Convective rain cells: Radar-derived spatiotemporal characteristics and synoptic patterns over the eastern Mediterranean



[1] This paper examines the spatiotemporal characteristics of convective rain cells over the eastern Mediterranean (northern Israel) and their relationship to synoptic patterns. Information on rain cell features was extracted from high-resolution weather radar data. The radar-gauge adjustment, validation, cell segmentation and tracking techniques are discussed at length at the beginning of the paper. Convective rain cells were clustered into three synoptic types (two winter lows—deep Cyprus lows and shallow lows—and one tropical intrusion, Active Red Sea Trough) using several NCEP/NCAR parameters, and empirical distributions were computed for their spatial and temporal features. In the study region, it was found that the Active Red Sea Trough rain cells are larger, live for less time and possess lower rain intensities than the rain cells generated by the winter lows. The Cyprus low rain cells were found to be less intense and slightly larger on average than the shallow low rain cells. It was further discovered that the preferential orientation of the rain cells is associated with the direction and velocity of the wind. The effect of distance from the coastline was also examined. An increase in the number and area of the rain cells near the coastline was observed, presumably due to the sea breeze convection. The mean rainfall intensity was found to peak near the shore and decrease with distance inland. This information is of great importance for understanding rain patterns and can be further applied in exploring the hydrological responses of the basins in this region.

1. Introduction

[2] Space-time characteristics of convective rain cells are a key element in understanding rainfall-runoff processes and making hydrological predictions, as the catchment response is highly sensitive to these cell properties [Syed et al., 2003; Fiener and Auerswald, 2009]. To fully examine these characteristics, high-resolution rainfall data are necessary (e.g., 1 km2 and 15 min); however, dense networks of rain gauges supplying the necessary resolution are hard to find [Krajewski et al., 2003]. One possible solution to this dilemma may be found in using recorded rainfall data from remote sensing systems such as weather radar, which have been used for meteorological research from the end of the Second World War [Byers and Braham, 1949], and recently in several locations worldwide, such as southwest France [Féral et al., 2000], southern Israel [Karklinsky and Morin, 2006], northern Australia [May and Ballinger, 2007], northeast Italy [Capsoni et al., 2008] and northeast Spain [Barnolas et al., 2010]. Féral et al. [2000] distinguished between inland rain cells and cells over the Atlantic Ocean, fitting an elliptic geometrical shape depicting the major and minor axes and the orientation for each rain cell. Karklinsky and Morin [2006] conducted their research in a different climate area than Féral et al. [2000], one which ranged from Mediterranean to arid. They derived additional rain cell parameters and their distributions, such as central location, area, maximal rain intensity and spatial integral of rain intensity. May and Ballinger [2007] compared the height, size and lifetime of convective cells in a buildup or a break season, where convection is continental in nature, to monsoonal regime, where convection is oceanic in nature. Capsoni et al. [2008] examined the effect of different thresholds of rain intensity on rain cell life cycle and its statistical parameters. Barnolas et al. [2010] analyzed the geometry of convective structures associated with heavy rainfall using GIS in the western basin of the Mediterranean Sea.

[3] The catchment hydrological response depends on several parameters of convective rain cells such as their intensity, location, coverage area, speed and direction. The nature of this dependency can be investigated by inputting rain cells derived from weather radar into hydrological models computing runoff discharges or flash flood occurrences. Morin et al. [2006] demonstrated that a single intense rain cell, passing over the catchment near its tributaries and toward the outlet during a rain event, may trigger a high flash flood in the catchment. Yakir and Morin [2011] also considered the cell's velocity to be an important factor affecting the time the cell stays over the catchment and hence the amount of rain falling over it and the generated flood magnitude. They demonstrated that the catchment response is very sensitive to the convective cell's starting location, velocity and direction, and that small changes in these features can double and triple the peak discharge. This sensitivity was similarly demonstrated in a paper by Smith et al. [2000]. Rozalis et al. [2010] applied a hydrological model for flash-flood prediction and found that intrastorm rain intensity distributions significantly affect the flow magnitude, as exemplified by the different hydrological responses simulated for two storms with similar amounts of rain but different rain intensity distributions over time. Zoccatelli et al. [2010] found that a considerable degradation of the modeled runoff is expected for small catchments (less than 160 km2) when the intrastorm rainfall variability is neglected.

[4] The main objective of the current study is to characterize the spatial and temporal diversities of convective rain cells derived from weather radar data for different synoptic systems along the northwestern Israel coastline at the eastern Mediterranean (EM).

2. Study Area

[5] The study area (102 × 73 km2, centered over coordinates 34.7E 32.5N) was located on the northwestern Israeli coastline of the Mediterranean Sea (Figure 1). Approximately 64% of the study area was offshore and 36% was inland. The Dalya and Taninim catchments, located in the center of the study area, drain toward the Mediterranean Sea. The terrain near the coastline is mostly flat, ascending moderately eastward to 490 m above sea level at the mountain ridge's highest point.

Figure 1.

Map indicating the radar system's location, the study area and the radar mesh over it (visible radar mesh), the Dalya and Taninim Basins (bright polygons) and the validation rain gauges (triangles, numbered 1–13).

