Analysis of Winds Affecting Air Pollutant Transport at La Plata , Argentina

An hourly wind analysis for the populated area of La Plata city (with high industrial, power station and vehicular activities) is presented and discussed. Euclidean distance and minimum covariance determinant (a robust correlation coefficient) are employed, as similarity approaches, in order to compare observed wind direction frequency patterns at two monitoring sites during 1998-2003. A preliminary assessment of two sectors, namely Sector 1 (NNW-N-NNE-NE) and Sector 2 (ENE-E-ESE), relevant for the transport of industrial air pollutants towards population exposed, is discussed taking variances into account and employing a locally weighted smoothing approach (LOESS). Both similarity approaches allowed gain insight of wind patterns. The distance approach showed good similarity between sites while the correlation approach showed an uneven picture depending on the wind direction. Most of the differences are explained in terms of the sea-land breeze effect but also differences in terrain roughness and data quality are taken into account. Winds from sectors 1 or 2 (analyzed during 1998-2009) may occur more than 50% of the time, most of the differences regarding the influence of the day and the season on these sectors are attributable to sea-land breeze phenomena. The LOESS proved to be appropriate to analyze the stability with time of both sectors and to discard possible remaining patterns; results are in accordance with studies that assess the interannual variability for different variables in La Plata river area. The robust correlation coefficient revealed, as an example, the linear character of dependence between winds from sector 2 and sulfur dioxide concentrations. Wind velocities and calms are also discussed.


Introduction
Urban areas are major sources of air pollution which is the result of processes of accumulation, dispersion, transformation and removal of contaminants.Pollutant emissions affecting air quality in cities have an adverse impact on human health [1].The city of La Plata and its surroundings are densely populated areas (approximately 800,000 inhabitants) with high industrial and vehicular activity.Glassmann and Mazzeo [2] made a regional study of air pollution potential in Argentina and concluded that La Plata is located in a zone with poor at-mospheric self-cleansing capacity.In spite of these facts, no governmental air monitoring network is installed.In recent years several works contributed to assess different parameters characterizing air pollutants in the area [3][4][5][6][7] but references to hourly distribution of winds and sealand breeze effects have been rare.The characterization of wind direction frequency patterns associated with industrial sources is essential for describing the transporttation of air pollutants and settles the basis for a further assessing of environmental impact on human health.
Enriching previous reports this article make available a detailed summary of wind direction frequencies while discusses their importance for air pollutant modeling.
One purpose of this paper is to compare observed wind direction frequency patterns covering the period 1998-2003 at two monitoring sites (named Point A and J Figure 1).A previous report [8] analyzed similarity between both sites employing cluster and multidimensional scaling analysis.The present study is intended to gain insight of hourly occurrences of winds regarding air pollutants.To this end two different approaches are involved: one quantifies proximity using the squared Euclidean distance and the other quantifies correlation applying to the minimum covariance determinant (MCD) estimator.The well known squared Euclidean distance between two patterns expresses how far these two objects are from each other.It is in fact a dissimilarity measure [9] because the distance increases as the similarity decreases.Although non conventionally used in atmospheric sciences, the MCD coefficient (a linear robust correlation estimator) was adopted instead of the traditional Pearson's coefficient because it reduces the influence of outliers [10].Both similarity approaches are intended to reveal different aspects of wind pattern characteristics.While MCD allows comparing pattern "shapes", Euclidean distance allows comparing "sizes" [11].One wind pattern will be considered similar to another insofar as they are positively correlated and close one to each other.
In a previous work [12] two sectors relevant for the transport of air pollutants from the Industrial Pole to populated areas covering the period 1997-2000 have been emphasized, namely: Sector 1 (NNW-N-NNE-NE) transporting pollutants towards the city center and Sector 2 (ENE-E-ESE) transporting pollutants to main residential areas (see down left corner in Figure 1).In the present work the period under study is expanded and the discussion is intensified.The second purpose of this paper is to assess the presence of these sectors and to consider their daily and seasonal variations associated with the daily and annual cycles.Also their stability with time is evaluated by employing a local weighted non-parametric regression method (LOESS) increasingly used in environmental sciences [13,14].The period involved to analyze both sectors is 1998-2003 for sites A and J (selected because of the availability of simultaneous data) and 1998-2009 for site J (the largest data set available).SO 2 is often taken as a witness gas of industrial activity.An increasing annual trend was detected close to the Industrial Pole between 1998 and 2000, surpassing 26 ppbv for the year 2000 [15].Being industrial air pollutant hourly data scarce, we employed observed SO 2 concentrations published in a previous report for correlation purposes.
Additionally, observed seasonal averaged wind velocityies at both sites are compared taking into account differences in instrument exposure.Observed calms are provided for context.

