1. Introduction
Thunderstorms are typically initiated by a convergence mechanism in moderate-to-high amounts of low-level moisture and instability. Once a vigorously growing cumulus cloud reaches the cumulonimbus stage, a convective thunderstorm is imminent that can produce heavy rain, lightning, strong winds, and/or hail (Roberts and Rutledge 2003). The unpredictable diurnal development, behavior, and evolution of these short-lived storms occur almost daily over summer (JJA), especially in humid regions like the Southeast United States (SEUS). Despite advances in meteorological research, the predictability of such storms in weak synoptic environments remains low relative to those in synoptic environments with stronger dynamic forcing mechanisms (i.e., quasigeostrophic lift and/or strong cold fronts). Although recent improvements in ensemble-based convection-allowing numerical weather prediction (NWP) models, in part through the assimilation of an array of meteorological observations (radar, satellite-based, aircraft, surface observations, etc.), have increased forecast skill, the resultant forecast skill for precisely predicting convective storm initiation (CI) remains relatively low (Coniglio et al. 2019). While CI is far less predictable than other meteorological phenomena, first-time CI events could be influenced by local land surface and meteorological phenomena. As a means of determining some of the underlying factors that lead to CI on typical summertime afternoons when synoptic-scale forcing is weak, an analysis of initial storm development locations was made to evaluate if any consistent spatial patterns emerge. If such patterns in CI exist, the study results could help benefit weather forecasters across the SEUS during the summer season by providing maps on where first-day CI may be more or less likely to occur across a given region.
Previous studies have investigated these mechanisms related to weakly forced CI. Two key discoveries were that upslope winds over heterogeneous terrain can induce mechanical lifting of air (e.g., Kuo and Orville 1973; Weckwerth et al. 2014) and land surface heterogeneity can produce daytime differential heating boundaries that support localized convergence, which can lead to convective initiation (Wilson and Schreiber 1986; Yan and Anthes 1988; Segal and Arritt 1992; Hong et al. 1995; Trier et al. 2004). A combination of these two mechanisms can act in unison to enhance CI (e.g., Wilson and Schreiber 1986). The SEUS has considerable mesoscale terrain features, such as the foothills of the Appalachian Mountains (spanning from New York to northern Georgia/Alabama) that cause variability in boundary layer depth and lead to differential heating. Given the proven significance of these convective catalysts, this study is motivated because a dual-summer radar-based CI analysis that incorporates geographical (e.g., Gambill and Mecikalski 2011) and nongeographical (meteorological) variables over the SEUS has yet to be performed. This study examines explicitly how distinct geographical features such as elevation (<1000 m) and land use influence weakly forced CI events over the SEUS. The meteorological variables of interest include antecedent rainfall, which can affect overland evapotranspiration (ET), and forecast model analysis fields [e.g., from the Rapid Refresh (RAP) model]. Convective studies on echo identification utilizing radar and/or satellite have been conducted over Atlanta, the Black Hills, Colorado High Plains, eastern New York/western New England, Brazilian Amazon, and Mexico (Haberlie et al. 2015; Kuo and Orville 1973; Klitch et al. 1985; Wasula et al. 2002; Lima and Wilson 2008; Giovannettone and Barros 2008).
With this central idea, several primary questions guide this study: 1) What are the cumulative spatial distribution patterns of summertime CI events across the SEUS? 2) Do higher amounts of rainfall over previous days correspond to greater frequencies of CI occurrence? 3) How important are the geographical features relative to the meteorological features in weakly forced CI over the study period? The primary goal of this 2020–21 summertime analysis is to help explain SEUS summertime CI in weak synoptic environments, which can enhance CI predictability. CI is defined in this and other previous studies on the topic as the first occurrence of a 35-dBZ echo at the −10°C level from a convective cloud, and this CI threshold has been used to track convective echoes over time (Roberts and Rutledge 2003; Mueller et al. 2003).
The guiding hypothesis here is that SEUS summertime weakly forced CI events would occur nonrandomly, with the static geographical features (terrain and land use) being more important overall in dictating whether weakly forced CI occurs than the meteorological features, especially early in the diurnal cycle. However, the connection between certain features of both types (e.g., topography and wind direction) would still manifest concerning CI occurrence, to a degree, both spatially and in the statistical distributions. Under this hypothesis, the meteorological features would also not show as significant of an overall importance due to their nonstatic and ever-changing temporal nature. If none of the examined features show any correlation with weakly forced CI occurrence over the study period, then it is to be inferred that CI events in synoptically weak conditions are randomly distributed.
To address the above hypothesis and associated science questions, this study analyzed gridded radar data during JJA to quantify locations of CI across a region of the SEUS, specifically in and surrounding the state of Alabama, where mesoscale forcing such as land–sea breeze circulations does not influence CI. The CI maps were then assessed for any nonrandom spatial patterns, and those patterns were then evaluated against various land surface, meteorological, and rainfall fields to determine if any of the observed CI patterns correlated to those fields. The end result is an analysis on which if any land surface, meteorological, and rainfall fields statistically explain why CI might be more favored to occur in some locations versus others, on summertime days with synoptically weak forcing.
2. Background
The summer season (late May–September) over the eastern United States brings lush vegetation, humid conditions (dewpoint temperature Td ≥ 17°C; Gambill and Mecikalski 2011), and near-daily thermodynamic support for deep convection. Ordinary or weakly forced cell storms typically occur in these environments with moderate-to-high convectively available potential energy (CAPE) > 1000 J kg−1 and low vertical wind shear < 20 kt (1 kt ≈ 0.51 m s−1) (Markowski and Richardson 2010). This type of convection is known for producing weak outflows (or “gust fronts”) from localized cold pools of dense air that propagate outward from storms. These conditions result in high bulk Richardson numbers, implying a dominance of outflow over inflow and a tendency for the storms to be short-lived (Weisman and Klemp 1982; Rotunno et al. 1988; Markowski and Richardson 2010) in comparison to longer-lived, larger-scale storms such as supercells (e.g., Wilson 1966). Ordinary cell storm occurrences peak in the early to midafternoon, in concert with maximum solar heating (e.g., Lima and Wilson 2008; Rickenbach et al. 2015). Since ordinary cells tend to be strongly aligned in the vertical direction (lack of vertical updraft tilt associated with wind shear), eventually, the increase in evaporative cooling and hydrometeor loading caused by precipitation accelerates the downdraft. The average lifetime of these weakly forced storms is ∼30–60 min, based on the duration of the air ascension to the anvil top, plus the average time it takes precipitation to reach the ground (Markowski and Richardson 2010). However, when this type of convection is weakly supported by wind shear or high instability, it can persist for ∼2–4 h.
Precipitation climatology studies over the SEUS have shown clear diurnal patterns (e.g., Haberlie et al. 2015; Rickenbach et al. 2015). Rainfall from mesoscale precipitation systems prevails in winter and spring (Figs. 1a,b), while maximum rainfall amounts attributed to isolated convection are seen in summer (Fig. 1c). Not only are spatial relationships in storm type observed in SEUS climatology but also are trends in intensity and duration. Brown et al. (2019) showed the dominance of short-duration precipitation events across the SEUS in the 1960–2017 period, with a rise in the number of short, intense rain events in more recent years. For some explanation, Findell et al. (2011) found that summer convective rainfall patterns in the eastern U. S. and Mexico are largely influenced by local surface latent heat fluxes, suggestive of a positive soil moisture–precipitation feedback process along the lines of this study’s main hypothesis (see also Tao et al. 2019).
Seasonal average precipitation amounts (mm day−1) due to isolated precipitation features over the Southeast U.S. from 2009 to 2012. Scale range of average isolated precipitation is depicted in the color bars. The four subplots depict each season: (a) winter (DJF), (b) spring (MAM), (c) summer (JJA), and (d) fall (SON). Figure adapted from Rickenbach et al. (2015).