[6] The study area has a Mediterranean climate; its rainy season lasts from October to May, while June to September is dry and hot. The annual mean rainfall is 550 mm near the coastline, decreasing inland with distance from the sea. The major winter rainy synoptic system in the region, called the Cyprus low, is an extratropical cyclone that passes over the EM [Goldreich, 1998] and contributes on average 83% of the total rainy days [Saaroni et al., 2010]. The Cyprus lows and the other shallow lows (known together as the “winter lows”) can be divided into several classes depending on the location of their centers and their depths [Alpert et al., 2004b; Osetinsky, 2006]. The remaining rain events are associated with the Active Red Sea Trough (RST) synoptic system, which is more common during the transition seasons [Goldreich et al., 2004]. The RST is defined as a surface trough extending from eastern Africa along the Red Sea toward the Middle East, and is responsible for most of the flash flood events in southern Israel [Kahana et al., 2002].

3. Radar Rainfall Estimation

3.1. Radar and Rain Gauge Data

[7] Data from the Shacham (EMS) Mekorot company weather radar system located at Ben Gurion Airport, 30 km east of Tel Aviv and 25–120 km south of the study area (Figure 1), were used in this study. Data from this radar have been used extensively for climatology and hydrology studies over the last decade [see Morin et al., 2001, 2009; Goldreich et al., 2004; Karklinsky and Morin, 2006; Morin and Gabella, 2007; Rozalis et al., 2010; Yakir and Morin, 2011]. The radar is a C-band (5.35-cm wavelength), non-Doppler system with a maximal transmitting power of 250 kW, a temporal resolution of about 5 min per volume scan, and a spatial polar resolution of 1.4° × 1 km in space (Figure 1). The radar typically scans at about 13 elevation angles; because radar beams at the elevations lower than 2 degrees are partially blocked or contaminated by ground clutters, only data from the first available elevation angle above 2 degrees (mean elevation of 2834 m with a standard deviation of 743 m above sea level over the study area; elevation changes from 873 m to 4190 m) were used for the analysis. The examination of reflectivity values for the entire period of interest revealed pixels with substantial ground clutters or beam blockage which were manually excluded from the analysis.

[8] Radar data for twelve hydrological years (1991/1992–1997/1998, 1999/2000–2002/2003 and 2004/2005) were analyzed in this study, with a total of 191,586 radar volume scans. Data from two hydrological years (1998/1999 and 2003/2004) were excluded from the study due to substandard quality (see validation section below). Daily rainfall data derived from 26 rain gauges were used for the radar-gauge adjustment (all within 100-km distance from the radar), and another 13 rain gauges in the surrounding of the Dalya and Taninim basins (see locations in Figure 1) were used in the validation process. All rain gauges are operated by the Israel Meteorology Service (IMS).

[9] The selected hydrological years possess a high diversity of precipitation amounts: the year 1991/1992 was the wettest year recorded in Israel; three hydrological years (1994/1995, 2001/2002 and 2002/2003) are considered wet years with an annual rainfall above average, while three other hydrological years (1993/1994, 2000/2001 and 2004/2005) are considered dry years; and the remainder of the hydrological years contained roughly average rainfall (Table 1).

Table 1. Rain Gauges and Measured Annual Rain Depth (mm) in the Study Area
NameEn KarmelOferRamat Ha ShofetRamot MenasheMayan ZeviAmiqamEven YizhaqRegavimKefar GliksonBinyaminaBarqayGan ShemuelHa ZoreaAverage
Station number12345678910111213 

3.2. Radar-Gauge Adjustment

[10] Rainfall intensity data (R, mm h−1) were initially calculated using weather radar reflectivity data (Z, mm6 m−3) by a fixed Z-R power law relationship, and then readjusted using the weighted regression (WR) method as suggested by Gabella et al. [2001]. The Z-R relationship of Z = 316R1.5 was applied for this study, following a previous study that used the same relation in our region [Morin and Gabella, 2007]. A lower threshold of 10 dBZ (dBZ = 10*Log10Z) for noise filtering and an upper threshold of 250 mm h−1 (61 dBZ) to reduce unrealistic strong returns from hail particles (similar values presented by Morin and Gabella [2007] and Gabella et al. [2011]) were set. Accumulated annual rain depth in rain gauges and in radar pixels above those gauges were computed for each hydrological year and the WR coefficients were derived, as detailed in Morin and Gabella [2007]:

display math

where Fdb is the correction factor, which is ten times the logarithmic radar-to-gauge depth ratio (P and G, respectively), a0 and aD are the coefficients to be determined by simple linear regression, and logD is the logarithmic distance from the radar site in reference to a 60-km distance:

display math

where d is the distance from the radar in kilometers.

[11] Rain gauges, for which a substantially low fit obtained from equation (1) was found for a given year, were removed from the analysis. The WR-derived coefficients for the 12 years of data were (average ± standard deviation):

display math

[12] Table 2 summarizes the WR coefficients, the R2 and the number of rain gauges used for each hydrological year.

Table 2. WR Coefficients (a0 and aD), R2 and the Number of Rain Gauges Used for Radar-Gauge Adjustment
Hydrological Yeara0aDR2Number of Rain Gauges

3.3. Validation

[13] Yearly rain depths from the 13 validation rain gauges were compared to the yearly rain depths derived from the radar for each hydrological year. Each rain gauge was classified as: (1) in close agreement (when the absolute difference between the cumulative radar annual rain depth and the gauge values was less than 100 mm); (2) overestimated; or (3) underestimated (when the radar annual rain depth was more or less than the gauge data by 100 mm, respectively). In addition, fractional standard error (FSE) function was applied to assess the quality of the quantitative precipitation estimates, as presented by Vignal et al. [2000]:

display math

where N is the number of rain gauges examined and P and G are the annual rain depths derived from the gauges and from the radar, respectively. The FSE scores, presented in Table 3, were found to be high (>0.35) in two hydrological years— 1998/1999 (score of 0.37) and 2003/2004 (0.44)—and thus were excluded from analysis.

