to aid CC formation. Forest–nonforest (e.g., pasture) boundaries also alter wind flow and help to form mesoscale circulations (Walker et al. 2009) because of aerodynamic roughness differences between forests and crops, causing low-level convergence.
It is important to note that the land cover classes in which CCs most frequently occurred were not all the five most common classes found over the SEUS, with the exceptions being classes 5 (mixed forests) and 14 (cropland–natural vegetation mosaic) (Fig. 2b). The most prolific or widespread classes in the database are 0 (water), 14 (cropland–natural vegetation mosaic), 5 (mixed forests), 8 (woody savannas), and 12 (croplands) as indicated by the red circles, which portray domain percentage in Fig. 2b. The remaining widespread classes ranked low in CC% for both years averaged together. Therefore, these CC% per land class results are not a restatement of the most occurring land classes.
A consistent pattern seen for each 5–7-day period is that class 0 (water) is the least likely to be associated with CCs. For all time periods, Fig. 2b shows water had a CC% of only ~27% as indicated by the blue diamonds. All other land cover classes had CC% values >50%. The most likely cause of this is that CCs are less frequent over oceans compared to over land. This led to the presumption that the water class may have a higher CC% ranking if ocean pixels were removed from processing, leaving only land and water pixels consisting of inland lakes and wider rivers (>1 km). To test this, a subset of the SEUS domain was taken from 31.0°–37.5°N and 97.0°–81.5°W for 8–14 July 2006. No ocean pixels were contained in this subset region. This resulted in land cover class rankings only slightly different from those found over the entire SEUS domain for this period (see Table 3). Classes 2 (evergreen broadleaf forest), 9 (savannas), 4 (deciduous broadleaf forest), and 8 (woody savannas) remained in the top five within the subset. Class 13 (urban) moved into the top position, although previously for 8–14 July 2006, it ranked low. The high ranking of class 13 (urban) for the subset and for all six periods averaged together suggests that urban areas increase CC frequency of occurrence despite their low ranking for 8–14 July 2006. Class 0 (water) only moved up one position in rank. Inland lakes and rivers aid in the growth of CCs downwind because of moisture fluxes, land-to-water heat flux difference, and by differential friction causing surface convergence.
With respect to the null hypothesis (i.e., land cover variability has no relationship to CC distributions), the distribution seen over water would be expected to prevail for other land classes. Using the Kolmogorov–Smirnov test, the class 0 (water) distribution of CC numbers per score is approximately exponential. For example, for periods 8–14 July 2006 and 15–19 July 2007, the mean number of pixels for all score categories (0–8) over ocean is 297 226, whereas the SDs are 578 006 and 477 222, respectively. Lilliefors (1969) states that in a perfect exponential distribution, the mean and SD would be the same. Since the class 0 (water) distributions are nearly exponential, the ocean is considered to be the null case. The nature of the distributions from all other land classes, however, as seen in Fig. 2a (for land classes 1–16), differs considerably from that of class 0 (water); hence land cover is influencing where CCs do and do not predominate.
c. Elevation gradient relationships
Figure 1a suggests that CCs have a greater tendency to develop within mountainous regions, especially the Appalachians and along the coasts, albeit the highest elevations have lower CC occurrences compared to those possessing higher topography gradients. In Fig. 4a, averaged CC scores are plotted with respect to elevation gradients in histogram format. Recurring patterns emerge: lower-scored CCs are spatially uniform over all elevation gradients. Overall, CCs were most associated with land with lower gradients because lower gradients are more prevalent overall. However, despite the very low total percentage (<10%) of the land possessing gradients >0.05 (red dots in Fig. 4b), the CC% increases from 70% to 100% as gradients increase to 0.45 (i.e., a disproportionately high number of CCs occur over relatively small regions of higher gradients). The conclusion then is that flow up gently sloping terrain and differential heating, a result of elevation gradients, enhances the development of CCs. No critical gradient threshold is linked to a jump in CC frequency; CC frequency increases as elevation gradients increase with a nearly exponential distribution still occurring per score.
