On the importance of synoptic classification methods with respect to environmental phenomena



One of the most important goals of synoptic climatology is to analyse the relationships between atmospheric circulation and surface environmental conditions. Since manual (subjective) classification might reflect differently environmental phenomena as compared to computer assisted (semi-objective) classification, a comparison between both classifications was performed for the Eastern Mediterranean (EM) for 10 years. The overall frequencies of the 19 recurring synoptic types characterising the circulation regime over the EM are similar for both methods (correlation coefficient = 0.96, p-value < 0.0001). Nevertheless, comparing the classifications on a day-to-day basis for the three most common types showed a large degree of disagreement between the two. For synoptic types with a deep horizontal pressure gradient both classifications yielded the highest agreement (over 50%), improving the consistency between both classifications.

Yarnal's (1993) ‘environment-to-circulation’ approach was tested on three surface environment processes: air pollutants, desert dust intrusions, and flash floods. Our results indicate that the weak pressure gradients associated with high ozone levels make the classification more difficult for both methods. Regarding dust outbreaks, classifications point on the importance of the cold cyclone location rather than its pressure gradient. As for flash floods, the flow pattern at the surface level is insufficient to predict atmospheric conditions prone for their occurrence, suggesting that upper air data is an essential factor for determining such highly convective events.

Finally, two new approaches to evaluate the quality of classifications are presented, demonstrated on two frequent circulation systems persisting over the EM. A qualitative approach compares the composite mean sea level pressure maps from each classification, while a quantitative approach compares the resultant winds at a central site.

This study demonstrates that comparing different synoptic classification methodologies and different environmental applications can lead to additional and valuable insights on the interaction between the environment and synoptic-scale circulation. Copyright © 2011 Royal Meteorological Society

1. Introduction

Barry and Perry (1973) defined synoptic climatology as the study of the relationship between the atmospheric circulation and local or regional climates. Later, Yarnal (1993) showed in his book that contemporary synoptic climatologists often utilise climatically related variables and refined the Barry and Perry definition stating that synoptic climatology relates the atmospheric circulation to the surface environment.

In order to understand this relationship, synoptic climatologists first classify the atmospheric flow to discrete circulation patterns that typify significant modes of the atmospheric circulation. After these classes have been determined, based on Yarnal (1993) deduction through a literature review, one of the two fundamental approaches to classification is adopted: ‘circulation to environment’ or ‘environment to circulation’. In the first approach, an atmospheric circulation classification is performed and then related to an environmental phenomenon. In the second method, the circulation classification is carried along specific environment-based criteria set for a particular environmental phenomenon. These two approaches were revealed as efficient tools relating circulation to climatically related variables such as air pollutants (e.g. McKendry, 1994; Moulin et al., 1998; Brook and Johnson, 2000; Schwarzhoff and Reid, 2000; Yarnal et al., 2001), heat wave mortality (Rainham et al, 2005) and epidemiology (Goldberg, 2007). Spatial synoptic classification clustered into weather types was recently used as an effective tool to evaluate air quality predictions (Hu et al., 2010). Similar studies relating atmospheric circulation to the surface environment have been performed for the Eastern Mediterranean (EM): for elevated surface ozone concentrations (Dayan and Levy, 2002), dust outbreaks and visual ranges (Dayan et al., 1991; Moulin et al., 1998; Dayan and Levy, 2005; and Dayan et al., 2007) and heavy rain inducing flash-floods (Kahana et al., 2002; Dayan and Morin, 2006; Ziv et al., 2006).

Several map-pattern classification methods exist and continue to be developed to automated techniques. This development is mainly attributed to our understanding of the role of atmospheric circulation on the planetary boundary layer environment and the rapid expansion of computer resources supporting such new techniques.

Manual map-pattern classification predates computers and is the most familiar method among synoptic climatologists. Its main drawbacks are the inherent subjectivity, a possible inconsistency in categorisation and its being labour intensive. However, its major weakness, i.e. its subjectivity, is also its main advantage bringing experience and intuition of the skilled forecaster into the classification process. Second, manual classification is flexible, enabling the possibility to consider the weather circulation on several preceding and succeeding days rather than looking at each day in isolation (Jones et al., 1993).

Contemporary automated classification methods have been broadly reviewed by Yarnal (1993). Among the most popular is the correlation-based method due to its efficient categorisation by a computerised pattern-recognition algorithm. However, this method requires the investigator to make few subjective decisions, e.g. on the size of the grid and number of grid points determining the number of map patterns, the correlation threshold affecting within-group similarity and between-group differentiation. Another caveat is their limited ability to identify patterns, often important though, infrequent for specific surface environmental phenomena (Huth, 1996). Since, actually, all the methods have some inherent subjectivity, Yarnal and White (1987) suggested to abandon the term ‘objective classification’ and to call, instead, all automated classifications ‘computer assisted’.