Table 3. Validation Results: Rain Gauge Status (Compared to the Annual Precipitation Depth Derived From the Radar), Along With Their Corresponding FSE Score
Hydrological YearNumber of Gauges: OverestimatedNumber of Gauges: in RangeNumber of Gauges: UnderestimatedFSEValid

4. Rain Cell Identification

[14] Rain cell identification is a procedure in which the convective rain cells are spatially determined (using the segmentation method) and temporally analyzed (with a cell tracking algorithm). This procedure is discussed at length in the next two subsections.

4.1. Segmentation and Feature Extraction

[15] In the segmentation process each radar image (Cartesian map with 0.5 × 0.5 km2 pixel size transformed from the original radar polar data) is partitioned into rain segments representing rain cells, using the procedure described below. Pixels are set to null value if one of the following is found to be true: (1) pixels have rain intensity lower than a predefined lower threshold (10 mm h−1); (2) five or more pixels in the 3-by-3 pixel neighborhood are null; or (3) the pixels are spurs.

[16] The rain intensity threshold, corresponding to about 40 dBZ, guarantees that only the convective part of the rain will remain for the rain cell derivation. This threshold was previously used by Rigo and Llasat [2004], Kyznarová and Novak [2009], Barnolas et al. [2010] among others for the same purposes. A lower threshold of 1 mm hr−1 was also applied to check the sensitivity of the results against this assumption (not presented) and similar conclusions were obtained.

[17] The continuous clusters of pixels remaining in the image are then segmented and labeled. Rain segments with an area lower than a predefined area threshold (9 km2) are removed in order to avoid small cells that are potentially noise. Most of the segments represent individual convective rain cells, but some of the large segments signify convective structures [Rigo and Llasat, 2004; Barnolas et al., 2010]; in the current study we refer to all segments as convective rain cells. An example of the segmentation process is presented in the upper part of Figure 2a, where rain cell segments are delimited by black lines.

Figure 2.

(a) (top) Three consecutive radar images presenting the rain intensities and rain cells' segmentation. (bottom) The cell tracking results are presented: (left) each cell receives a unique ID (and color) and an ellipse is assigned (red ellipse); those ellipses are projected to the next time step according to the computed motion vector (dashed ellipse). (middle) A new cell is born (light blue), one cell is tracked from the previous image (orange) and two cells (blue and green from the first image) are merged into one cell (green, in the middle image). (right) The two upper cells are tracked (green and yellow) while the blue cell from the previous time step is terminated and does not appear in this image. (b) The ellipse is fitted to the rain cell after segmentation and major and minor axes are computed. (c) Two cells from time step t (left) are merged to become one cell at time step t + 1 (right), following the example given at the text. (d) Illustration of the ‘track’, ‘merge’ and ‘split’. Red ellipse marks the rain cells true position at time t (left to the arrow) and at time t + 1 (right); dashed ellipse marks the projected location of the cells from time t to time t + 1.

[18] Our method of segmentation differs from the one introduced by Karklinsky and Morin [2006] and applied in the papers of Morin et al. [2006] and Yakir and Morin [2011]. They have suggested finding the local maxima for each rain cell and then expanding the region by adding neighbor pixels until the lower threshold is met. This way only one peak is possible for each segment. The method presented here, in contrast, allows several peaks for a single rain cell. Other methods of segmentation and cell identification are presented in the literature [e.g., Dixon and Wiener, 1993; Johnson et al., 1998; Féral et al., 2000; von Hardenberg et al., 2003; Bonelli and Marcacci, 2008].

[19] An ellipsoid is fitted for each segment and major and minor axes are found (Figure 2b). The major axis is defined as the longest direction from the segment center, and the minor axis is set perpendicular to it. The rain cells are then characterized using the following parameters: ellipsoid central location, segment area (km2), major and minor radius lengths (km), ellipticity (ratio of minor-to-major radius lengths), cell orientation (angle of the major radius in degrees, relative to the west-east direction; positive counterclockwise) and maximal and mean rain intensities (mm h−1) over the segment.

4.2. Cell Tracking

[20] Cell tracking algorithms allow monitoring of rain cells progress and the prediction of their future position. These algorithms use various sources of information such as radar, satellite [Sieglaff et al., 2011] or lightning [Kohn et al., 2011] data and they have been integral tools of nowcasting methodologies for many years (for example, the SCIT algorithm by Johnson et al. [1998] or in Dixon and Wiener [1993]). The convective cell tracking algorithm presented here temporally links between rain cells in consecutive images (two successive images with a time gap not exceeding ten minutes) based on the conceptual model CELLTRACK developed by Kyznarová and Novak [2009], with modifications made for the current study.

[21] Motion vector, composed of velocity and direction components, is determined for each sequential image as follows:

(1) The radar image at time t is shifted 3 degrees clockwise from the north with a velocity of 3 m s−1;

(2) The correlation between the shifted radar images at time t and the radar image at time t + 1 is computed using Pearson's correlation;

(3) Step 1 is repeated for different directions (increments of 3 degrees) and velocities (increments of 3 m s−1, up to a velocity of 36 m s−1);

(4) The best correlation coefficient for time t is recorded, along with the corresponding velocity and direction components. The motion vector thus describes the average rain cell's progress between two consecutive images.