Citation: Weather and Forecasting 39, 1; 10.1175/WAF-D-22-0216.1
Outflow collisions over favorable land or orographic areas have also been observed as a convective aide (Wilson and Schreiber 1986). In Lean et al. (2009), an isolated thunderstorm initiated in the south of England due to the combination of a convergence line induced by terrain, low-level moisture associated with a nearby front, and diurnal heating. Through radar, satellite, and surface observation data, Goggins et al. (2010) identified various frontal boundaries that could act as CI mechanisms during one summer in the National Weather Service (NWS) Birmingham County Warning Area. In Goggins et al. (2010), convection was categorized into autoconvection (a single boundary initiated without the aid of other boundaries) and convection initiated from the interaction of multiple boundaries. Out of all identified convergence boundaries in Goggins et al. (2010), topographical “boundaries” were found to be convective ∼49% of the time, a significantly lower percentage than for the identified synoptic-scale fronts. This contrasts with Koch and Ray (1997), which found that topographic boundaries were convective ∼80% of the time (likely due to sea breeze fronts near the coast, which were placed into the topographical category). In Lima and Wilson (2008), outflow propagation from existing convection was found to be a more prevalent CI mechanism in the late afternoon, implying that different mechanisms prevail over different diurnal periods. Convergent boundaries can also result from a land–sea diurnal pressure gradient (e.g., sea breeze).
Hirt et al. (2019) noted that by adding horizontal and vertical wind perturbations into 1-km cloud-resolving models, more CI events were induced over higher elevation gradient regions versus lower-gradient regions, with slower wind perturbations increasing CI likelihood further due to the maintenance of stronger updrafts (i.e., intense perturbations can weaken a convective updraft). Gambill and Mecikalski (2011) discovered that, on average, Geostationary Operational Environmental Satellite (GOES)-based SEUS convective cloud (CC) frequencies are positively correlated with elevation gradient despite the low number of high-gradient locations relative to flatter areas [see Figs. 2 and 4 in Gambill and Mecikalski (2011)]. Similarly, convective echo formation in parts of Mexico is dictated at the regional scale by topographic heterogeneity, minimal ocean–mountain range distance, and the orientation of the mountain ranges (Giovannettone and Barros 2008). Not accounting for regional topographical patterns has been found to lead to an oversimplification of general precipitation patterns (Brown et al. 2019). As with other regions, overall, SEUS cloud minima and maxima tend to shift with the time of day, with trends toward cloud minima in valleys and maxima near 1500 LT, and maxima over higher-terrain features after 1100 LT (Gibson and Vonder Haar 1990). Lima and Wilson (2008) found that steep terrain was the dominant early afternoon CI mechanism in central Brazil. Lake breezes can also locally drive CI. In North Alabama, Lake Wheeler commonly produces summertime thermal lake breeze circulations that result in precipitation adjacent to the lake (Asefi-Najafabady et al. 2012).
Land type and heterogeneity can locally enhance thermal buoyancy to influence CC formation (Gambill and Mecikalski 2011). For example, forest and savanna have been correlated with higher CC percentages relative to other classes in the SEUS (Fig. 2.3a in Gambill and Mecikalski 2011). Using land surface models, varying soil moisture and latent heat flux gradients cause thermal solenoidal circulations that result from land heterogeneity on scales of 10–20 km; however, they are dampened at smaller scales due to turbulent mixing superseding the pressure gradient force at that scale (e.g., Avissar and Liu 1996; Collins and Tissot 2015). Heating indices have quantified the influence of these thermally driven circulations by integrating land-use and GOES-R data, which are influenced by vegetation changes over time, the estimation of latent heating, and crop irrigation impacts (Walker et al. 2009).
ET significantly impacts low-level moisture content and local weather patterns in ensuing days, especially given high surface soil moisture and dense vegetation. ET is defined by the U.S. Geological Survey (USGS) as the process by which water is either lost to the atmosphere, evaporated from groundwater reservoirs, or transpired from plants. Recent precipitation helps raise areal ET levels, and higher antecedent rainfall totals are concurrent with greater low-level (below ∼900 hPa) moisture to help fuel surface-based convection. Other factors that exhibit a positive correlation with ET include temperature (plant stomata are more or less open) and surface wind (increased evaporation of moisture from plants). Relative humidity is negatively correlated with ET rate (lower relative humidity = better conditions for evaporation). While water bodies (lakes, rivers, oceans) are a dominant producer of atmospheric moisture (90%) globally, transpiration over land can contribute as much as 10% to low-level moisture.
3. Datasets
a. Radar observations
For CI identification, 2D radar reflectivity data at the −10°C level were obtained from the Multi-Radar and Multi-Sensor (MRMS) online archive (Zhang et al. 2011). The MRMS product suite integrates 146 WSR-88D radars and 30 single-polarization C-band radars across the continental United States and Canada. The MRMS −10°C isotherm surface is derived from a 3D Reflectivity Cube over each grid point using vertical temperature data from the RAP model analysis. Quality control eliminates non-hydrometeorological echoes from insects or frontal boundaries. MRMS data have a grid spacing of ∼1.11 km × 1.01 km (0.01° latitude/longitude) on a latitude–longitude grid projection. The data were gathered for each case day during peak solar heating (1600–0000 UTC).
b. RAP model analysis
Archived hourly 13.54-km grid spacing RAP model analysis data were obtained from the National Centers for Environmental Prediction (NCEP). The following RAP feature variables were used in this study and are listed with their CI feature subcategory: (i) 10-m wind direction (wind), (ii) local standard deviation of 10-m wind direction (Θstd; convergence), (iii) 10-m wind speed (wind), (iv) 2-m temperature (Ts; instability), (v) 2-m dewpoint temperature (Td; moisture, instability), (vi) surface-based lifted index (instability), (vii) surface-based CAPE (instability), and (viii) 950-hPa vertical motion (ω950; instability). These features have been shown to be important on convectively active days (Collins and Tissot 2015).
c. Elevation data
d. Land-use data
The 2019 Moderate Resolution Imaging Spectroradiometer (MODIS) MCD12Q1 land cover dataset was also used in this study. MCD12Q1 contains multiple subclassification sets, including six separate supervised global classifications from annual MODIS reflectance data (Friedl and Sulla-Menashe 2019). The Type 1 MODIS classification subdataset was selected [the annual International Geosphere–Biosphere Programme (IGBP) classification], containing 17 land class labels and one unclassified label (Table 1). The data are on a sinusoidal grid at ∼463-m grid spacing. Latitudes and longitudes were derived from the native coordinates. As with elevation data, land class data were upscaled by rebinning to MRMS resolution. Unlike the elevation data, the MODIS data cannot be simply interpolated as it is a classification dataset rather than a continuous field. A different rebinning technique was applied that finds the mode (most occurring land type) of each grid point and surrounding points. The rebinned grid contains the assigned land cover mode at each grid point. Since the standard deviation of the MODIS land-use data is an ill-defined indicator of land variability, it was not included as a feature.
MODIS land-use classifications and their respective numerical assignments.
e. Antecedent rainfall data
Daily web-archived antecedent rainfall from the NWS Advanced Hydrological Prediction Service (AHPS; Lin and Mitchell 2005) was obtained, as used in Walker et al. (2009), to assess local differential heating and moisture boundaries. The AHPS data are assimilated and mosaicked radar, rain gauge, and satellite rainfall data as obtained by several NWS River Forecast Centers (RFCs) nationwide and are available at a 4-km resolution on a polar stereographic grid projection. The files contain “observed” daily precipitation (sample interval from 1200 UTC on the previous day to 1200 UTC on the current day), “normal” climatological precipitation values, the difference between the two variables, and the respective relative frequencies. In this study, only the “observed” rainfall was analyzed. For each selected case day, the data of the previous 5 days were retrieved for daily 1-, 2-, and 5-day totals. To properly match each case day with the correct antecedent rainfall data, datetime similarity indexing was done between the MRMS and antecedent rainfall files. The data were cubically interpolated to the MRMS grid resolution using the Clough–Tocher curve minimization scheme, which has higher computational efficiency than other interpolation methods and continuous interpolated surface [Renka and Cline (1984)]. Any inadvertent negative values in the interpolated data were set to zero inches. From the five assigned files, three separate features were formed: 1-, 2-, and 5-day antecedent rainfall. The three rainfall totals remained static on the grid over their assigned case day.
4. Methodology
a. Study domain and event collection
The spatial analysis domain is shown in Fig. 2a and captures the diverse geographical features of Alabama, which allows for the identification of spatial trends of weakly forced CI occurrence with respect to any of these geographical regions. For the other statistical analyses, the entire domain was utilized for an increased sample size. The domain was chosen to capture maximum elevation features, large urban areas, and a diversity of land types and avoid major water bodies (e.g., Gulf of Mexico) since diurnal sea-breeze-induced CI is not a focus. Elevation and land-use data remained grid static for the entire study period.