Currently, it seems that there is no unique classification method being the best and recommended one (Huth et al., 2008). As Frakes and Yarnal (1997) concluded in their paper: “Each classification method offers advantages and disadvantages, and selecting the appropriate technique depends on a balance among the needs of the research, the skills of the investigator and the nature of the data”.

Few synoptic classifications were performed for the EM region. Koplowitz (1973) carried out the first ‘computer assisted’ classification of the Sea-Level-Pressure (SLP) field pattern. Ronberg (1984) performed an automated categorisation, reducing upper and surface data by Principal Component Analysis to several weather types that were afterwards sub-divided subjectively into four distinct groups affecting the EM. Alpert et al, (2004) performed a blended manual and automated classification for daily synoptic systems over the EM based on a Discriminant Analysis methodology.

Since manual classification can lead to different environmental implications as compared to computer-assisted classifications, a comparison between a subjective (manual) versus semi-objective (computer assisted) classification was performed for the EM for 10 years (1995–2004). In order to examine how these varying classification methods reflect the relationship to surface environmental conditions, several environmental phenomena characterising the EM (elevated surface ozone concentrations, mineral dust outbreaks, and flash flood events) were analysed along the “environment to classification” approach.

Basic results of the comparison between both classifications are given in the next section. An assessment of the implications of the synoptic classifications on environmental phenomena featuring the EM is presented in Section 3. Section 4 demonstrates two new methods allowing a preliminary evaluation for the performance of the classification methods. The paper concludes by summarising the results.

2. Comparison of the manual and computer assisted classifications

Both manual and computer assisted classifications applied different techniques and principles in the categorisation process of daily synoptic patterns. The manual classification relied solely on expert judgment and was based on SLP maps of 12 UTC (14 local time), with isobars intervals of 2.5 hPa for a grid covering the area between 20°N, 15°E to 50°N, 45°E. The chosen time of these maps reflects best the evolution of environmental phenomena analysed. This classification process is performed routinely at the Synoptic Meteorology Laboratory at the Department of Geography, Hebrew University of Jerusalem.

Whereas the manual classification relies exclusively on expert judgment, in the computer assisted classification done by Alpert et al. (2004) a panel of experts had subjectively predefined synoptic categories for one year. This year then served as a training set for the Discriminant Analysis program that automatically classified the entire dataset. The dataset includes four equally weighted parameters: geopotential height (H), temperature (T) and two horizontal wind components (U and V) at 1000 hPa. The size of the grid used by this methodology as compared to the manual methodology is considerably smaller − 27.5°N, 30°E to 37.5°N, 40°E. This grid includes 25 grid points with an interval of 2.5°. Hence, 4 fields (H,T,U,V) at 25 points yield 100 variables for one pressure surface only. In the final stage, Euclidean distances from the expert-predefined days were calculated. The predefined day to which the minimum distance was found determined the class to which the specific day should be assigned (Alpert et al., 2004).

Data for both classifications was extracted from the NCEP/NCAR Reanalysis website (http://dss.ucar.edu/pub/reanalysis/) along a model developed by Kalnay et al., 1996 and Kistler et al., 2001. This operational model assimilates meteorological observations from global stations and calibrates its output so as to minimise the model errors in relation to the available data (see Kalnay et al., 1996, for details).

It should be stressed that classifications differ slightly in the number of synoptic categories used. While the computer assisted classification attributed each day to one of 19 categories, the manual classification used four additional patterns, including an ‘undefined’ one. The latter category was included in order to avoid the categorisation of an indistinct synoptic pattern. Altogether, the total number of days attributed to the four additional patterns used by the manual classification merely amounted to 259 (∼7%) for the entire 10-year period.

Both classifications are comparable since they are fundamentally based on SLP 12 UTC maps, defined subjectively by panels of experts, and share the same 19 synoptic categories. Classification results for both methodologies are presented in Table I which exhibits the frequencies of each synoptic pattern for the 19 categories defined by Alpert et al. (2004). The calculation of the frequencies (as percent) for the manual classification was adjusted so as not to include days of additional synoptic systems used by this methodology. The names of the synoptic patterns are widely used in most of the EM countries as mentioned by Alpert et al. (2004). Nomenclature of the synoptic patterns was determined along the location of Israel, as representing the EM region.