The life cycles of the rain cells are divided into five categories:

(1) Birth—a rain cell was created at time t;

(2) Track—a rain cell at time t was detected along the motion vector at time t + 1;

(3) Split—a rain cell at time t was split into at least two cells detected at time t + 1;

(4) Merge—at least two rain cells at time t were merged into one cell at time t + 1; and

(5) Death—a rain cell at time t was not detected at time t + 1.

[22] Figure 2d illustrates the “track,” “merge,” and “split” categories. The correlation between rain cells at time t and rain cells at time t + 1 and the selection of a category is determined by projecting rain cells at time t according to the motion vector assuming different scenarios and computing overlapping areas between the ellipsoids of the projected cells and the ellipsoids of the cells at time t + 1 (s-score):

display math

where S is the ellipsoid area, it is a rain cell (or rain cells) at time t, * marks its projection into the next time step, jt+1 is a rain cell (or rain cells) at time t + 1 and ∩ marks the intersect of the two. Different cell tracking categories and cell matching combinations result in different s-scores and the one with the highest value is selected. In the original algorithm developed by Kyznarová and Novak [2009] another variable is used in equation (5) representing the size of the area not shared by either the observed rain cell at time t or by the projected rain cell from time t − 1. This variable was excluded from the current study in order to expedite calculations.

[23] Example for the cell tracking procedure: if the upper cell in Figure 2c at time t (solid ellipse in the left side of Figure 2c) has an ellipsoid area of 40 km2 and the lower cell has an ellipsoid area of 50 km2, the cell detected at time t + 1 (solid ellipse in the right side of Figure 2c) has an ellipsoid area of 150 km2, and the overlapping ellipsoid area of the lower and upper cells projected in time t + 1 using the motion vector (dashed ellipses in the right side of Figure 2c) and the cell at time t + 1 is 85 km2, then: the s-score for “merging” will be inline image. Alternatively, the s-score for the “track” category (meaning, one of the two cells must “die” and the other is “tracked”) will be 0.27 for the upper cell and 0.29 for the lower one. Since the merging scenario provides the highest s-score among all other scenarios it will be selected in the cell tracking algorithm. Three radar images are presented in Figure 2a along with the matching cell tracking results and illustration of the “track,” “merge,” and “split” categories are presented in Figure 2d.

[24] Evaluation of the automatic cell tracking procedure was performed for the storm of 6 December 2001. This storm (consists of three rain events, see next section) lasted for just under twelve hours (00:03 to 12:00), with 126 radar images recorded, 857 individual rain cells detected and 193 tracks derived with the cell tracking algorithm. Data derived from the auto-tracking algorithm were compared to data extracted by a manual tracking procedure. The automatic algorithm overestimated the “birth” and “death” classes (+11.6% and +9.9%, respectively), with underestimation reported for the “split” and “merge” classes (−14.2% and −6.3%). The underestimation of the “split” class explains the overestimation of the “birth” class, and the underestimation of the “merge” class explains the overestimation of the “death” class. The algorithm estimated the “track” class well, with a minor difference of 1.1% detected. In addition, the auto-tracking procedure was evaluated against a manually tracking procedure for three storms and its Probability of Detection (POD) and False Alarm Rate (FAR) are presented in Table 4 (as in Dixon and Wiener [1993]). There is a good fit between the manually tracking and the auto-tracking algorithms, as indicated by the average POD of 0.88 and FAR of 0.15.

Table 4. Probability of Detection (POD) and False Alarm Rate (FAR) for Three Storms
Date (dd/mm/yyyy)Synoptic TypeCorrect AssignmentsMissed AssignmentsIncorrect AssignmentsPODFAR
6/12/2001 00:03–12:00Shallow and deep lows34443470.890.12
6/02/1992 14:19–19:19Deep Cyprus low799120.900.13
14/03/1996 11:59–17:11Red Sea Trough6013230.820.28
Total 48365820.880.15

5. Synoptic Classification

5.1. Cluster Analysis Classification

[25] The 191,586 radar images derived from the 12 years of data were collated for a total 1,113 rain events. A rain event begins when rain cells first appear on the radar image and ends when there is an intermission of more than one hour before the next rain cell appears, meaning that a storm lasting several hours or even days can potentially be divided into several rain events. A hierarchical agglomerative cluster analysis (CA) technique using Ward criterion was applied in order to classify the rain events into several synoptic types. This method was chosen as it is one of the most commonly employed and it yielded the best results in previous climatologically studies [Unal et al., 2003; Vrac et al., 2007]. The classification was performed using four variables obtained for the time of each rain event from the six-hour NCEP/NCAR reanalysis [Kalnay et al., 1996]: sea level pressure (SLP), specific humidity at 700 hPa (SHUM700), geopotential height at 500 hPa (HGT500) and zonal wind at 850 hPa (UWND850). The data for these variables were extracted for a location near the study area (35E 32.5N). Additional parameters were tested, including meridional wind component, lifted index, convective precipitation rate and temperature at different levels. However, all of these parameters were omitted as no better CA results were obtained.