(top) The red box indicates the spatial domain over which radar (for determining a CI event) and feature data were collected for each case. It encompasses eastern Mississippi and the majority of Alabama. Coordinate domain bounds are (90.00°, 85.45°W) (longitude) and (31.50°, 35.00°N) (latitude). (bottom) Map of Alabama counties for reference in the spatial analysis. The black stars show Birmingham (BHM), Huntsville (HSV), Muscle Shoals (MSL), Montgomery (MGM), and Decatur (DUC). Regional names are included, as referred to in the paper text.
Citation: Weather and Forecasting 39, 1; 10.1175/WAF-D-22-0216.1
Convective day selection involved the use of synoptic weather maps and radar data to determine whether a day was conducive for widespread weakly forced CI. The NEXRAD base radar reflectivity mosaic (Iowa Environmental Mesonet) data were examined from mid-May to early September during summers 2020 and 2021 for the occurrence of scattered-to-widespread weakly forced convection that appeared as individual cells of higher radar reflectivity. Given the lack of other discovered public MRMS archives at the time of analysis and the 1-day retention period of the archive used in this study, the MRMS data had to essentially be captured “semi-live” (within 24 h). There is an Amazon bucket archive (https://noaa-mrms-pds.s3.amazonaws.com/index.html) that currently retains data back to October 2020 (at the time of writing), so it does not have archived data for summer 2020. For each prospective case day, Storm Prediction Center morning mesoanalyses were used to confirm the presence of calm synoptic conditions over the study domain. Brown and Arnold (1998) defined a weak synoptic environment by the following criteria in their investigation of convective cloud clustering near land surface–induced boundaries over Illinois: no frontal mechanisms within 500 km, 500-hPa winds < 7.5 m s−1 (∼15 kt), surface winds < 5 m s−1 (∼10 kt), and surface dewpoint temperatures (Td) > 17°C. These criteria were applied in this study with a slight extension of the 500-hPa wind threshold (<20 kt) over the study domain. A total of 36 case days were selected (Table 2). Given that the summers of 2020 and 2021 were not unusually dry, this sample represents typical nondrought summers. All features for each case day were subcategorized into hierarchical clusters based on their interfeature correlations (Fig. 3), with the Ward linkage distance between clusters calculated based on variance minimization (Müllner 2011). Notable collinear feature clusters are elevation–elevation gradient, wind speed–wind direction standard deviation, antecedent rainfall, and Td–lifted index (LI)–surface-based CAPE. The negative correlation of LI with Td and CAPE is due to its negative scale with instability (Fig. 3). Other broader clusters include land use–elevation–elevation gradient and wind direction–Ts–ω950.
The 36 case dates selected for the CI analysis. Days are sorted by month.
(a) A hierarchal clustering dendrogram of similar features selected for the present study using the Ward’s linkage algorithm, with linkage distance (smaller distance at node = stronger clustering) on the y axis and (b) corresponding Spearman correlation matrix (showing feature intercorrelation) of those same features. Common clusters are displayed as a single color in the hierarchy, with the y axis as the linkage distances. Feature data from the 1600–1900 UTC period (early day group) across all case days were used in the making of these two visuals.
Citation: Weather and Forecasting 39, 1; 10.1175/WAF-D-22-0216.1
b. Data rescaling
For each case day, hourly RAP model fields from 1600 to 0000 UTC were obtained. With differing RAP model and MRMS radar temporal resolutions, each hourly analysis (except for 0000 UTC) was assigned to the three 15-min intervals after the analysis time at the top of each hour (00th–45th min). The 45th–60th-min interval was assigned the following hour analysis since this interval is closer to that hour. As an example, the 1915–1930 UTC interval was assigned the 1900 UTC analysis and the 1945–2000 UTC interval was assigned the 2000 UTC analysis. An assumption made during this process is that each hourly analysis is steady state (static) over its assigned intervals.
The RAP data were reshaped to the MRMS grid spacing using a cubic interpolation with the piecewise Clough–Tocher triangulation curve minimization scheme (Renka and Cline 1984). After rescaling, the grid spacings matched the radar data domain shape (350 points meridional, 455 points zonal, 389 km × 430 km, respectively) before conducting the feature intercomparison. The zonal distance across the grid varies slightly due to Earth’s curvature but is still regularly spaced in latitude/longitude coordinates. Thus, for the feature datasets that are coarser than the MRMS grid (fewer data points), an interpolation scheme was applied. On the contrary, for the datasets finer than MRMS (more data points), a rebinning method was applied.
c. CI definition and event counts
The 35-dBZ at −10°C CI definition was used on a point-by-point basis over the MRMS grid space. The 35-dBZ precipitation intensity threshold has been applied in prior studies to build CI nowcasting algorithms (Mueller et al. 2003; Mecikalski et al. 2015), to define minimum storm area for tracking convective echoes (e.g., Lima and Wilson 2008; Patou et al. 2018), and as a thunderstorm tracker used for determining CI and monitoring hail cores over the European Alps (Nisi et al. 2016, 2020). Some studies have defined CI using a radar reflectivity threshold of 30 dBZ at the 1-km height level (Schreiber 1986; Wilson and Schreiber 1986); however, convective cloud-top glaciation and ice nucleation processes often occur at the −10°C level at the time of CI. Both are strong indicators of ongoing and/or deepening convection and the formation of convective precipitation (e.g., Schreiber 1986; Roberts and Rutledge 2003; Mecikalski and Bedka 2006).
Over each 15-min interval, a grid point is assigned a binary value of “1” if CI occurred at any point during that interval (e.g., precipitation echo intensity at −10°C reached or exceeded 35 dBZ in at least one instance). To minimize positive CI instances with existing echoes that span multiple grid points, a CI event is only counted at a grid point if no adjacent points reach or exceed 35 dBZ. Once all data were collected, it was grouped into early (1600–1900 UTC), middle (1900–2200 UTC), and late (2200–0000 UTC) day CI bins to effectively compare the dominant CI mechanisms in each of these timeframes. Although convergent boundaries are a common CI mechanism, no boundary identification was performed; this is obviously not to say, however, that there were not any gust front–type features or other kinds of boundaries present.
d. Random forest model for feature importance analysis
One study component is a quantitative evaluation of feature importance with respect to the binary occurrence of CI using machine learning. Two datasets are needed in a machine learning model: the features (sample data) and the target class(es) (the variable to be predicted or classified, in this case CI). Given an initial set of data to train the model, machine learning classification is intended to produce a set of predicted target outcomes from a separate set of test data. The random forest model provides a predictor importance assessment, which is composed of multiple decision trees. During the model training, each decision tree is trained on various subsamples of the input training data rather than the entire dataset at once. When each tree is formed, ∼20% of the input training datasets are not included (so-called “bootstrapping”). With an initial specified set of model hyperparameters, the bootstrapped training prediction/classification sets of each subsample are then averaged over all decision trees to optimize the accuracy of the model, reduce overfitting, and cancel out prediction errors resulting from the variance of the decision trees (Pedregosa et al. 2011).
One metric of computing feature importance is based on Gini impurity (Breiman 2001), which ranks features based on the average decrease in node impurity over all decision trees and their relative node-splitting contribution. A second method valid for classification is permutation-based importance (Breiman 2001), which is a metric representing the average importance obtained through feature shuffling and rerunning of the model over a set number of iterations. This latter method avoids issues associated with Gini impurity, such as being strictly confined to a training set and an inherent bias toward numerical features (many unique values). Both methods were implemented for comparison in this study.
The optimal combination of hyperparameters for random forest model configuration was determined through an exhaustive grid search over a specified parameter space with threefold cross-validation. Fine-tuning the model through multiple grid searches is unnecessary, especially if the mean cross-validated model scores do not show a clear positive correlation with the model complexity (e.g., greater tree depth, a higher number of trees). Also, given the large analysis sample size (>106 total samples across all case days), the necessity of multiple rounds of hyperparameter optimization is not fully justified. Instead, a threefold stratified K-fold cross-validation technique was applied to a single hyperparameter grid search on the training data of the late (2200–0000 UTC) day CI group to assess the degree to which the cross-validation scores improve with increases in the maximum tree depth and the number of trees (test parameter ranges of 5–20 branches and 50–200 trees, respectively). The optimal combination of hyperparameters was 200 decision trees and a maximum tree depth of 20, which was subsequently used for all model runs.
Before implementing the random forest model, the following adjustments were made to the features: conversion of Ts and Td to degrees Celsius (°C), rounding of the interpolated lifted indices to the nearest integer and masking negative (unrealistic) antecedent rainfall totals to a value of zero. The following model runs for each time group were conducted: 1) all features included with four different bootstrapping random states over the study domain (Gini importance), 2) three most important and two least important features from the first four runs were excluded at a single random state over the study domain (Gini importance), and 3) one feature from each hierarchal cluster in Fig. 3a to calculate permutation-based importance.