Table I. Summary of frequencies of the EM synoptic systems for the years 1995–2004. Three most frequent systems are marked in bold font
Synoptic patternManual classificationComputer assisted classification
 N (days)Percent of the 10-yr periodN (days)Percent of the 10-yr period
Red Sea trough with eastern axis52915.652514.4
Red Sea trough with western axis391.2391.1
Red Sea trough with central axis2918.63309
Weak Persian trough66219.559216.2
Moderate Persian trough2687.947212.9
Deep Persian trough802.4591.6
High to the east1283.81173.2
High to the west5411659216.2
High to the north1765.21815
High over Israel2086.11313.6
Deep low to the east441.3521.4
Deep Cyprus low to the south10.120.1
Shallow Cyprus low to the south120.430.1
Deep Cyprus low to the north361.11042.8
Shallow Cyprus low to the north1293.81604.4
Cold low to the west290.9922.5
Shallow low to the east17151584.3
Sharav low to the west290.9140.4
Sharav low over Israel210.6300.8

Table I indicates that both classifications yielded similar frequencies for the 19 synoptic patterns analysed throughout the period, resulting in a correlation coefficient of 0.96 (p-value < 0.0001). However, total frequencies do not indicate whether classifications exhibit high similarity on a finer annual basis. Therefore, we examined whether both classifications obtained the same trend of annual frequencies for the three most frequent categories (Red Sea trough with eastern axis, Weak Persian trough, and High to the west). The Red Sea trough (RST) is an extension of a low surface pressure from a tropical depression toward the Red Sea occurring during transitional seasons and the winter. Persian trough is a surface thermal barometric system centred in the Persian Gulf and extending from the southwest Asian monsoon. This trough begins to occur in the middle of June and remains until mid-September. About 20% of the days belonging to the transitional seasons are characterised by a closed high-pressure system over the eastern Mediterranean Basin (i.e. High to the west), although this synoptic system can prevail also during the winter and to a lesser extent during the summer.

Time series of frequencies for the mentioned synoptic systems are shown in Figure 1a)–c). These three time series indicate that although the correlation between classifications is high for the entire 10-year period, it is clear that both methodologies yield only partial agreement when examined on an annual basis. The correlation coefficients for both classification methods for the three synoptic systems are 0.20 (p-value = 0.58), 0.46 (p-value = 0.18) and 0.58 (p-value = 0.08) for a), b), and c), respectively, in Figure 1. According to both methods, a positive trend is apparent for the weak Persian trough leading to warm conditions over the region. This finding indicates a general trend of warming over the EM, consistent with previous studies (Ziv et al., 2005; Saaroni et al., 2003, 2010).

Figure 1.

Annual trends in frequencies of the three most common synoptic patterns for 1995–2004: a) RST with eastern axis; b) Weak Persian trough; c) High to the west

However, in order to derive the implications of environmental phenomena as related to atmospheric circulation, a much finer temporal resolution is required. Table II summarises the differences between classifications on a daily resolution. A date which was classified differently for both classification methodologies is considered a ‘mismatch’. The percent of mismatches was calculated by summing the number of mismatches per category, and arbitrarily dividing it by the frequency of the category in the manual classification.

Table II. Mismatches between classifications. High agreements are marked in bold font, low agreements in italics
Synoptic patternNaMismatches (days)Mismatches (%)b
  • a

    Days in manual classification.

  • b

    Mismatch days as percent of manual observations.

Red Sea trough with eastern axis52931660
Red Sea trough with western axis393795
Red Sea trough with central axis29116958
Weak Persian trough66234853
Moderate Persian trough26814755
Deep Persian trough807493
High to the east1288667
High to the west54137469
High to the north17613376
High over Israel20817082
Deep low to the east442761
Deep Cyprus low to the south100
Shallow Cyprus low to the south121192
Deep Cyprus low to the north361747
Shallow Cyprus Low to the North12910380
Cold low to the west291448
Shallow low to the east17112774
Sharav low to the west292793
Sharav low over Israel211571

Results obtained by the comparison of mismatches indicate that for synoptic systems characterised by a deep horizontal pressure gradient, i.e. Deep Cyprus low to the north and Cold low to the west, both classifications yielded the highest agreement (over 50%), in spite of the low frequencies of these synoptic categories. This finding is consistent with Frakes and Yarnal (1997) stating that replication rate of manual classification approaches 90% during the winter when pressure patterns are well defined with steep gradients. Typical SLP maps of the Deep Cyprus low to the north and the Cold low to the west are shown in Figure 2a) and b), respectively. The SLP composites are calculated for days in which both classifications yielded identical results.

Figure 2.

Composite SLP (hPa) maps for synoptic patterns of deep pressure gradient yielding the highest agreement for both classification methods: a) Deep Cyprus low to the north (N = 19); and b) Cold low to the west (N = 15)

However, the agreement between classification results of synoptic types having important environmental implications in the EM, i.e. RST with western axis, Deep Persian trough and Sharav low to the west did not exceed 7%. The first type is associated with torrential rain inducing flash floods over desert areas (Kahana et al., 2002), the second is responsible for long-range transport of pollutants (Erel et al., 2007) which is not dealt with in this study, and the last is prone for dust outbreaks (Israelevich et al., 2002, 2003; Dayan et al., 2007). Average SLP maps for the RST with western axis (Figure 3a)), Deep Persian trough (Figure 3b)) and Sharav low to the west (Figure 3c)) are shown for days of agreement between classifications.