[26] 882 rain events (80% of the total number) lasting less than six hours each (to match the NCEP/NCAR temporal resolution) were chosen for the hierarchical CA process. The CA results were tested using linear Discriminant Analysis (DA) method in order to determine the optimal number of clusters. The same NCEP/NCAR data used for the CA were chosen as covariates and the results of the CA analysis were used to categorize the covariates into groups. In a perfect match the DA should have given the same results as the CA. Best DA results were found for CA of three clusters: the first cluster contained 606 rain events (DA predicted 525 of them, or had an 87% success rate), the second cluster had 222 rain events (with 212 predicted, or a 95% success rate) and the third cluster included 54 rain events (all predicted by the DA). The events in the first class were characterized by higher SLP than the other groups, while in the second cluster the events had a strong 850 hPa u-wind component and a low geopotential height at 500 hPa. In the third cluster the events had a high specific humidity component at 700 hPa and a high 500 hPa geopotential height (see comparison in Figure 3 and Table 5).

Figure 3.

NCEP/NCAR type-classification variables, boundaries delimited one standard deviation around the mean (dashed line) for each parameter.

Table 5. Mean and Standard Deviation (in Parentheses) for the NCEP/NCAR Type-Classification Variables
Synoptic TypeNumber of EventsUWND850 (m s−1)SLP (hPa)SHUM700 (kg kg−1)HGT500 (m)
(1) Shallow lows6066 (4)1016 (3)0.002 (0.0008)5597 (71)
(2) Cyprus lows22212 (4.3)1010 (3)0.002 (0.0007)5521 (70)
(3) RSTs541.5 (4.3)1011 (3)0.005 (0.0007)5705 (81)

5.2. Synoptic Types

[27] Synoptic maps, covering the Mediterranean, were plotted for each class described above. The rain-bearing Mediterranean cyclones, known as the “winter lows,” are the most significant synoptic systems during wintertime in Israel [Alpert et al., 2004a]. The first two types belonged to this system: Type 1 was synoptically characterized as a shallow low in light of both the sea level pressure and the 500 hPa geopotential height (Figure 4a). The center of the low was located to the north of Israel, bringing moderate westerly winds into the EM coastline. Type 2 was synoptically described as a classical Cyprus low, a deep cold low centered above Cyprus supported by a deep upper level trough (Figure 4b). The Cyprus low system triggers stronger westerly winds near the EM coastline than do the shallow lows. This second-type CA product matches the one presented by Ziv et al. [2006] for Cyprus low systems affecting the EM (see their Figure 2).

Figure 4.

Synoptic maps representing: (a) shallow low over the North of Israel (type 1); (b) Cyprus low (deep low over Cyprus, type 2); and (c) the Active Red Sea Trough (type 3). Contours represent sea level pressure and colored zones stand for 500 hPa geopotential height.

[28] Type 3 was linked to the Active Red Sea Trough (RST), which occurs mainly during the autumn [Alpert et al., 2004a] and spring. The RST is defined synoptically as a low sea level pressure trough extending from eastern Africa along the Red Sea toward the EM [Ashbel, 1938]. The sea level pressure and the 500 hPa geopotential height that derived from the CA procedure (Figure 4c) resemble the RST compositions made by Kahana et al. [2002] (see specifically their Figures 2 and 3). The wind pattern in our scenario changed spatially over Israel, with southwestern winds in the south of Israel becoming southeastern winds in the north of Israel along the EM coastline.

6. Spatiotemporal Characteristics of Convective Rain Cells

6.1. Empirical Distributions of the Rain Cells

[29] Empirical distributions of spatial and temporal convective rain cell characteristics are presented in Figures 5, 6 and 7. A total of 42,582 convective rain cells (5,740 tracks) were analyzed: 27,293 cells (3,767 tracks) of them relate to the shallow low synoptic type; 12,125 cells (1,638 tracks) count for the Cyprus low type; and 3,164 cells (335 tracks) relate to the RST. The analysis of the spatial rain cell characteristics (Figures 5 and 6c) was done for each individual rain cell regardless its life cycle status. Whereas, the cells life time frequency analysis (Figures 6a and 6b) was calculated from complex tracks (i.e., full life cycle, including splits and merges). In this case, if a cell splits, then the two ‘new’ cells are continue to be referred to by their original identification; if cells merge the life time of all but one end at that point, while the newly created merged cell continues its life cycle. The most significant findings were:

Figure 5.

Empirical distribution of characteristics of the synoptic types of convective rain cells: (a) major radius; (b) minor radius; (c) area; (d) orientation (along the east-west axis); (e) area's mean rain intensity; and (f) max rain intensity. Median (M), mean (μ) and maximum values for some of the figures which were truncated (max) are presented.

Figure 6.

(a) Track-based empirical distribution for each synoptic type of rain cell's life span with median (M), mean (μ) and maximum values (max). (b) Correlation between the accumulated mean rain depth and the cell's life span. (c) Empirical distribution of the rain cell's ellipticity. (d) Frequencies of “track,” “merge” and “split” tracking categories relatively to the total number of rain cells for each synoptic type.

Figure 7.

Rose diagram presenting the distribution of the motion vector (velocity and direction) for each synoptic type. The mean and standard deviation of the distributions and the number of samples used are summarized in Table 6.

[30] 1. The convective rain cells preserve the same shape for the three synoptic types, where the major radius is on average twice as long as the minor radius (Figures 5a and 5b), and the distributions of the ellipticity factor are similar for the three types with an identical mean factor of 0.53 (Figure 6c).