5. Results
a. Atmospheric analysis
The Birmingham, Alabama (BMX), sounding at 0000 UTC 20 July 2020 shows modest instability (<1000 J kg−1; Fig. 4a), typical of weak synoptic days in the SEUS. A well-mixed convective boundary layer is seen with near-adiabatic temperatures to the lifting condensation level (LCL). Note the extremely shallow temperature inversion near the surface as temperatures began to cool along with the small amount of CIN (listed to the right of the sounding), resulting from the waning solar radiation in the evening hours. Wind speeds throughout much of the troposphere are <20 kt, with winds < 10 kt up to 400 hPa.
(a) A sounding at 0000 UTC from Birmingham (BMX) on 20 Jul 2020 with the surface-based parcel profile (black line) and corresponding CAPE (red shade). Sounding data were obtained from the University of Wyoming archive (weather.uwyo.edu/upperair/sounding.html). Spatial plots at 1800 UTC 20 Jul 2020 include the following: (b) MRMS reflectivity at −10°C (dBZ), (c) NCEP Reanalysis (NARR) 2-m dewpoint (°C) with 10-m wind barbs (kt), and (d) NARR 500-hPa geopotential height (m) with 500-hPa wind barbs (kt). NARR data were provided by the NOAA/OAR/ESRL PSL, Boulder, CO: https://psl.noaa.gov/data/gridded/data.narr.html.
Citation: Weather and Forecasting 39, 1; 10.1175/WAF-D-22-0216.1
The conditions at the surface and 500 hPa at 1800 UTC (1300 LT; Figs. 4b–d) are also representative of the synoptic situations from all sampled days. MRMS −10°C reflectivity at ∼1800 UTC shows the spatial distribution of convective echoes. In Fig. 4b, weakly forced convection formed by 1800 UTC over Huntsville, Alabama; Birmingham; and other scattered areas in the Cumberland Plateau, with light surface winds of ≤5 kt being present (Fig. 4c). These small wind magnitudes have a rather insignificant impact on the motion of ongoing convective storms. Overlaid in Fig. 4c is 2-m Td, with widespread values > 19°C (∼66°F) across most of the domain. In central and northern Alabama, CI events were clustered over and near locations with higher Td values. Figure 4d shows 500-hPa wind speeds, which are no greater than 15 kt and more geostrophic to the south, indicative of the presence of an upper-level ridge. The shallow geopotential height gradient across Alabama (increase of ∼30 m from southwest to northeast) stems from the lack of a significant synoptic-scale pressure gradient. A defined shortwave trough propagated across the upper Midwest and Ontario on 20 July 2020, with weak synoptic conditions prevalent over the SEUS.
b. Elevation and land use
Although Alabama does not have mountain ranges of the caliber seen in the western United States, it does have modest elevation features scattered throughout the state (reaching 735 m). A swath of ridges and valleys, denoted here as the “valley–ridge” region, extends from near Birmingham to just north of the Piedmont region to the south (Fig. 5a). Higher elevation gradients are found across the valley–ridge and Piedmont regions, as well as across north-central and northwest Alabama as compared to areas south. Isolated higher-elevation features are also present near Huntsville (Fig. 5b). To the west is the Highland Rim region, spanning northwest Alabama and the Tennessee River Valley. Closer to the Gulf coastal plain in the southern part of the domain, elevation is relatively low, and no significant topographical features exist.
Spatial plots of (a) CRM elevation and (b) calculated elevation gradient. Units are in meters for elevation and meters/pixel for elevation gradient. Circled regions are Highland Rim (red), Cumberland Plateau (brown), valley–ridge (yellow), Piedmont (maroon), and coastal plain (purple). (c) MODIS land-use data over Alabama. Classifications are grouped as follows: forestland (blue), shrubland/savanna (dark green), cropland/vegetation and urban area (light green), and water bodies (yellow).
Citation: Weather and Forecasting 39, 1; 10.1175/WAF-D-22-0216.1
Rebinned MODIS land-use classifications are shown in Fig. 5c. Although water only makes up 0.528% of the 159 250 total grid points, larger distinct water bodies (yellow/light blue shading) include Lake Wheeler (Highland Rim region), Lake Guntersville (Cumberland Plateau region), and Weiss Lake (valley–ridge region). The light green shading represents developed urban areas (Huntsville, Birmingham, and Montgomery, Alabama) and cropland/vegetation (prevalent in northwest Alabama), respectively; these make up 0.907% and 2.61% of all grid points. Savannas and woody savannas (classes 8 and 9) are by far the most prevalent classes, together occupying 55.5% of all grid points. Each forest class has a unique shade of blue in Fig. 5c. Needleleaf forests (evergreen, deciduous) make up just 0.399%, while broadleaf forests (evergreen, deciduous) make up 23.3% of all grid points. Mixed forest makes up 14.8% of all grid points. As a result of the vastly greater frequencies of certain land classes relative to others, a potential bias arises with regard to total CI event count for each land class. Regridding the MODIS data to the MRMS resolution may have also played a role in this since the rebinning technique used found the most common land type over each bin and assigned it to the rebinned grid point. Since the overall event counts for the savanna classes will be significantly greater than the other classes, the utilization of a “relative CI percentage” for the static datasets was more appropriate for this analysis.
c. Cumulative CI event heat maps
Figures 6a–c show a set of “heat maps” of CI events for each time group with the MRMS grid resolution decreased by 75% (sums were calculated in ∼4 km × 4 km bins over the domain). In Fig. 6a, several focus areas of higher event density appear for the early time group (27 917 early CI events, 2326 events per 15-min interval on average). The first is region 1 in Fig. 6a. Although no significant elevation features exist in this area, it has a higher land proportion of forest than surrounding areas (Fig. 5c). Another CI focus area occurs over parts of the valley–ridge and Cumberland Plateau regions, extending from Shelby into Dekalb Counties. From Fig. 5b, this higher-density swath roughly parallels a southwest–northeast path of higher elevation gradients. Region 2 (Fig. 6a), which has some of the highest elevations in Alabama, has the densest event clusters for the early CI time group across the entire domain. Lesser clustering is seen in the Piedmont region. A third area of higher event density is in far northwest Alabama, to the west and south of Florence–Muscle Shoals (region 3). No prevalent land type or large elevation features exist here, except for a few minor spots of enhanced elevation no higher than 120 m. The southern third of the domain lacks early CI events relative to the northern parts (4913 early CI events below 32.5°N, only 17.6% of all early events). To get a better sense of small-scale spatial trends around Huntsville and the Tennessee Valley area, a zoomed-in visualization is shown in Fig. 6a. A narrow strip of higher event density extends west–east through Huntsville (region 4). Another area of higher CI event density is in northeastern Jackson County (region 5), where significant elevation gradients and heavy forest cover exist. A small cluster just south of Lake Wheeler (region 6), west of Decatur, is also seen where the dominant land type is cropland with flatter elevation. On the north side of the lake, there is a lack of CI events, most notably in the early and middle time groups. This could be attributed to a prevalence of background winds with a southerly component over all case days, which supports localized convergence with the lake breeze south of the lake.
Spatial heat maps of CI events tallied in ∼4 km × 4 km bins for (a) early (1600–1900 UTC), (b) middle (1900–2200 UTC), and (c) late (2200–0000 UTC) time groups over all case days. The region shown is the Alabama portion of the domain. The right plots focus on the Tennessee River Valley region (blue box; includes Huntsville). Other CI regions of interest are indicated by the black boxes. Color scale represents the total CI events per histogrammed bin (each of which contains multiple grid points) over all 36 case days in that specific time group.