Figure 3.

Composite SLP (hPa) maps of synoptic types having important environmental implications in the EM: a) RST with western axis (N = 2); b) Deep Persian trough (N = 6); and c) Sharav low to the west (N = 2)

3. Classification results related to environmental phenomena

Environmental phenomena are mostly the result of distinct synoptic-scale events that interact with smaller scale flows (e.g. land–sea breeze or anabatic–katabatic winds), rather than the outcome of purely local weather conditions (Kahana et al., 2002). Accordingly, this study examines conditions on the large synoptic scale, enabling the smaller scales to come into effect and result in the discussed environmental phenomena.

In order to examine the synoptic circulation systems associated with significant environmental conditions at the surface for each classification method, the ‘environment-to-circulation’ approach was used. For this sake, environment-specific criteria were set for three environmental phenomena characterising the EM: elevated concentrations of air pollutants, desert dust intrusions, and flash floods. The circulation pattern was derived for the selected dates meeting these criteria for each phenomenon.

3.1. Air pollutants—ozone

Numerous studies analysed the relationships between atmospheric circulation categories and air pollutants in general. More specifically, this relationship has been examined with respect to secondary pollutants (e.g. ozone) (Wolf and Lioy, 1980; Heidorn and Yap, 1986; Comrie and Yarnal, 1992; McKendry, 1994; Pryor et al., 1995; Schwarzhoff and Reid, 2000; Dayan and Levy, 2002). The satisfactory results of these circulation-ozone studies are due to the typical formation time of this secondary pollutant and its dependence on insolation and ambient temperature, two factors governed by the large-scale meteorology. These result in a fitting of the spatio-temporal distribution of ozone to the spatio-temporal distribution of barometric systems in the synoptic scale. This compatibility leads to distinctive seasonal and diurnal variations in ambient ozone concentrations obtained for the particular modes of atmospheric circulation.

In order to point at the differences in the categorisation of synoptic patterns associated with elevated surface ozone concentrations between both classification methods, data from the Ben-Shemen monitoring station located in the central part of Israel were analysed. This site was selected since it is a rural site located downwind of the largest metropolitan area in Israel: Tel Aviv (Figure 4). Transport of ozone from the coastal urban area inland was demonstrated in previous studies by Luria et al. (1984) and confirmed by Lifshitz et al. (1988) and Weinroth et al. (2008) as one of the mechanisms leading to high ozone levels often measured during the summer in the foothills of the Judean Hills.

Figure 4.

Study area showing the location of the Ben-Shemen air quality monitoring site, Beit Dagan radiosonde launch site, and major cities in Israel

Relying on 4-year data (2001–2004), we defined a ‘high-ozone day’ as the occurrence of 4 consecutive 30-min average concentrations of ozone between 13:30 and 15:00 PM (local time) each exceeding a threshold of 70 ppbv, (60% of the Israeli Air Quality Standard) which is most likely to be found during late noon (Dayan and Levy, 2002). In total, 29 days were found to have elevated surface ozone concentrations. Frequencies of the defined synoptic patterns for these days are presented for both classification methods (Table III).

Table III. Frequencies of synoptic patterns associated with elevated surface ozone concentrations
Synoptic patternFrequency (Days)
 Manual classificationComputer assisted classification
  • a

    Seven days of elevated surface ozone concentrations were attributed to the undefined category.

Red Sea trough with eastern axis210
Red Sea trough with western axis10
Red Sea trough with central axis64
Weak Persian trough34
Moderate Persian trough02
High to the west23
High to the north40
High over Israel32
Deep Cyprus low to the north01
Shallow Cyprus low to the north11
Sharav low to the west02
N (for the 19 comparable categories)22a29

The most frequent synoptic pattern leading to elevated surface ozone concentrations along the computer assisted classification is the RST with eastern axis. In comparison, the manual classification points at the RST with central axis as the most frequent category. Nevertheless, the composite SLP maps (not shown) showed no difference for both categories, suggesting that both classifications are in agreement and attribute the same flow pattern to different classes, i.e. that the division of the RST to three distinct categories is too refined when assigned to elevated surface ozone concentrations. The resembling composite maps stem from the inherent weak pressure gradient associated with the RST. This latter feature complicates the attribution of flow patterns to a distinct category.

Since the RST in its three modes (eastern, central, and western axes) is characterised by dry and warm air mass influencing the region (Dayan and Levy, 2002), these modes are all prone to elevated surface ozone concentrations.