[31] 2. The RST rain cell areas are larger on average than the areas of the winter lows (about 40 km2 to 64 km2, Figure 5c), but the RST rain intensities (both maximum and mean distributions, Figure 5e and Figure 5f) are lower than the rain intensities of the winter lows. RST events are characterized by medium-level clouds [Goldreich et al., 2004] originating in the Red Sea (about 300 km south of the study area) and therefore their intensities tend to decrease during their trajectory over land (thus the empirical distributions presented here are relevant only for the northern part of Israel). In contrast, the clouds associated with the winter lows tend to be low-level clouds (and therefore smaller than the RST clouds) with much shorter paths over land.

[32] 3. The rain cells' orientations (Figure 5d) were different for the winter lows and the RST: while the preferential orientation of the rain cells connected with winter lows is mainly west-east to southwest-northeast (with up to 30 degrees tilt range), the RST cells have two significant preferential orientations—one west-east to southwest-northeast and the other toward the north-south axis.

[33] 4. For each synoptic type, the two-component motion vector time series distributions (see cell tracking subsection) are presented in Figure 7 and Table 6. The shallow low events are characterized by relatively low velocities (mean velocity of 10.3 m s−1) and directions that were primarily western with minor southwest deviations, while the Cyprus low events are composed of higher velocities (mean of 15 m s−1) and a more apparent southwest direction (motion vector data were smoothed out prior to distribution computation).

Table 6. Mean and Standard Deviation (in Parentheses) for the Direction and Velocity Components of the Motion Vectors
Synoptic TypeSample Size (N)Velocity (m s−1)Direction (deg)
(1) Shallow lows1644810.3 (5)279.8 (42.4)
(2) Cyprus lows659915 (5.1)260.9 (32)
(3) RSTs146914.3 (7.2)275.4 (92.9)

[34] 5. Similar to the orientation distributions, the RST events comprise two significant preferential vectors—the main one has a strong (over 30 m s−1) southwesterly component and the smaller one is characterized by a moderate south-southwest composition (with velocities not exceeding 30 m s−1).

[35] 6. Regardless of their synoptic type, the derived cell life time distributions (Figure 6a) indicate that most of the convective rain cells “lived” for only five minutes, meaning that they are presented in only one radar image. The Cyprus low rain cells lasted longer on average than did the RST rain cells, and the durations observed for the shallow lows were even longer than the durations observed for the Cyprus lows. As expected, good correlations were found between the cells' life spans and the cells' accumulated mean rain depth (Figure 6b)—the longer the rain cells lasted, the more rainfall accumulated per cell. It should be noted that the cell life time parameter (Figure 6a) might be biased toward lower values resulting from the incomplete cell tracks captured when cells enter or exit the domain and from the overestimation of “birth” and “death” classes and underestimation of “split” and “merge” classes (see section 4.2). In addition, we emphasize that the computed cell life time parameter depends on the rain intensity and cell area thresholds set for the rain cell identification procedure (see section 4.1 above).

[36] 7. The “Track” category occurs substantially more often (80% on average) than the “merge” (14.9%) and “split” (5.1%) categories for the three synoptic types (Figure 6d). Differences were found between the frequencies of the “merge” and “split” categories for the different synoptic types: the “merge” category varies between 11% for the shallow low to 19.4% for the RST; and the “split” category varies between 3.5% for the shallow low to 7.5% for the RST.

[37] The means of all the above parameters for the three synoptic types were compared statistically using the nonparametric Kruskal-Wallis test [Kruskal and Wallis, 1952]. For all cases a p-value smaller than 0.0001 was found, indicating a very strong significant difference.

6.2. Effect of Distance From the Coastline

[38] Rain cells generally moved westward in the study area, crossing the Mediterranean coastline on the way (Figure 1). The means of several newly formed rain cells' parameters were analyzed in order to examine how the distance from the shoreline affected the rain cells. For this, the study area was divided into 2 × 2 km2 pixels and the shortest distance to the shoreline was calculated from their centers. We then analyzed the characteristics of the new cells whose centers lay within each of the pixels. The findings from the analysis were: The average number of new cells offshore increased as the distance to the land decreased (Figure 8a). The number of new cells peaked near the coastline and decreased considerably about eight kilometers inland. The orientation of rain cells moving away from sea toward the land (from 40 to 55 degrees from the east-west axis) changed counterclockwise (Figure 8b), and the trend was inverse as the orientation changed clockwise with increasing distance from the shoreline inland (from 55 to 40 degrees offset). The mean area of new cells (Figure 8c) slightly decreased from 60 km from the shoreline toward the land (from 24 km2 to 20 km2), with a major leap detected at the coastline (from 20 km2 to 25 km2). A positive linear trend (R2 = 0.79) was detected for the mean maximum rain intensity (Figure 8d) for the distance of 60 to about 30 km offshore (22 to 25 mm h−1). From 30 km offshore to the coastline the mean maximum rain intensity stayed at approximately 25 mm h−1, and a sharp negative trend appeared from the coastline inland, reducing the mean rain intensity to 20 mm h−1 20 km from shore. In order to verify the above trends represent a real change in cell characteristics rather than the effect of distance from the radar (that is partially correlated with distance from shoreline) a similar analysis was conducted for a sub sample of the data taken from a narrow range of radar distances. The same trends were found for the sub-sample although they were noisier due to the small sample size.