A clearer pattern emerges in Fig. 4b, where elevation gradients are plotted versus their average CC% for all six periods [Eq. (2)]. Overall, as elevation gradients increase, the CC% increases. Alternately, an elevation gradient of 0, such as that over the ocean, produces a much lower CC% value in comparison to gradients >0. Although a 0 gradient is the most frequent as indicated by the red circles in Fig. 4b, they were the least conducive to CCs as indicated by the blue diamonds in the same figure.
d. NDVI relationships
In Fig. 1d, NDVI is plotted over the SEUS for the 8–14 July 2006 period, with locations with averaged CC scores ≥ 5 shown as squares. Averaged CC scores and NDVI were plotted as a histogram in Fig. 5a. NDVI < 0, which correspond to water and no data values, were reassigned values of 1.1. Figure 5a shows two domain frequency maxima at ~0.7 and 1.1 NDVI, as also indicated by the red circles in Fig. 5b. In Fig. 5a, the same maxima were seen for all averaged CC scores. As NDVI values increase, the frequency of higher scores (CC%) increases. This relationship is also presented in Fig. 5b, where NDVI is plotted versus the average CC% for all time periods [Eq. (2)] as blue diamonds. Figure 5b suggests that CCs are more frequent over heavier, denser vegetated areas. Much of the SEUS is covered by NDVI values >0.5 during the time periods, and so the relationship between NDVI and CCs is less clear as noted by the slow increase in CC% with NDVI. The continued increase in CC% after the domain percentage (red circles) decreases for NDVI values ≥0.7 leads to the conclusion that high NDVI enhances CCs. This high-NDVI–CC relationship suggests that active and dense vegetation produces increased evapotranspiration, and in turn leading to higher boundary layer water vapor (parcel equivalent potential temperature), updrafts with increased potential energy, lower cloud-base heights, and subsequently more CC growth; this is in line with previous research (Pielke 2001; Hong et al. 1995). While there is a maximum domain percentage of NDVI over water (NDVI = 1.1) in Fig. 5a, water was not very conducive to CCs as is evident by the CC% (Fig. 5b) and the above land cover analysis (i.e., the white space in Fig. 1a denoting a lack of CCs). The nature of the distributions for positive NDVI (Fig. 5a) differ considerably from that of 0 NDVI (top row in Fig. 5a), and as with land cover, this suggests that higher NDVIs are associated with greater CC occurrence.
e. Land surface variability
Figure 6a shows LSV [Eq. (3)], plotted over the SEUS for 8–14 July 2006, which is representative of the other time periods studied. Higher LSV values are noted along higher elevation gradients, especially along the Appalachian Mountains and coastal regions, which is where higher scores were located as indicated by squares (Fig. 6a). Lower LSV values occur around the Mississippi River region, for example.
In Eq. (4), a is averaged scores for CCs varying from 0 through 8, and b is LSV values 0 through 3.0. Overall, higher averaged scores were associated with higher percentages of pixels with higher LSV values. The highest LSV value obtained for each week was ~2.0, and the average LSV was ~0.3. This indicates that large SDs (i.e., high variability) in all three land datasets are not necessarily associated with CC development and proliferation. Either minor SDs in all three land datasets, or a large SD in a single land dataset, appear sufficient to aid in CC occurrence.
Figure 6b shows several important aspects of the relationship between LSV and CC scores. First, most CCs are associated with near-zero LSV (the result of the high domain percentage of ocean). Second, there is an increase in frequency of CCs as LSV increases between 0.5 and 0.7. Third, as LSV increases beyond 0.7 the number of CCs diminishes, which is related to the limited region defined by high LSV (>1.0). Fourth, the local maximum between 0.5 and 0.7 for all CCs, regardless of the score associated with them, implies that as land surface heterogeneity increases, CC frequency increases. Or, as LSV increases, the likelihood of CC occurrence is enhanced. Regions that possess increased variability in land surface characteristics (lakes among forests, intermingled with agricultural land, with significant topography) will support an enhancement in thermally buoyant updrafts, or substantial variability in the local lifted condensation level, such that CC development is more widespread (i.e., our second hypothesis). This is along the lines of previous analyses (Clark and Arritt 1995; Rabin and Martin 1996; Lynn et al. 1998; Lynn and Tao 2001). Our conjecture based on Fig. 6a is that landscapes with large variability in vegetation types and topography, and possessing scattered water bodies, are more conducive to CC occurrence compared to homogeneous landscapes, in line with our first hypothesis.
5. Discussion and conclusions
Here we review the main results and discuss the more speculative aspects from this study. CCs were analyzed over six 5–7-day time periods, beginning with an initial dataset of ~8.7 × 108 GOES 1-km-size pixels, for defined tranquil weather periods over a large region of the SEUS. CCs were identified by way of eight IFs of CCs as observed in GOES IR data, which pertain to several attributes of nonprecipitating CCs, with higher scores associated with larger CCs. This study makes use of the method of MB06 algorithm and the CCM of Berendes et al. (2008), which scores only CCs per pixel using thresholds for the eight IR fields. This study then analyzes CCs’ frequency of occurrence with respect to land surface components and land surface heterogeneity.