3.2. Dust outbreaks

The frequency of dust intrusions over the EM is strongly related to the climatology of depressions affecting the region (Moulin et al., 1998). Synoptically, the majority of dust storms which occur above the EM over the year are caused by an increased surface pressure gradient. They are usually associated with the passage of a cold low-pressure system. Dayan et al. (2007) found that the Cyprus low is the main contributor to the suspended dust over this region, being four times larger than that of the second one, the Sharav cyclone. This is explained by the low occurrence of Sharav cyclones crossing the region (Alpert et al.2004), in addition to their weaker wind speeds restricting mobilisation and transport of dust to the EM as compared to cold cyclones. This striking finding contradicts the widely accepted idea that the Sharav cyclone is the major synoptic-scale system responsible for dust transport to the study region. Further details on this typical spring synoptic system are given by Alpert and Ziv (1989).

In order to determine the differences in the attribution of synoptic patterns to dust outbreaks, we followed the same methodology used with respect to ozone concentrations, for the period of 2000–2004. We defined dust outbreaks as days for which concentration of suspended PM10 exceeded an average 30-min threshold of 100 µg m−3 in at least 18 of the 31 monitoring stations in Israel. Since some of these days were consecutive, the centres of dust outbreaks were defined as the date in which the maximal concentration was recorded in any of the stations.

53 ‘high dust days’ met the threshold value for the 5-year period. Out of these days, 22 days were considered as ‘severe dust days’, where PM10 values exceeded an average 30-min concentration of 1000 µg m−3 in at least 5 of the network stations. Results of the classification for both thresholds are presented in Table IV.

Table IV. Frequencies of synoptic patterns associated with dust concentrations exceeding the lower threshold (100 µg m−3) and the higher threshold (1000 µg m−3)
Synoptic patternFrequency (Days)
 Manual classificationComputer assisted classification
 > 100 µg m−3> 1000 µg m−3> 100 µg m−3> 1000 µg m−3
Red Sea trough with eastern axis5070
Red Sea trough with western axis2111
Red Sea trough with central axis4060
Weak Persian trough0020
Moderate Persian trough0022
Deep Persian trough0021
High to the east4010
High to the west4210
High to the north1010
High over Israel1011
Deep low to the east1011
Deep Cyprus low to the south0011
Shallow Cyprus low to the south0010
Deep Cyprus low to the north32105
Shallow Cyprus low to the north10762
Cold low to the west2183
Shallow low to the east4100
Sharav low to the west3100
Sharav low over Israel3322
N (for the 19 comparable categories)47185319

As Table IV indicates, both classifications found that cold lows to the north (either deep or shallow modes) are the synoptic patterns most frequently causing dust outbreaks, of which are characterised by dust concentrations exceeding 1000 µg m−3. Both classifications also agree that Red Sea troughs are a major source of more moderate dust episodes in the EM.

When the classification results for severe dust days are examined, cold lows to the north of Israel remain the synoptic patterns most commonly related to dust outbreaks by both manual and computer assisted classifications. The manual classification attributes most severe dust outbreaks to the shallow mode of the cold low to the north, whereas the computer assisted classification categorised the most frequent synoptic system in this case as belonging to its deep mode. Average SLP maps for both classifications are shown in Figure 5.

Figure 5.

Composite SLP maps (hPa) characterising severe dust outbreaks over the EM. Maps are based on all days meeting the environmental criteria (7 for the manual classification, 5 for the computer assisted classification): a) manual classification (shallow Cyprus low to the north); and b) computer assisted classification (deep Cyprus low to the north)

The average maps demonstrate that the methodologies indeed yield different results. Even though two of the averaged dates are overlapping, each classification attributes elevated mineral dust concentrations to cold lows characterised by a different gradient pressure. The similarity in the frequencies of cold lows to the north associated with elevated dust concentrations for both modes and both thresholds implies that such outbreaks occur for both modes of this synoptic system.

Since the pertaining associated winds for both synoptic types are westerly to southwesterly over the EM, and averages of dust concentrations are very similar for both classifications (3049 and 2905 µg m−3 for the manual and computer assisted methodologies respectively), we conclude that the most important factor for dust transport by cold lows to the region is wind direction rather than wind speed. Contrary to the conclusion reached in the previous section, stating that refinement in the RST classification for elevated surface ozone concentrations is unnecessary, the sub-division of cold lows to the north into the two modes enriches our understanding on the relationship between circulation and dust episodes in the EM.