Characteristics of the Region under Study
La Plata is located in eastern Argentina (35˚S, 58˚W) on the estuary of the De La Plata River, which is one of the most important rivers in South America (its basins covers 3,200,000 km 2 ) and is part of the boundary between Argentina and Uruguay.The city centre is located about 11 km far from the Río de La Plata bank in the typical "Pampa" prairies (see Figure 1).The average elevation is 15 m above the mean sea level.La Plata River estuary is wide enough to have relevant sea-land circulations.From a geographical point of view the diurnal cycle of the sea-breeze is expected to occur between NNW and ESE (clockwise).According to Thornthwaite's [16] classification La Plata climate is "wet, mesothermal with null or small water deficiency".The annual mean temperature is 16˚C.January is the hottest month with 22.4˚C and July the coldest month with 9.9˚C.The annual average relative humidity is 70% with a minimum in January (60%) and a maximum in July (85%) [17].According to the National Meteorological Service for the last three decades (1981-2010) predominant wind directions for 8-direction wind roses registered at La Plata Airport (see Figure 1) are E, NE and N [18][19][20].
A large Industrial Pole (including oil refinery plants, a major shipyard, steel processing plants, etc.) located between the river and the city (see Figure 1) concentrates major industrial air pollutants sources for the city and surroundings.A new thermal power station (with a capacity of 560 MW) constructed in the vicinity areas of the industrial complex is announced to be put in operation during 2012.

Instrumentation and Data Characteristics
Monitoring site labeled as Point A is located in a urban area and belongs to Universidad Tecnológica Nacional (UTN).It operates a Weather Monitor II Euro Version ® meteorological station (Davis Instruments, CA).Monitoring site Point J is in a semi-rural area belonging to Estación Agrometeorológica Julio Hirschhorn-Universidad Nacional de La Plata (UNLP).It operates a GroWeather Industry ® meteorological station (Davis Instruments CA).Both models take wind directions every 22.5˚ completing 360˚ of the compass (16 directions) with an accuracy of ±7˚ of the read out.The detection limit and resolution for wind velocities are 1.6 km•h -1 in both cases.The heights above the ground for the anemometers and weather cock were 12 m at Point A and 5 m at Point J.The distance from Point C (at the river bank see Figure 1) to Point A is around 9.8 km while to Point J is around 18 km.
Sets of simultaneous data at points A and J correspond to 1998-2003.Data at Point J covering the period 2004-2009 were also employed.The data set belonging to site A had a deficiency in NNE records.This drawback was found to be due to an obstacle that prevented a correct observation.Both monitoring sites provided complete data sets with the exception of Point J during winter 2000 which records were very poor; missing data were replaced by the median of the 4 adjacent years (the choice of this number of years was arrived at as a consequence between bias and variance); the same procedure applied to summer, autumn and spring yielded the smallest quadratic error compared with the average and the weighted average.Data at Point A were recorded every 15 minutes while data at Point J were recorded every hour.The difference in data quality is due to the fact that the institutions from which the data were obtained use them for different purposes; these sets do not conform a monitoring network.Throughout this paper hourly averages imply hourly blocks (for example, 00:00-00:59 hr is equivalent to "hour 0" local time).Regarding seasons, summer included December of the precedent year and January and February of the current year, whereas autumn included March, April and May, winter included June, July and August and spring included September, October and November.

Similarity Analysis
Seasonal hourly patterns of wind direction frequencies for both monitoring sites were compared by considering D E 2 (squared Euclidean Distance) and MCD (minimum covariance determinant).
The squared Euclidean distance is a metric that allows assessing proximity between two objects (vectors).
where x and y are in our case objects of dimension 24.
Recall that Pearson product-moment sample correlation coefficient "r" can be expressed as: where is the estimated covariance between variables x and y, and are the estimated variances of x and y, and Var y x μ and y μ are the estimates of their respective means.This statistic is widely used for summarizing the relationship between two variables or group of variables that define an object.The statistic r expresses the degree of association of two variables [21] and constitutes a standardized measure of linear dependence between them [22].A value of r close to 1 or -1, indicates that each of the variables can be accurately predicted by a linear function of the other.The sign indicates the direction of the relationship between the Y and the X.
Two drawbacks have been traditionally accounted for the application of r [23,24]: the sensitivity of this statistic to outliers due to the fact that classical average and covariance matrices are extremely sensitive to atypical observations [25] and the inability to recognize nonlinear relationships.Outliers may play the role of inflating or deflating the r estimate since there are "good" and "bad" leverage points [25].
In order to minimize the influence of outliers we employed the MCD correlation coefficient introduced by Rousseeuw [10,26,27] that considers robust alternatives for the location and scatter estimates given in Equation (2).
MCD computes average and scatter estimates regarding a subset of h of the n data (2 < h < n) which attain the smallest determinant of the covariance matrix.Then, the location and scatter MCD estimators are given by the average and covariance matrix of the optimal subset.MCD will provide an estimate of the strength of the correlations for the data of interest free of the influence of outliers, so that a low value of the estimate would indi-volved is poor.The fast-algorithm for computing MCD is complex and it is explained in detail in [28].MCD computations have been carried out by the use of the statisticcal software package SCOUT Version 1.0 from US EPA [29].MCD properties such as breakdown value, affine equivariance and influence function are described in [30,31] and [32].In order to have a wide coverage and supposed a 10% of contamination in our data, a value of h = 0.8 was chosen.Thereby the breakdown point (BDP) was about 20%, which is a value close to that recommended by [33] in order to have a balance between robustness and efficiency [25].
Distances and corr cate that the linear relationship between the objects in-elation coefficients are different appr