Generally, CC% was higher in association with certain land surfaces, specifically larger elevation gradients, forest-type land cover classes, and higher NDVIs (>0.5). CC% increased as elevation gradient increased, with no critical gradient magnitude being found. A strong relationship between CCs and land surface heterogeneity is quantified, namely higher CC% with increased heterogeneity, or variability.
Elevation gradients are the most important forcing mechanism for CCs, relative to land surface cover, cover variability, and cover type (and corresponding variations in NDVI). Elevation alone is of lesser importance in CC generation, as consistent with the previous studies cited here. From Figs. 1a,c, CC development occurs over elevation gradients, especially along the Appalachian Mountains, but not necessarily over higher elevations. This is justified as the standard equation for vertical velocity (w = dz/dt, where z is height) is relative to the magnitude of the topographic gradient, not just the magnitude of the elevation. Small CCs (with scores 1 or 2) were more likely to develop over all elevation gradients, whereas higher-scored, larger CCs were more associated with higher gradients. Overall, as elevation gradient increases, the CC frequency of occurrence also increases.
CCs were shown to develop over certain land cover classes more than others. Forest classes were the most conducive to CC development. The forest classes comprised four of the five top classes over which CCs were most prolific for the entire SEUS, yet these were not always the more spatially prominent land cover classes in the SEUS. Others [e.g., 12 (croplands)] were widespread, but CCs were not consistently highly associated with them. This finding is also consistent with the result that the growth of CCs is increased over regions with larger elevation gradients, because forests are more likely to be found along sloped topography in the SEUS, where urban development and agriculture are more difficult. As mentioned, the higher frequency of CCs over forests is likely enhanced by higher latent heat values increasing instability due to forests’ ability to store and release moisture, as well as because of differences in aerodynamic roughness values, that is, forests compared to crops. Differential heating and upslope flow are also likely contributing factors to enhancing CC occurrence.
An additional main result is that the distributions of CCs with scores 0–8 per land class vary considerably from exponential, whereas their distribution over ocean is closer to exponential. We therefore conclude that land cover, NDVI, and topography gradients are altering CCs’ distributions, as seen in Figs. 2a, 4a, and 5a. In terms of which of these three factors are more or less important, Figs. 2b, 4b, and 5b show (blue diamonds) CC% as a function of the varying fields as used within LSV. From these three plots, the more smoothly varying relationship seen between topography gradients and CC% (Fig. 4b) suggest that elevation gradient variability is quite important within LSV, with NDVI versus CC% (Fig. 5b) being of lesser importance given only a generally increasing CC% as NDVI variability increases. For land cover class, CC% is relatively invariant across land classes (Fig. 2b), suggesting that land cover variability is the least important field comprising LSV, and hence only subtly influencing CC patterns. However, forest-type classes, although not the most prominent over the SEUS, possessed the highest CC%, suggesting that sloped topography (a high variability in topography) and CC occurrence are coupled as forest classes predominate over highly sloped land.
Other conclusions include 1) CC occurrences over inland water bodies (e.g., lakes, rivers) were not conducive to CC development. However, as mentioned above, the combination of land cover types and topography gradient likely contributes to CC development and growth by forming mesoscale circulations similar to those of sea breezes, as a result of contrasting land surfaces (Walker et al. 2009; Asefi-Najafabady et al. 2010). 2) CC development over urban areas was high once ocean pixels were removed from the analysis in the subset. And 3), our analysis of NDVIs and CCs showed a positive relationship given the high forest-to-CC correlation noted above. Overall, CCs occurred more frequently with increasingly higher NDVI, despite the gradually decreasing percentage of pixels with NDVI values ≥0.7.
We conclude that regions possessing increased variability in land surface characteristics (lakes among forests, significant topography intermingled with agricultural land) enhance CC development. When land surface factors work in concert with synoptic and mesoscale convergent forcing, we speculate that the predictability of CC development (and likely convective storm occurrence) can be increased, such that the first CCs to develop on a given day are more likely to occur over heterogeneous terrain.
This research was supported by NASA Advanced Satellite Aviation Weather Products (ASAP) subcontract under NNL07AA00C. Comments from David Schultz, as well as two anonymous reviewers, significantly improved the quality of this paper.