3.3. Flash floods

Kahana et al., (2002) used a flood database for the Negev Desert, southern Israel, for a 30-year period (1965–1994) together with atmospheric data retrieved from the US National Meteorological Center to investigate the synoptic conditions with which the major floods in this region are associated. A systematic classification revealed that several synoptic types produce the majority of the largest Negev floods. 20 of the 52 major flood cases (38%) were attributed to a well defined synoptic type: an Active Red Sea trough (ARST) which is essentially a variation of a RST with a western axis. The ARST is defined as a surface trough extending from East Africa through the Red Sea toward the Eastern Mediterranean, accompanied by a pronounced trough at the 500 hPa level over eastern Egypt (Kahana et al., 2002).

Following the same approach implemented for the previous two environmental applications, we analysed the frequencies of synoptic patterns causing flash floods in the semi-arid to arid south of Israel. Since the number of floods is small for the decade analysed, we expanded the sample by adding another year (1994) for which a synoptic classification was undertaken for days when a flood occurred. Dates of recorded flash floods characterised by recurrence intervals of at least two years were obtained from the Israeli Hydrological Service (personal communication). The results obtained by both classifications are presented in Table V. Owing to the limited number of storms and their large inter-annual variation (Kahana, et al., 2004), results should be dealt with caution.

Table V. Frequencies of synoptic patterns associated with flash floods for the Negev Desert, southern Israel
Synoptic patternFrequency (Days)
 Manual classificationComputer assisted classification
Red Sea trough with eastern axis34
Red Sea trough with western axis30
Red Sea trough with central axis33
High to the east01
High to the west21
High to the north01
Deep cold low to the north02
Deep cold low to the west21
Shallow cold low to the east23
N (for the 19 comparable categories)1516

Both classifications find Red Sea troughs as the most frequent synoptic patterns causing flash floods in southern Israel. According to the manual classification, a third of the RSTs were classified as RSTs with a western axis, consistent with a study performed by Kahana et al. (2002) for 30 years. However, this type was not identified by the computer assisted classification. Further analysis indicated that out of the three RSTs with western axis identified by the manual classification, two were classified as RSTs with central axes by the computer assisted classification on days when a flash flood occurred. This finding suggests a possible weakness and an inherent bias in the identification of this surface synoptic type by the computer assisted classification.

An ARST is characterised not only by surface conditions but also by conditions prevailing in upper atmospheric layers. However, both classification methods rely on SLP data solely, which hampers the identification of the ARST as producing heavy rainfall inducing flash floods.

The main implication of this finding is that upper air data is essential for both methodologies in order to adequately identify ARST's leading to floods in southern Israel.

4. Comparison of the quality of classifications

Since “currently there is no apparent objective or subjective means for determining a “best” classification method” (Huth et al., 2008), we take an initial step to compensate for this problem, demonstrating both a qualitative and a quantitative approach to compare the quality of classifications. These approaches were implemented with respect to two frequent circulation systems persisting over the EM, and reflect the quality of classifications on both regional and local scales: RST with central axis (Figure 6a)) and High to the west (Figure 7a)). These patterns were chosen due to their high frequencies (∼9% and ∼16% for the RST with central axis and High to the west, respectively, along both classifications) and consensus on their typical wind flow, relevant for the demonstration of the quantitative approach.

Figure 6.

RST with central axis composite SLP maps (hPa) for observations comprising the three groups: a) reference group (N = 122); b) manual classification only (N = 169); and c) computer assisted classification only (N = 208); as well as d) difference between a) and b); and e) difference between a) and c)

Figure 7.

High to the west composite SLP maps (hPa) for observations comprising the three groups: a) reference group (N = 167); b) manual classification only (N = 374); and c) computer assisted classification only (N = 425); as well as d) difference between a) and b); and e) difference between a) and c)

Both quality evaluation methods rely on the same fundamental approach: based on the 10-yr synoptic categorisation, days in which these categories were identified by either of the classifications were selected for the analyses. These days were divided to three groups for each synoptic type. The first group consists of days of agreement for both classifications (i.e. same classification for a given date, hereinafter ‘reference group’). The second and third groups comprise days where the analysed synoptic system was identified by only one of the classification methods.

Days belonging to the reference groups were considered as having the highest chance for being ‘properly classified’, therefore, used as reference for the comparison between the remaining two groups in each of the analyses.

4.1. Qualitative comparisons

The comparison of classifications starts with the qualitative approach, which provides a regional perspective on the quality of classifications, covering the entire EM. This comparison is based on the composite SLP synoptic maps plotted for each of the three groups (Figure 6a)–c)) for RST, and Figure 7a)–c) for High to the west). Examining these maps can help determining which of the maps representing each of the classifications resembles more the map representing the reference group. To ease the visual comparison, the difference between the reference group and each of the other groups is also shown (Figure 6d) and e)) for RST with central axis, and Figure 7d) and e) for High to the west.