Win
2 at points A and J 1 and 2 was performed as a cases a nonparametric method based on local w oaches to measure similarity [34].Two objects can be highly correlated (MCD ≈ 1 or -1), but their distance could be high enough to consider them as different (e.g. they could have a significantly different mean).On the other hand, the two objects can be poorly correlated (MCD ≈ 0) but D E 2 could be very small and so they can both be considered as describing the same characteristics (although their differences may be due to different causes).Furthermore, according to [9] D E 2 is very sensitive to additive and proportional translations while MCD is wholly insensitive to them; finally, both estimates share sensitiveness to mirror images translations.Then, for the purpose of comparing wind frequency patterns measured at two monitoring sites both correlation and proximity were considered of interest"

Season on Sectors 1 and 2
d frequencies for Sectors 1 and are considered.In order to discriminate and quantify "daily" from "seasonal" variations within the series, an "average day" was estimated by averaging hourly the corresponding hours of the day for all the data.The averaged day was later substracted to each original series.The remaining curve had still the influence of the season.The seasons through the years under study were then averaged to obtain the "average season".The average season was later on substracted to the remaining curve (the curve resulting after the first subtraction) in order to obtain the residuals.Finally, the variances involved at the different steps of subtractions were considered in order to evaluate the degree of contribution of the "day" and the "season" to the total variation.
A trend analysis for Sectors prospective study in order to save the inexistence of simultaneous larger hourly data collections published in the area.
In both eighted regression, usually called "LOESS" or "LOW ESS" [35,36] was employed.For a sequence   x y the procedure computes at each given x within th e of e rang the i x 's a value   ˆŷ y x  as follows.Call I a window with pan h aroun ses to zero at x =   has no explicit form and does not belong to an rametric family.A small window span h yields a small bias but a possibly large variance, while the contrary happens with a large h; therefore h must be chosen to strike a balance between bias and variability.The nonparametric fit allows visualizing trends but it is important to discriminate whether it reveals actual data features or simply statistical artifacts.To this end the means corresponding to different time intervals were compared by estimating their standard deviations.Here it must be taken into account the lack of independence in the data.If 1 2 , , x x  is a stationary sequence with variance 2  σ and y given pa ted, where V is the "inf 23]: ; k ρ is the k-th order autocorrelation.An a he ata suggested that their dependence could be well represented by a first-order autoregressive process, with Finally, the estimated for er-each window as the square root of the variance for the mean.

Exposur
An empirical approach often used in air pollutant disp sion calculations is the power law velocity profile [38,39] given by: where: e wind velocity "corrected e wind velocity observed at a g z is w ht for the observed wind velocity.urface ro according to terrain roughness and atmospheric stability given by p;

Results and Discussion
the height that is desired to obtain the "corrected" ind velocity; r h is the heig The exponent p increases with increasing s ughness and increasing stability.Various researches reported p values between 0.07 and 0.60.Tables given in chapter 3 of [40] were used in order to select the most appropriate value for p at points A and J.In order to avoid differences due to the effect of altitude wind velocities were corrected considering a reference height of 10 m.Besides, differences in roughness (see Section 2.2) between points A and J, were saved by affecting each point with correction factors of p = 0.25 and p = 0.15 respectively.A neutral atmospheric stability class was considered due to the fact that the seasonal averages involve day and night.