Citation: Weather and Forecasting 39, 1; 10.1175/WAF-D-22-0216.1
The middle time CI group consists of a vastly greater number of overall event counts compared to the early group in Fig. 6b (50 002 middle CI events, 4167 events per 15-min interval on average). It should be noted that some of these middle CI events likely stemmed from outflow boundary interactions. In addition, the north–south event distribution is more balanced than the other two groups (25 808 and 24 194 middle CI events above and below 33.25°N, respectively). The increase in weakly forced CI formation is evident during the midafternoon hours, with one focus area located over Jefferson, Shelby, and Blount Counties (region 7 in Fig. 6b) at the southwest end of the Appalachian foothills. A notable trend from the early CI time group is that those areas which stood out in terms of event density during the early CI group now exhibit lower counts relative to surrounding areas. Atmospheric recovery of instability from previous convection likely influenced this spatial trend, with neighboring areas that had less early convection now experiencing increased events. Western Jackson County has the highest-count cluster of the domain (region 8), where several bins have more than 20 event counts. Of the 4241 middle CI events over the Tennessee Valley inset region in Fig. 6b, this cluster encompasses 1530 of those events [contributing 36.1% of all events inside this inset region, with the cluster defined within the latitude/longitude bounds (34.5°, 34.9°N), (86.7°, 86.0°W)]. Another higher-density area of events is present in the southeast corner of the domain, where weak clustering had already occurred in the early CI time group (Region 9). The area of higher event density in northwest Alabama, which also had relatively higher event densities in the early time group, shifted eastward to overlay the Florence–Muscle Shoals area. In the Tennessee River Valley, the higher event density within Madison County has shifted southward from the early group (region 10), while the signal south of Lake Wheeler is now less evident. Meanwhile, the area north of the Tennessee River continues to have extremely low CI events relative to surrounding areas.
A significant drop in overall domain event density occurs in the late CI time group because of the decrease in solar heating and instability (Fig. 6c). The stout drop in event counts is evident over the entire domain (24 469 late CI events, 3059 events per 15-min interval on average). Compared to the early CI time group, a proportion of CI events nearly twice as high is found in the southern part of the domain (7660 late CI events below 32.5°N, making up 31.3% of all late domain CI events). Most notable is the cluster over Jefferson and Shelby Counties, including the Birmingham metropolitan area (region 11 in Fig. 6c). This is indicative of an urban heat island (UHI) effect influencing CI occurrence in the early evening hours [988 late CI events in this cluster defined within the latitude/longitude bounds (33.25°, 33.75°N), (87.20°, 86.60°W), composing 4.04% of all late CI events], which is certainly a nonnegligible fraction considering the size of the cluster relative to the domain. Focusing on the Tennessee River Valley, a marked decrease in CI events from the earlier time frames occurs over region 10. A smaller cluster of higher events appears south of the Huntsville area near the Tennessee River.
d. Feature importance analysis
Tables 3–5 show feature importance values of all three CI time groups using models with training data from all 36 case days over the entire domain. The goal was to establish the key features contributing to CI frequency throughout the analysis period. When incorporated into a random forest classification model, each feature (or predictor) has a unique degree of importance in determining the class labels of that model. Instead of a spatial analysis, CI frequency distribution, or a CI versus non-CI feature sample comparison, all predictors are utilized for training the model with a constant set of hyperparameters. Since only the importance rankings were desired (and not CI predictions), the entire dataset was used for the model training. A higher Gini importance value (unitless scale from 0 to 1) in a feature implies higher average node purity and a more balanced feature split during the class labeling. Due to the bias of the Gini importance toward high cardinality values (e.g., RAP model fields), separate categorical and noncategorical feature comparisons are necessary for the Gini importance metric.
Early day group Gini feature importance values over the entire domain for four different random states, a run with the three most important and two least important features from the left four runs, and average permutation feature importance for a singular random state over five repetitions. Bolded numbers indicate top-ranking features relative to the others within a given predictor importance column. Here, Θstd is the standard deviation of the wind direction, CAPE is the convective available potential energy, and ω950 is the 950-hPa vertical motion from the Rapid Refresh forecast model analysis.
By order of Gini ranking in the early group (Table 3), the three highest-ranking features across all four bootstrapping states are surface-based CAPE, 2-m Td, and Ts. The dewpoint temperature Td is primarily a moisture indicator while the other two are instability indicators. Although surface-based LI is also an instability indicator, it ranked lower because of its limited scale range compared to the top three features. Considering the two topography features, elevation consistently ranks nearly double in Gini importance magnitude than elevation gradient. The two least important features, in order of lowest Gini importance, are land use and 1-day antecedent rainfall across all four random states. With land use, the main contributor to its bottom ranking is likely its limited categorical scale range. In the run excluding the top- and bottom-ranking features (set to a singular random bootstrap state in all three time groups), the three top-ranking features are LI, 10-m wind speed, and elevation. These top-ranking features are a mix of feature types including instability, wind, and topography. The two bottom-ranking features are 10-m wind direction and elevation gradient. Out of the five selected features for the early permutation run, the top two features are 1-day antecedent rainfall and Θstd. These results contrast with the early Gini runs, where these two features are ranked in the bottom and middle tiers, respectively.
In the middle time CI group, the Gini importance rankings do not stray greatly from the early group (Table 4). One minor difference does appear as the third and fourth rankings are swapped (ω950 and Ts). Again, all top-three features are some form of an instability indicator. The bottom two rankings continue to be land use and 1-day antecedent rainfall. For the antecedent rainfall features, 1- and 2-day antecedent rainfall importance results are lower than the early CI group, while 5-day antecedent rainfall importance is higher. Gini importance of elevation is also lower than in the early CI group, supporting the idea that this feature is a more CI dominant mechanism earlier in the day. In the middle CI group exclusion run, the three most important features are surface-based LI, Ts, and 10-m wind speed, while the bottom two are 10-m wind direction and 2-day antecedent rainfall. Compared to the exclusion run of the early day CI group, two of the top three features and the bottom-ranking feature remain the same. Out of the five features included in the middle day permutation run, the top two features are 1-day antecedent rainfall (as in the early day group) and 10-m wind direction. Elevation ranks in the bottom two in permutation importance for both the early and middle day groups, contrary to the strong early spatial and statistical correlations in this study.
Importance results for the late day CI group are shown in Table 5. In order of Gini importance, the top three features across all four bootstrapping states are 2-m Td, ω950, and surface-based CAPE. These are the same top-ranking features as the middle day group, and only now, ω950 has ascended to second-most important. One notable trend is a marked decrease in importance of the surface-based LI from the earlier time groups. Elevation is more important than in the middle day CI group but less important than in the early day CI group, while elevation gradient Gini importance is actually higher than in both the early and middle groups. As with the other two time groups, the two lowest-ranking features are land use and antecedent rainfall. For the late day exclusion run, the three most important features are wind speed, elevation, and surface-based LI. The two least important features are 10-m wind direction and 2-day antecedent rainfall. Out of the five features included in the late permutation run, the top two late features are wind direction and Θstd. The most notable trend from the early and middle day groups is the decreased permutation importance of 1-day antecedent rainfall, and with the top two being wind direction features, it suggests an increased significance of thunderstorm outflow for the later time period. Moisture (2-m Td) is also less important than in the earlier time groups.
6. CI statistical relationships to important features
The following section overviews aspects of several fields highlighted in the feature importance analysis with respect to their relationships with CI distributions. While none of the features outlined below reach a level of statistical significance when explaining CI frequency distributions, important patterns are still worthy of discussion.
a. Antecedent rainfall
Since antecedent rainfall grid distributions evolve day-to-day, cumulative CI versus non-CI (1 vs 0) distributions can measure differences between events and nonevents. To effectively compare the much larger non-CI to the smaller CI distribution in each time group, a random sample of the non-CI events with the same sample size as the CI events was used, enabling fair comparisons between the distributions and time groups. Since the distributions of 1-, 2-, and 5-day antecedent rainfall were similar, only the 5-day antecedent rainfall is discussed here.
Early, middle, and late CI versus non-CI distributions for 5-day antecedent rainfall over the whole domain are shown in Fig. 7. In the early distribution, the CI sample mean is greater than the non-CI sample mean (0.85 vs 0.76 in.). Discrepancy between the two middle samples is greater than between the early samples, evidenced by further separation between the middle CI and non-CI means (0.92 vs 0.76 in.). The mean difference between the late CI and non-CI samples is nearly identical to that of the early distribution (0.85 vs 0.76 in.). Regardless of the decrease in contrast of the two samples compared to the middle distribution, the CI sample prevails as having higher antecedent rainfall on average relative to the non-CI sample across all three time groups, also the case for the 1- and 2-day distributions (not shown). Last, although the CI sample mean is greater for all three time groups, the middle CI group emerges for the three antecedent rainfalls with the widest CI versus non-CI mean differences.
CI vs non-CI 5-day antecedent rainfall boxplots for all three time groups. For comparison, a random sample of non-CI events of the same size as the corresponding CI sample was extracted from the dataset. Means (green triangles), medians (orange lines), and confidence intervals (notches around the median) are shown. Whiskers are 1.5 times the interquartile range (IQR).