Starting with the RST with central axis, Figure 6a)–c) indicate that the synoptic maps indeed exhibit high similarities, and the difference maps (Figure 6d) and e)) show only small absolute deviations of up to 1.5 hPa from the reference group. It is interesting to note that the difference of the computer assisted classification from the reference group (Figure 6e)) follows a spatial pattern similar to the RST with central axis itself. This pattern indicates that, on average, the computer assisted classification does produce the ‘correct’ spatial pattern of this synoptic pattern, though with biased values of pressure. However, this comparison does not enable to determine unequivocally the superiority of either of the classification methods with respect to this synoptic type.

Considering the second synoptic system analysed—High to the west—the composite maps (Figure 7a)–c)) suggest that the manual classification exhibits stronger similarity to the reference group than the computer assisted classification, as both maps show a well defined pressure core of the high. This is further emphasised by the differences maps (Figure 7d) and e)) where the absolute differences in pressure between the computer assisted classification and the reference group (–3–0.5 hPa) are larger than those obtained for the manual classification (−1.5–1 hPa), indicating that the latter better classifies this synoptic system than the computer assisted classification.

4.2. Quantitative comparisons

The quantitative comparisons are based on flow parameters of wind speed and direction. These parameters have substantial environmental implications, having a dominant effect on the dispersion of air pollutants. Before the elaboration of this method, it should be stressed that since wind flow parameters are extracted locally, the quantitative comparison conducted in this paper serves as an example, and does not cover the entire EM.

The quantitative comparison was conducted as follows: wind speed and direction at 1000 hPa were extracted from the 12 UTC Beit Dagan (the Israeli Meteorological Service national launching site, Figure 4) radiosonde measurements for each day of the study period. The 1000 hPa level is within the boundary layer, and therefore relevant for most environmental applications. In order to avoid ground interactions with shallow atmospheric layers resulting in inconsistent wind flow, only days with recorded wind speeds exceeding 5 m s−1 were included in the analysis, resulting in a smaller number of observations than those used in the qualitative approach. Figures 8 and 9 show polar scatter plots of wind data from days belonging to each of the groups for RST with central axis and High to the west, respectively. Next, several parameters characterising the wind flow for each synoptic pattern were calculated (Tables VI and VII). These parameters include the averages of wind direction, east–west (U) and north–south (V) wind components, and the scalar mean wind speed (i.e. arithmetic mean), as well as their standard deviations. Additionally, the vector mean wind speed (i.e. (U2 + V2)0.5), was calculated. The difference between the scalar mean wind speed and the vector mean wind speed indicates the wind's tendency to blow from the same direction on different days, where small differences indicate greater persistency. Large deviations of the mentioned parameters from those calculated for the reference group should be interpreted as being of a rather weak classification. Consequently, these calculations enable to estimate quantitatively, and on a local scale, the extent to which each of the classifications deviates from the closest ‘agreed classification’ for each of the synoptic systems.

Figure 8.

Polar plots showing wind direction and speed for observations comprising the three groups for the RST with central axis type: a) reference group; b) manual classification group; and c) computer assisted classification group

Figure 9.

Polar plots showing wind direction and speed for observations comprising the three groups for the High to the west type: a) reference group; b) manual classification group; and c) computer assisted classification group

Table VI. Wind parameters characterising Red Sea troughs with central axis
 Days of agreementManual classificationComputer assisted classification
Average wind direction89°141°352°
Scalar mean wind speed (m s−1)
Stdev of scalar mean wind speed (m s−1)
Vector mean wind speed (m s−1)
Average V (m s−1)− 0.040.84− 2.64
Stdev V (m s−1)
Average U (m s−1)− 3.59− 0.690.38
Stdev U (m s−1)
Table VII. Wind parameters characterising Highs to the west
 Days of agreementManual classificationComputer assisted classification
Average wind direction281°273°299°
Scalar mean wind speed (m s−1)
Stdev of scalar mean wind speed (m s−1)
Vector mean wind speed (m s−1)
Average V (m s−1)− 0.94− 0.23− 2.23
Stdev V (m s−1)
Average U (m s−1)4.815.163.9
Stdev U (m s−1)

Starting with the RST with central axis, Table VI shows that the mean wind speed is similar for all groups, leaving wind direction as the important factor for the comparison. Interestingly, the reference group yielded a dominant southeasterly wind meeting the expected flow regime for this synoptic category, and a secondary northwesterly sector explained by the interaction of the sea breeze with the prevailing synoptic system (Figure 8a)). The uniformity of both wind speed and direction of the reference group (as displayed by the low standard deviations of all flow parameters shown in Table VI) justifies the selection of the ‘days of agreement’ group as a robust reference. The manual classification yielded a mean direction closer to that of the reference group, though characterised by large standard deviations in both U and V components and wind speed (Table VI and Figure 8a) and b)). This large scatter is partly compensated by the modal wind sector obtained (90–135°) which is identical to the modal sector obtained for the reference group. However, the lack of uniformity for both wind direction and speed reduces its applicability as the preferable method for the prediction of surface environmental conditions associated with this particular synoptic system.