Figure 8.

The effect of the distance from coastline on newly formed rain cells, in terms of: (a) average number of cells; (b) mean orientation; (c) mean area; and (d) mean maximum rainfall intensity.

7. Summary and Discussion

[39] Convective rain cells derived from weather radar data were analyzed for several key parameters, such as cell area, rainfall intensities, cell orientation and ellipticity. Cell tracking algorithms enabled the computing of some temporal characteristics, including cell life span and cell life cycle. The rain cells were also classified into three synoptic types according to atmospheric variables.

[40] The statistical analysis of convective storms might be affected by the study area size (102 × 73 km2) because of the border effect, i.e., some of the convective rain cells exit or enter the study area, thus their life cycle is not completely recorded. Overestimation of the “birth” and “death” classes and underestimation of the “split” and “merge” classes presented in section 4.2 are an indication for the incomplete life cycle record. However, the large sample of rain cells which had been tracked for more than one volume scan (16,596 cells) assures the reliability of the cells` parameter distributions. It also should be noted that the size of the current study area lies within the range of sizes presented in other studies, with larger areas analyzed by Bonelli and Marcacci [2008] (400 × 400 km2 in northern Italy) and by May and Ballinger [2007] (256 × 256 km2 in northern Australia) while smaller study area sizes have been analyzed by Capsoni et al. [2008] (a circular area with a 40 km radius) and by Karklinsky and Morin [2006] (80 × 70 km2 and 150 × 65 km2 in southern Israel). The analysis presented here well represents the coast area of northern Israel, extending eastward up to the foothills of the mountainous region (Figure 1).

[41] Synoptic classification was computed using CA with four NCEP/NCAR parameters: SLP, SHUM700, HGT500 and UWND850 (see section 4.1). These parameters were selected due to the fact that they represent the synoptic pattern in the study region well, especially the SLP and HGT500 (for examples of classification using those variables, see Zangvil et al. [2003] and Ziv et al. [2006]). The auto-classification conducted in this study using the CA method yielded three synoptic types: the shallow lows, the Cyprus lows and the RST; all were previously defined by researchers based on subjective observation of the synoptic parameters [Kahana et al., 2002; Alpert et al., 2004b; Osetinsky, 2006; Ziv et al., 2006] using the NCEP/NCAR reanalysis data. It has to be emphasized that in spite of the fact that the 12 years of data analyzed are composed of wet, average and dry years (Table 1 and section 3.1), the climatologically representation may not be perfect and, accordingly, the CA results are record-dependent.

[42] Elliptical shapes seem to depict the shape of convective rain cells well, as indicated from the mean ellipticity factor of 0.53 yielded in this study and from the results of others: Yakir and Morin [2011] reported an average ellipticity factor of 0.58 and Karklinsky and Morin [2006] reported a mean ellipticity of 0.63 (both studies were conducted in Israel); Barnolas et al. [2010] found that the mean major and minor axis values in Catalonia are 7.79 km and 3.66 km (respectively) associated with an ellipticity factor of 0.47; and Féral et al. [2000] reported a mean factor of 0.57 for both inland and oversea rain cells in France. It has to be noted that part of the variance discussed above is a result of the use of different segmentation methods for different studies. The preferential orientation of the elliptic cells is correlated with the motion vector direction: a west-east (with 10 degree average deflection toward the north) major-axis orientation was found for the three synoptic types corresponding to southwesterly motion vector direction, while an additional favored north-south orientation was found for the RST class with a corresponding southerly direction component. The elliptical shape, orientation and progress velocity of the small convective rain cells are therefore determined by the wind direction and velocity (interpreted here as the motion vector) which is controlled by the central location and depth of the low pressure system over the EM.

[43] The occurrence of the different synoptic types during the winter and their associated rainfall amounts have been investigated in the past [Alpert et al., 2004a, 2004b; Saaroni et al., 2010]. The current CA process determines that the Cyprus lows contribute 25% of the winter rain events, the shallow lows generate 69% and the RST account for the rest (only 6%). The convective rain cells generating these storms have been analyzed for their intensity, area and temporal distribution as the catchment's hydrology response is known to be sensitive to these parameters [Morin et al., 2006; Rozalis et al., 2010; Yakir and Morin, 2011]. The Cyprus low and the shallow low rain cells are similar in general, but the Cyprus low rain cells were found to be less intense and slightly larger than the shallow low rain cells; this phenomenon was discussed by Saaroni et al. [2010], who claimed that the deep lows were found to be more productive in the mountain regions (east of our study area) than along the coastline due to stronger wind associated with the deeper cyclones as also observed in the current study (see Figure 7 and Table 6). The shallow low rain cells tend to live longer than the Cyprus low rain cells (14 min versus 12.3 min) and to move slower.

[44] The RST rain cells act differently: their area is much larger—almost double the shallow lows on average; their rain intensities are lower, but with longer tail toward the extreme (Figure 5f); the rain cells' life spans are shorter (9.2 min); and their velocities can reach up to 40 m s−1 along the southwest-northeast axis.

[45] It seems that the shallow low rain cells possess the highest potential for generating flash floods in both the Dalya and Taninim basins as they are the most intense, they “live” longer and they move slower, and thus they potentially spend more time inside the catchment area than the rain cells originating from Cyprus lows and RST. The RST, the main cause for flash floods in the arid to semi-arid climate catchments in southern Israel [Kahana et al., 2002; Ziv et al., 2005], very rarely affects the hydrological system in the Mediterranean climate of central and northern Israel (e.g., the April 2006 rainstorm over Wadi Ara with a recurrence interval of more than 100 years [see Morin et al., 2007]).