Citation: Weather and Forecasting 39, 1; 10.1175/WAF-D-22-0216.1
b. Surface-based CAPE and 2-m Td
On average, CI events had higher antecedent rainfall and associated ET than nonevents, where ET can influence instability. Unlike the antecedent rainfall distributions, the surface-based CAPE distribution means (Fig. 8) reside closer to their respective medians. This instability feature can vary substantially, both spatially and temporally, resulting in the balanced nature of these distributions as opposed to most sample points clustering in the bottom quartiles. The early day CI group has a pronounced difference between the CI and non-CI means, even more than the middle and late day groups (2694 vs 2128 J kg−1). A key difference between the early and middle CI samples is that the upper quartile of the early CI sample extends further than the non-CI sample, up to near 4500 J kg−1. The means of both middle samples are lower than their early day counterparts, with a potential factor being low-level evaporative cooling or cold pools from existing convection which increases atmospheric stability. Regardless, the same CI/non-CI trend persists between the two middle samples (2336 vs 1963 J kg−1). In the late CI sample, the extent of the lower quartile is narrower than in the other two time groups, while the upper quartile is slightly lower than the middle group. The means of both late samples retract further during this timeframe, which can be attributed to both existing convection and, more importantly, the waning of the diurnal heating in the early evening hours. On average, the late CI sample still has higher surface-based CAPE than the late non-CI sample, the smallest discrepancy of the three time groups (2089 vs 1806 J kg−1).
CI vs non-CI surface-based CAPE boxplots for all three time groups. For comparison, a random sample of non-CI events of the same size as the corresponding CI sample was extracted from the dataset. Means (green triangles), medians (orange lines), and confidence intervals (notches around the median) are shown.
Citation: Weather and Forecasting 39, 1; 10.1175/WAF-D-22-0216.1
Since Td and surface-based CAPE tend to be interrelated, it should be no surprise that their CI versus non-CI trends are similar (Fig. 9). The early CI sample mean is higher than the early non-CI sample (23.2° vs 22.1°C). The Td distributions in the middle CI group are like their respective counterparts from the early CI group. The difference between the middle CI and non-CI means is slightly higher (22.8° vs 21.6°C), both individually lower than their counterparts from the early time group. Both of their spreads are higher than the early group. The CI sample having a higher mean Td continues into the late time group (22.7° vs 21.8°C), the smallest difference of the three time groups. A wider extent of the non-CI samples in the Td distributions is more apparent than in the surface-based CAPE distributions. The maximum mean differences of CAPE and Td in the early time group suggest the heightened importance of atmospheric instability, where the higher mean antecedent rainfall of the CI samples could partially relate to this trend.
CI vs non-CI 2-m dewpoint temperature (Td) boxplots for all three time groups. For comparison, a random sample of non-CI events of the same size as the corresponding CI sample was extracted from the dataset. Means (green triangles), medians (orange lines), and confidence intervals (notches around the median) are shown.
Citation: Weather and Forecasting 39, 1; 10.1175/WAF-D-22-0216.1
c. 950-hPa vertical velocity
The box distribution plots for ω950 are not shown; however, the key results are synthesized here. The extent of the early CI sample is slightly wider than the early non-CI sample, with all samples spanning both positive and negative values. On average, the early CI sample has a slightly more negative ω950 compared to the early non-CI sample (−0.03 vs −0.02 Pa s−1). Compared to the early time group, the difference between the middle day distributions is more evident (−0.090 Pa s−1 for CI vs −0.010 Pa s−1 for non-CI). Trends from the early group are also not similar, with the non-CI sample mean only increasing by ∼0.01 Pa s−1. The CI versus non-CI gap widens even further in the late day group, shown in their respective means (−0.13 vs −0.01 Pa s−1). While the late group CI mean drops by ∼0.04 Pa s−1 from the middle group, the late group non-CI mean holds relatively steady. These trends show that the CI samples typically have lower ω950 (slightly stronger upward motion, or less subsidence) compared to the non-CI samples, which coincides with the higher average surface-based CAPE for the CI samples.
7. Discussion and conclusions
This study was centered around the initial hypothesis that summertime weakly forced CI events in the SEUS occur nonrandomly and that static features are most important in dictating weakly forced CI occurrence in weak synoptic environments; meteorological features would show a lesser correlation with CI partially due to their nonstatic nature. If no features are correlated with these CI events, then they are mostly expected to be randomly distributed across the region. It is noted that based on the heavily overlapping distributions in the boxplots (Figs. 7–9) and that the importance values are low overall (<0.20; Tables 3–5), none of the above findings are statistically significant; however, they do demonstrate that the features analyzed influence CI occurrence in synoptically weak environments. Tests performed to quantify the degree of nonrandomness using a distance-based statistical method for planar point patterns (Clark and Evans 1954) concluded that over the entire geographical domain of this study, the CI distributions did not differ significantly from a purely random distribution. However, across subdomains, namely, those shown in Fig. 6, a degree of clustering was found in the CI distributions. This clustering, while subtle at times, reflects the ability of the predictor importance analysis to isolate some increased importance in some of the predictor fields analyzed.
With the entirety of the results discussed, the main conclusions are as follows. First, several spatial patterns emerge in the cumulative CI heat maps, with a few standing out. Northeast Alabama shows CI event clusters in several different areas in the early and middle day groups. Jefferson County, including Birmingham proper, contains the largest cluster in the late group (likely attributed to an UHI effect). Second, there is an early CI occurrence relation with elevation, as apparent in the spatial distributions and its feature importance ranking in the early day CI group. Thus, the significance of elevated differential heating and orographic lift is seen in the early afternoon hours. Third, on average, weakly forced CI events occurred in moister and more unstable conditions than in non-CI instances across all three time groups, which can lead to a lowering of cloud LCLs and/or increased instability, making it easier for ascending air parcels to reach their LFC (especially for short-lived storms which tend to have more narrow updrafts). The moisture and instability indicators ranked high among all features in the Gini importance rankings. Fourth, other minor CI clusters are seen elsewhere across all time groups, again indicating more of a nonrandom nature of the spatial distribution of CI events. Both static and nonstatic features also show discernible statistical discrepancies between instances with and without CI events (e.g., antecedent rainfall and instability). The overall conclusion from this study is that summertime CI events in weakly forced synoptic environments in the SEUS occur largely in a random manner; however, the land surface and other meteorological conditions do show a degree of influence in CI occurrence.
Figure 10 synthesizes the results into a conceptual model that depicts a typical summertime weakly forced convective pattern, with horizontal south-to-north veering of the wind direction from a dominant southerly component to a dominant westerly component in a CI location (left red-circled area). Of course, this wind pattern is different over all CI days and will shift the favorable CI areas from day-to-day, as it has been shown here that static features are not the sole mechanism. Figure 11 conveys one such combination favorable for CI, containing locally higher amounts of antecedent rainfall and latent heat flux if it is collocated with or near higher terrain and a favorable low-level wind direction. This combination of enhanced latent heat flux from ET and relatively steep elevation may result in mesoscale differential heating circulations favorable for CI (e.g., Segal and Arritt 1992; Walker et al. 2009). On the other hand, a location may have locally high antecedent rainfall but be situated in an unfavorable spot when it comes to whether orographic lift can occur (Fig. 10; red-circled area in lower right).
A 3D conceptual model that relates different features based on the results in this study, which in tandem can cause nonrandom CI spatial distributions. Red circles represent areas of locally enhanced antecedent rain relative to surrounding areas, while the blue arrows show the direction of the background 10-m wind. The north (“N”) direction is indicated by the gray arrow. Formation of storms in the most favorable area is shown by the thunderclouds symbols.
Citation: Weather and Forecasting 39, 1; 10.1175/WAF-D-22-0216.1
A schematic of a mesoscale differential heating circulation over land. The different components are labeled accordingly, with the green slope representing an elevation feature. The term L is the length scale,
Citation: Weather and Forecasting 39, 1; 10.1175/WAF-D-22-0216.1
This study has several sources of error and limitations. First, the presence of mesoscale convergent boundaries (e.g., thunderstorm outflow) can locally enhance CI event frequency, acting as the dominant mechanism particularly in the later hours; this kind of mechanism was not directly accounted for in any CI feature analyzed in this study. Another limitation is missing CI counts at a grid point over a 15-min interval. Since the methodology was designed such that a grid point can only have a maximum of one tally per interval and CI can occur at any instance (even within the MRMS interval of 2–3 min), missing CI counts can underestimate the true CI count. In contrast, there were also likely instances where a CI tally was wrongly counted when it was part of an existing echo (even with the applied methodology). Therefore, these two effects would tend to balance the other.