The computer assisted classification yielded a bi-modal wind sector, concentrated mostly between 315 and 360° and 90 and 135°. This bi-modality is further emphasised when combining the computer assisted classification and the reference group (Figure 8a) and c)). However, its average wind direction deviates from that of the reference group by almost 100°, where the manual classification deviation is half that value.

Even though this classification does not yield a homogeneous wind sector, it seems to be a preferable method to relate environmental phenomena affected by wind field characteristics for this synoptic circulation system. This stems from the apparent bias of this classification evidenced also in the qualitative approach, enabling its correction.

Turning to the other synoptic pattern analysed—High to the west—the comparison yielded different results. Contrary to the polar plots obtained for the RST with central axis, the plots for this synoptic type do not point at distinct differences between classifications (Figure 9a)–c)). Some disagreement can be noted with respect to the average wind direction obtained for the manual classification, being closer to that of the reference group (Table VII). These results indicate that as opposed to the previous synoptic type analysed, we cannot point at any of the classification methods as being preferred over the other.

5. Conclusions

The synoptic-scale flow in the atmosphere plays a leading role for many environmental phenomena, either directly or indirectly by allowing meso-scale processes to come into effect (Ulrickson and Mass, 1990; Kalthoff et al., 2003). These phenomena include, for example, flash floods, dust outbreaks, and elevated ozone spells.

In order to compare between a manual and a computer assisted classifications of atmospheric circulation, their environmental outcomes in terms of episodes of elevated surface ozone concentrations, dust outbreaks and flash floods were examined over a period of 10 years (1995–2004) over the EM region. Since these environmental phenomena are mainly the result of the larger synoptic scale, an environment-to-circulation approach was implemented. Among the 19 recurring synoptic types characterising the circulation regime over the EM, the overall frequencies of the three most common synoptic types—RST with eastern axis—Weak Persian trough and High to the west were similar for both classifications. The mismatches on a day-to-day basis for these 3 types were 60, 53, and 69% respectively. However, for deep horizontal pressure gradient synoptic types, i.e. Deep Cyprus low to the north and Cold low to the west, both classifications yielded the highest agreement (over 50%). The observed trend of increase in the annual frequency for the Weak Persian trough indicates a warming over the EM during the 10 years analysed.

Classification results of synoptic types having important environmental implications in the EM, i.e. RST with western axis, Deep Persian trough and Sharav low to the west, differed substantially (over 90% disagreement) for both methodologies.

The essential results on the classification methods as reflected by environmental phenomena are: (1) With respect to elevated surface ozone concentrations, no substantial differences between classifications were found. This result stems from the high similarity in weather conditions associated with the three modes of the RST, implying a too-refined categorisation for this pollutant. Furthermore, the weak pressure gradient associated with these flow patterns makes their classification more complicated for both methods; (2) It was found that the location of the cold cyclone leading to dust outbreaks in the EM is more important than its pressure gradient; (3) As with ozone episodes, the location of the RST axis is not indicative of atmospheric conditions prone for flash floods occurrence. However, it should be noted that both classifications lack upper air data which is an essential factor for determining such highly convective events.

Two methods are presented for evaluating the quality of the classifications, a qualitative and a quantitative. These are applied to two highly frequent synoptic patterns. The conclusions derived from the evaluation indicate the following: (1) The computer assisted classification results are inherently clustered in their wind direction and speed for the several synoptic systems; (2) In contradiction to the manual classification, the computer assisted method can be improved by re-examining, subjectively, the days suspected as biased for certain synoptic types; (3) Quasi-stationary synoptic systems of large dimension such as High to the west yield resembling wind field characteristics for both classification methods.

Since the results obtained by both methods (qualitative and quantitative) for evaluating the quality of classification are exclusive to each of the synoptic types analysed, further research should broaden the comparison to the various synoptic systems prevailing over the EM.

This study demonstrates that comparing different synoptic classification methodologies and different environmental applications can lead to additional and valuable insights on the interaction between the environment and synoptic-scale circulation.


The authors would like to thank Eliav Schmulewitz of the Synoptic Meteorology Laboratory at the Department of Geography, Hebrew University of Jerusalem, for his contribution in the processing of data used in this article. The second author is grateful for additional funding given by the Amiran Fund, Hebrew University of Jerusalem. The third author gratefully acknowledges post-doctoral fellowship support from the Environment and Health Fund, Jerusalem, Israel. We wish to thank the two anonymous reviewers for their valuable comments and suggestions.