[46] The effect of distance from the coastline for the newly formed cells was examined. The average number of new cells increased from 60 km offshore to the coastline where the number of new cells peaked, and from the coastline toward the land a descending trend in the number of new cells was detected. The increase in convective rain cells near the shore is presumably the result of a sea breeze [Miller et al., 2003] from the Mediterranean toward the land, with a possible contribution from the coastal fronts as described by Rosenfeld and Nirel [1996] and Goldreich et al. [2004]. Therefore we can assume that the sea breeze convection is also responsible for the rain cell area increase detected near shore. The mean cell orientation changed counterclockwise when the cells were offshore and then clockwise from the coastline toward the land. This is in contrast to the constant counterclockwise trend found by Karklinsky and Morin [2006] for the southern part of the Israeli coastline. The highest mean maximum rain intensities were observed near the shore, with an intense decrease reported with distance inland. This phenomenon of decrease in precipitation as a function of the distance from the Mediterranean inland is well documented [Sharon and Kutiel, 1986; Goldreich, 1994], and seems to be moderated by the orographic effect 16 km inland from the coastline. Since the basins in the study region were generally oriented from east (upstream) to west (downstream) these observations imply that rain cells with higher rain intensities pass over the outlets of the basins, while the upstream section of the catchment, which is farther away from the seashore, experiences fewer rain cells with lower rain intensities. Further investigation utilizing hydrological modeling is required in order to assess the hydrological consequences of these gradients over the studied catchments; such research must also take surface and meteorological conditions (such as geology, land use, evaporation, soil wetness capacity, etc.) into account.

[47] Information on space-time characteristics of convective rain cells can be combined to form weather generators which will be used for downscaling methods of General Circulation Models (GCM) or Regional Circulation Models (RCM) that typically have too coarse a resolution for hydrological prediction. Wilby et al. [1998], Wilks and Wilby [1999] and Wilks [2010] gave examples of downscaling methods from dynamical downscaling (using a mesoscale grid within the GCM) through empirical downscaling (using regression methods from large-scale GCM to local weather stations) to stochastic downscaling (Markovian procedure). In most cases the data are downscaled to a daily resolution of several tens of square kilometers, which may be a poor resolution considering the hydrological response of medium and small basins. Downscaling to a resolution appropriate for hydrological prediction (for example, 15 min and 1 km2) must take into account the small-scale structure of rain fields, and particularly patterns of convective rain cells. The information gathered here can be used for this purpose in weather generators to create high-resolution rainfall ensembles under different climate scenarios. Those rainfall ensembles can subsequently be applied to hydrological models for simulating future stream flows and rainfall patterns under climatic change.

8. Conclusions

[48] This study provides information on the spatial and temporal features of convective rain cells over the eastern Mediterranean region. Several principal observations arose from our research:

[49] 1. The three synoptic types—shallow low, Cyprus low and RST (Section 5.2.)—are clearly defined by the four NCEP/NCAR reanalysis parameters: sea level pressure, specific humidity at 700 hPa, geopotential height at 500 hPa and zonal wind at 850 hPa.

[50] 2. The shape of the convective rain cells is described effectively using ellipsoid-shaped segments.

[51] 3. The Cyprus low rain cells are less intense, are slightly larger and move faster than the shallow low rain cells, as they are less productive along the coastline.

[52] 4. The RST rain cells are larger but contain lower rain intensities than the winter low rain cells due to their different trajectories. The RST rain cells' mean life span is also shorter than the mean life span observed for the winter lows (though most of the rain cells “lived” for a mere five minutes).

[53] 5. The preferential orientation of the winter low rain cells is west-east to southwest-northeast and the RST orientation has two modes: west-east to southwest-northeast and north-south. This is associated with the direction and velocity of the motion vector (presumably representing the wind vector) estimated from radar data at an elevation of 2000–2800 m above sea level.

[54] 6. The sea breeze convection is presumably responsible for the increase in the average number of new cells and the increase in the rain cell area near the coastline. The mean cell orientation changes counterclockwise as the cells are offshore and then clockwise from the coastline toward the land.

[55] 7. The mean rainfall intensities peak near the shore and decrease with distance inland, presumably due to the orographic effect.

[56] 8. The RST rain cells have a minor effect on the hydrological systems in the study region (in contrast with southern Israel, where the RST cells are responsible for the majority of the flash flooding).

[57] Based on this research, the next step should be the development of a high-resolution weather generator for creating rainfall ensembles under different climatology scenarios, using information on rain cell characteristics derived from this study. These rainfall ensembles can be applied to hydrological models for the studied Dalya and Taninim basins, simulating stream flow in general and flash floods in particular. The weather generator should be linked to RCM for predicting the rainfall ensembles and hydrological responses under the current and future climates for this area.


[58] Radar data were provided by E.M.S. (Mekorot Company) and rain gauge data were provided by the IMS (Israel Meteorology Service). The authors wish to thank Uri Dayan and Maya Bartov for their helpful comments and assistance in the synoptic classification. The study was funded by the Israel Ministry of Environmental Protection, the Israel Ministry of Agriculture and Rural Development and the Israel Science Foundation (grant 332/11).