In the MRMS data, there are inherent error biases in the individual radar components (e.g., transmitter, antenna, receiver) that affect the reflectivity return data despite the avoidance of other radar issues such as bright banding, beam broadening, and the “cone of silence” (area near/above a radar where the beam cannot reach when sampling data during scans) (Zhang et al. 2011). There are also discrepancies between the observed and RAP model temperature profiles, which can affect, even slightly, the vertical level at which the isotherm is found. In the MODIS data, the wetland classification tends to be underrepresented in the dataset, croplands are underrepresented in areas where the grid pixel sizes are much larger than average crop field sizes, and some grassland areas are incorrectly classified as mixed grass–tree savannas (Friedl and Sulla-Menashe 2019). This latter issue can be important to the study domain as the savanna land classification is the most prominent land class over Alabama. The AHPS antecedent rainfall has the following concerns involving the assimilated radar data: the presence of frozen hydrometeors, radar calibration error, varying validity of implemented Z–R relationship, and the presence of beam obstructions (DOC/NOAA/National Weather Service 2005).
Future work building on this study could include examining the impacts of additional features on CI occurrence. For example, the relation of 10-m wind direction to elevation gradient through the implementation of a normalized dot product can provide better indications of low-level orographic CI forcing (Nair et al. 2008). The land surface variability index from Gambill and Mecikalski (2011) could also be implemented to enhance CI distribution understanding, requiring MODIS land use, elevation data, and a vegetation dataset. Additionally, random forest runs can be conducted over smaller regions of interest to find and compare the most important features within them. The study advances our understanding of SEUS summertime weakly forced CI event patterns in synoptically weak and humid conditions, which can benefit CI nowcasting. Using the most important features assessed, convective indices could also be formed that more clearly describe the real daily CI likelihoods across a region.
Acknowledgments.
The funding for this research was provided by the University of Alabama in Huntsville (UAH) Earth Systems Science Center (ESSC) through ESSC Director Dr. John R. Christy.
Data availability statement.
For this study, the Multi-Radar Multi-Sensor (MRMS) −10°C radar reflectivity mosaic data were downloaded from the MRMS data web archive (mrms.ncep.noaa.gov/data/2D/Reflectivity_-10C/). Antecedent rainfall data were obtained from the Advanced Hydrological Prediction Service data download page (AHPS; water.weather.gov/precip/download.php). The Coastal Relief Model (CRM) elevation data were gathered from the National Centers for Environmental Information (NCEI) CRM page (www.ngdc.noaa.gov/mgg/coastal/crm.html). All of the Moderate Resolution Imaging Spectroradiometer (MODIS) MCD12Q1 data were downloaded from the USGS Land Processes Distributed Active Archive Center (lpdaac.usgs.gov/products/mcd12q1v006/), and the Rapid Refresh (RAP) model data were obtained from the University Corporation for Atmospheric Research RAP archive (UCAR; soostrc.comet.ucar.edu/data/grib/rap/).
REFERENCES
Asefi-Najafabady, S., K. Knupp, J. R. Mecikalski, and R. M. Welch, 2012: Radar observations of mesoscale circulations induced by a small lake under varying synoptic-scale flows. J. Geophys. Res., 117, D01106, https://doi.org/10.1029/2011JD016194.
Avissar, R., and Y. Liu, 1996: Three-dimensional numerical study of shallow convective clouds and precipitation induced by land surface forcing. J. Geophys. Res., 101, 7499–7518, https://doi.org/10.1029/95JD03031.
Breiman, L., 2001: Random forests. Mach. Learn., 45, 5–32, https://doi.org/10.1023/A:1010933404324.
Brown, M. E., and D. L. Arnold, 1998: Land–surface–atmosphere interactions associated with deep convection in Illinois. Int. J. Climatol., 18, 1637–1653, https://doi.org/10.1002/(SICI)1097-0088(199812)18:15<1637::AID-JOC336>3.0.CO;2-U.
Brown, V. M., B. D. Keim, and A. W. Black, 2019: Climatology and trends in hourly precipitation for the southeast United States. J. Hydrometeor., 20, 1737–1755, https://doi.org/10.1175/JHM-D-19-0004.1.
Clark, P. J., and F. C. Evans, 1954: Distance to nearest neighbor as a measure of spatial relationships in populations. Ecology, 35, 445–453, https://doi.org/10.2307/1931034.
Collins, W., and P. Tissot, 2015: An artificial neural network model to predict thunderstorms within 400 km2 South Texas domains. Meteor. Appl., 22, 650–665, https://doi.org/10.1002/met.1499.
Coniglio, M. C., G. S. Romine, D. D. Turner, and R. D. Torn, 2019: Impacts of targeted AERI and Doppler lidar wind retrievals on short-term forecasts of the initiation and early evolution of thunderstorms. Mon. Wea. Rev., 147, 1149–1170, https://doi.org/10.1175/MWR-D-18-0351.1.
DOC/NOAA/National Weather Service, 2005: Gridded observed precipitation data. Accessed 9 September 2021, https://water.weather.gov/precip/metadata/DailyObservedPrecip.xml.
Findell, K. L., P. Gentine, B. R. Lintner, and C. Kerr, 2011: Probability of afternoon precipitation in eastern United States and Mexico enhanced by high evaporation. Nat. Geosci., 4, 434–439, https://doi.org/10.1038/ngeo1174.
Friedl, M., and D. Sulla-Menashe, 2019: MCD12Q1—MODIS/Terra+Aqua Land Cover Type Yearly L3 Global 500m SIN Grid V006. NASA EOSDIS Land Processes DAAC, accessed 15 April 2021, https://doi.org/10.5067/MODIS/MCD12Q1.006.
Gambill, L. D., and J. R. Mecikalski, 2011: A satellite-based summer convective cloud frequency analysis over the southeastern United States. J. Appl. Meteor. Climatol., 50, 1756–1769, https://doi.org/10.1175/2010JAMC2559.1.
Gibson, H. M., and T. H. Vonder Haar, 1990: Cloud and convection frequencies over the southeast United States as related to small-scale geographic features. Mon. Wea. Rev., 118, 2215–2227, https://doi.org/10.1175/1520-0493(1990)118<2215:CACFOT>2.0.CO;2.
Giovannettone, J. P., and A. P. Barros, 2008: A remote sensing survey of the role of landform on the organization of orographic precipitation in central and southern Mexico. J. Hydrometeor., 9, 1267–1283, https://doi.org/10.1175/2008JHM947.1.
Goggins, G. D., M. L. Grantham, S. W. Unger, K. B. Laws, K. L. Pence, and L. C. Dawson, 2010: Analysis of summer convection over central Alabama. 15th Symp. on Meteorological Observation and Instrumentation, Atlanta, GA., Amer. Meteor. Soc., 9.2, https://ams.confex.com/ams/90annual/techprogram/paper_161408.htm.
Haberlie, A. M., W. S. Ashley, and T. J. Pingel, 2015: The effect of urbanisation on the climatology of thunderstorm initiation. Quart. J. Roy. Meteor. Soc., 141, 663–675, https://doi.org/10.1002/qj.2499.
Hirt, M., S. Rasp, U. Blahak, and G. C. Craig, 2019: Stochastic parameterization of processes leading to convective initiation in kilometer-scale models. Mon. Wea. Rev., 147, 3917–3934, https://doi.org/10.1175/MWR-D-19-0060.1.
Hong, X., M. J. Leach, and S. Raman, 1995: A sensitivity study of convective cloud formation by vegetation forcing with different atmospheric conditions. J. Appl. Meteor., 34, 2008–2028, https://doi.org/10.1175/1520-0450(1995)034<2008:ASSOCC>2.0.CO;2.
Klitch, M. A., J. F. Weaver, F. P. Kelly, and T. H. Vonder Haar, 1985: Convective cloud climatologies constructed from satellite imagery. Mon. Wea. Rev., 113, 326–337, https://doi.org/10.1175/1520-0493(1985)113<0326:CCCCFS>2.0.CO;2.
Koch, S. E., and C. A. Ray, 1997: Mesoanalysis of summertime convergence zones in central and eastern North Carolina. Wea. Forecasting, 12, 56–77, https://doi.org/10.1175/1520-0434(1997)012<0056:MOSCZI>2.0.CO;2.
Kuo, J.-T., and H. D. Orville, 1973: A radar climatology of summertime convective clouds in the Black Hills. J. Appl. Meteor., 12, 359–368, https://doi.org/10.1175/1520-0450(1973)012<0359:ARCOSC>2.0.CO;2.
Lean, H. W., N. M. Roberts, P. A. Clark, and C. Morcrette, 2009: The surprising role of orography in the initiation of an isolated thunderstorm in southern England. Mon. Wea. Rev., 137, 3026–3046, https://doi.org/10.1175/2009MWR2743.1.
Lima, M. A., and J. W. Wilson, 2008: Convective storm initiation in a moist tropical environment. Mon. Wea. Rev., 136, 1847–1864, https://doi.org/10.1175/2007MWR2279.1.
Lin, Y., and K. E. Mitchell, 2005: The NCEP Stage II/IV hourly precipitation analyses: Development and applications. 19th Conf. on Hydrology, San Diego, CA, Amer. Meteor. Soc., 1.2, https://ams.confex.com/ams/Annual2005/techprogram/paper_83847.htm.
Markowski, P., and Y. Richardson, 2010: Mesoscale Meteorology in Midlatitudes. Wiley-Blackwell, 430 pp..
Mecikalski, J. R., and K. M. Bedka, 2006: Forecasting convective initiation by monitoring the evolution of moving cumulus in daytime GOES imagery. Mon. Wea. Rev., 134, 49–78, https://doi.org/10.1175/MWR3062.1.
Mecikalski, J. R., J. K. Williams, C. P. Jewett, D. Ahijevych, A. LeRoy, and J. R. Walker, 2015: Probabilistic 0–1-h convective initiation nowcasts that combine geostationary satellite observations and numerical weather prediction model data. J. Appl. Meteor. Climatol., 54, 1039–1059, https://doi.org/10.1175/JAMC-D-14-0129.1.
Mueller, C., T. Saxen, R. Roberts, J. Wilson, T. Betancourt, S. Dettling, N. Oien, and J. Yee, 2003: NCAR Auto-Nowcast System. Wea. Forecasting, 18, 545–561, https://doi.org/10.1175/1520-0434(2003)018<0545:NAS>2.0.CO;2.
Müllner, D., 2011: Modern hierarchical, agglomerative clustering algorithms. arXiv, 1109.2378v1, https://doi.org/10.48550/arXiv.1109.2378.
Nair, U. S., and Coauthors, 2008: Biogeography of tropical montane cloud forests. Part II: Mapping of orographic cloud immersion. J. Appl. Meteor. Climatol., 47, 2183–2197, https://doi.org/10.1175/2007JAMC1819.1.
Nisi, L., O. Martius, A. Hering, M. Kunz, and U. Germann, 2016: Spatial and temporal distribution of hailstorms in the Alpine region: A long-term, high resolution, radar-based analysis. Quart. J. Roy. Meteor. Soc., 142, 1590–1604, https://doi.org/10.1002/qj.2771.
Nisi, L., A. Hering, U. Germann, K. Schroeer, H. Barras, M. Kunz, and O. Martius, 2020: Hailstorms in the Alpine region: Diurnal cycle, 4D-characteristics, and the nowcasting potential of lightning properties. Quart. J. Roy. Meteor. Soc., 146, 4170–4194, https://doi.org/10.1002/qj.3897.
NOAA/NCEI, 2001a: U.S. Coastal Relief Model Vol. 4—Central Gulf of Mexico. Accessed 22 May 2020, http://www.ngdc.noaa.gov/mgg/coastal/crm.html.
NOAA/NCEI, 2001b: U.S. Coastal Relief Model Vol. 3—Florida and East Gulf of Mexico. Accessed 22 May 2020, http://www.ngdc.noaa.gov/mgg/coastal/crm.html.
Patou, M., J. Vidot, J. Riédi, G. Penide, and T. J. Garrett, 2018: Prediction of the onset of heavy rain using SEVIRI cloud observations. J. Appl. Meteor. Climatol., 57, 2343–2361, https://doi.org/10.1175/JAMC-D-17-0352.1.
Pedregosa, F., and Coauthors, 2011: Scikit-learn: Machine learning in Python. J. Mach. Learn. Res., 12, 2825–2830.
Renka, R. J., and A. K. Cline, 1984: A triangle-based C1 interpolation method. Rocky Mt. J. Math., 14, 223–237, https://doi.org/10.1216/RMJ-1984-14-1-223.
Rickenbach, T. M., R. Nieto‐Ferreira, C. Zarzar, and B. Nelson, 2015: A seasonal and diurnal climatology of precipitation organization in the southeastern United States. Quart. J. Roy. Meteor. Soc., 141, 1938–1956, https://doi.org/10.1002/qj.2500.
Roberts, R. D., and S. Rutledge, 2003: Nowcasting storm initiation and growth using GOES-8 and WSR-88D data. Wea. Forecasting, 18, 562–584, https://doi.org/10.1175/1520-0434(2003)018<0562:NSIAGU>2.0.CO;2.
Rotunno, R., J. B. Klemp, and M. L. Weisman, 1988: A theory for strong, long-lived squall lines. J. Atmos. Sci., 45, 463–485, https://doi.org/10.1175/1520-0469(1988)045<0463:ATFSLL>2.0.CO;2.
Schreiber, W. E., 1986: Case studies of thunderstorms initiated by radar-observed convergence lines. Mon. Wea. Rev., 114, 2256–2266, https://doi.org/10.1175/1520-0493(1986)114<2256:CSOTIB>2.0.CO;2.
Segal, M., and R. W. Arritt, 1992: Nonclassical mesoscale circulations caused by surface sensible heat-flux gradients. Bull. Amer. Meteor. Soc., 73, 1593–1604, https://doi.org/10.1175/1520-0477(1992)073<1593:NMCCBS>2.0.CO;2.
Tao, C., Y. Zhang, S. Tang, Q. Tang, H.-Y. Ma, S. Xie, and M. Zhang, 2019: Regional moisture budget and land-atmosphere coupling over the U.S. Southern Great Plains inferred from the ARM long‐term observations. J. Geophys. Res. Atmos., 124, 10 091–10 108, https://doi.org/10.1029/2019JD030585.
Trier, S. B., F. Chen, and K. W. Manning, 2004: A study of convection initiation in a mesoscale model using high-resolution land surface initial conditions. Mon. Wea. Rev., 132, 2954–2976, https://doi.org/10.1175/MWR2839.1.
Walker, J. R., J. R. Mecikalski, K. R. Knupp, and W. M. Mackenzie Jr., 2009: Development of a land surface heating index-based method to locate regions of potential mesoscale circulation formation. J. Geophys. Res., 114, D16112, https://doi.org/10.1029/2009JD011853.
Wasula, A. C., L. F. Bosart, and K. D. LaPenta, 2002: The influence of terrain on the severe weather distribution across interior eastern New York and western New England. Wea. Forecasting, 17, 1277–1289, https://doi.org/10.1175/1520-0434(2002)017%3C1277:TIOTOT%3E2.0.CO;2.
Weckwerth, T. M., L. J. Bennett, L. Jay Miller, J. Van Baelen, P. Di Girolamo, A. M. Blyth, and T. J. Hertneky, 2014: An observational and modeling study of the processes leading to deep, moist convection in complex terrain. Mon. Wea. Rev., 142, 2687–2708, https://doi.org/10.1175/MWR-D-13-00216.1.
Weisman, M. L., and J. B. Klemp, 1982: The dependence of numerically simulated convective storms on vertical wind shear and buoyancy. Mon. Wea. Rev., 110, 504–520, https://doi.org/10.1175/1520-0493(1982)110<0504:TDONSC>2.0.CO;2.
Wilson, J. W., 1966: Movement and predictability of radar echoes. National Severe Storms Laboratory Tech. Memo. ERTM-NSSL-28, 30 pp..
Wilson, J. W., and W. E. Schreiber, 1986: Initiation of convective storms at radar-observed boundary-layer convergence lines. Mon. Wea. Rev., 114, 2516–2536, https://doi.org/10.1175/1520-0493(1986)114<2516:IOCSAR>2.0.CO;2.
Yan, H., and R. A. Anthes, 1988: The effect of variations in surface moisture on mesoscale circulation. Mon. Wea. Rev., 116, 192–208, https://doi.org/10.1175/1520-0493(1988)116<0192:TEOVIS>2.0.CO;2.
Zhang, J., and Coauthors, 2011: National Mosaic and multi-sensor QPE (NMQ) system: Description, results, and future plans. Bull. Amer. Meteor. Soc., 92, 1321–1338, https://doi.org/10.1175/2011BAMS-D-11-00047.1.