Sea-Land Breeze Presen
Observed wind frequencies covering the compass at sites A and J for the four seasons are analyzed on hourly basis.Figure 2 shows only summer and winter patterns due to space constraints; autumn and spring displayed in most cases intermediate behaviors.Considering the lack of meteorological studies in the area the sea-land breeze phenomenon appears as the only significant source of local atmospheric variability.Its influence is more pronounced during summer due to the higher temperature contrast between the land and the large water surface of the La Plata River.For this reason summer is considered the leading season to carry out the analysis.
An ove r E (e.g. Figure 2(a1)), N (e.g. Figure 2(a13)) and NE (e.g. Figure 2(a15)) are high respect to the rest of the directions throughout the seasons in coincidence with observations carried out at La Plata Airport (Section 2.1).According to Barros et al. [41] these three wind directions, originated by the western flank of the subtropical anticyclone of the South Atlantic Ocean (located around 35˚S, 45˚W), are of major importance for the De La Plata River basin.
During the s for S (Figure 2(a5)) and SSW (Figure 2(a6)) are significantly higher than those of the rest of the day.This is attributable to land-breeze because these wind directions are somewhat perpendicular to the coastline.During the first morning hours these frequencies decrease notably.As far as the influence of southern winds diminishes, wind frequencies from N (Figure 2 Differences between points A and J for wind direction frequencies involving the land-breeze are smaller than those involving sea-breeze.A weaker land-breeze is ex l stability [42] but also the flow of air from -42].Sea-breeze winds follow a pected mainly due to the nocturna to the city roughness that inhibits rotational pattern [43] clockwise, previously detected by Borque et al. [44] in a preliminary study during one day of March that revealed that the circulation rotates from NE to E between noon and dusk.This effect is observed land to water.The wind direction spectrum observed for the land-breeze appears restricted respect to that for seabreeze.The inland penetration should be encouraged in future studies. in hour 16 on (Figure 2

Similarity Analysis for Wind Direction
Frequencies during the Period 1998-2003 The four seasons of the year show by inspection pro-ximate patterns when comparing both sites (Figure 2).Warm seasons showed in general more differences between patterns than cold ones, as can be seen for summer and winter in Figure 2.This is attributable to the sealand breeze cycle that is more intense in warm  1).This metric gives an overall estimate of the proximity between patterns but does not distinguish if the differences are concentrated in a few hours or distributed throughout the day lar hour of the day) are also discussed in order to sh e complete picture of the proximity approach.NE and NNE exhibit relative high distances throughout the seasons, often between one-and two-fold standard deviation from the mean (of D E 2 ) (Table 1).Recalling that NNE has been deficiently measured at site A and that wind direction frequencies were expressed as a percentage for a given hour, it is possible to consider that the distortion for this direction would mainly affect the adjacent ones i.e.NE and N. Regarding these three directions the maximum individual differences throughout the day were: 12.9% in summer for NE at hour  (cloc general proxi ate for ll the n alues are b low the genera ean) able 1 g of directions involves cold fronts, frontal waves and instability lines.Considering that the area under analysis is mainly flat, the land-breeze is weak and that the directions involved are not influenced by sea-breeze phenomenon both weather stations show proximate patterns regarding wind occurrences.
Table 2 shows the MCD estimates for the same patterns discussed above.An overview of this table shows that linear relationship between patterns are somewhat alternated As stated before (Section 1), short distances as well as strong linear correlations are expected to be found at two monitoring sites with common characteristics: low flat lands.Nevertheless Point A is located close to the river bank in an urban area with low buildings, while Point J is located more inland in a semi-rural area.As Ratto et al. [8] concluded, local winds are influenced mainly by sealand breeze circulations.This determines more variations in site A with respect to site J regarding some wind directions.This physical phenomenon would affect observations at both sites in a different way.The daily cycle of the sea-land breeze responds to the atmospheric pressure anomaly field induced by the cyclic thermal contrast at the surface [42].In addition to sea-land breeze circulations as a cause of differences between sites, differences in data quality, terrain roughness and instrument exposures explain, in general, the degree of non-similarity observed at both monitoring sites.
While the distance approach generally depicts good similarity between sites as concluded in [8] the correlation approach gives an uneven picture.As evidenced by correlation analysis for some wind directions and seasons one curve for a particular wind direction can not be predicted from the other.This finding should be considered when air pollutant measurements at any site within the city need to be correlated with individual wind direction frequency observations from the point of view of sites A or J.

Influence of the Day and the Season on Sectors 1 and 2
Average values for the occurrences of winds from Sectors 1 and 2 for the two periods under study are presented in Table 3.Note that considering together Sectors 1 and 2 winds towards exposed population (Figure 1) are occurring most of the time (above 50%).In order to gain knowledge on both sectors variations due to daily and annual cycle are considered.Figure 3 summarizes the analysis carried out for secring 1998 Due to space constraints this secnts A tor 2 during 1998-2003 (at Points A and J) and du -2009 (at Point J). tor was particularly selected to show the complete steps of analysis in this section.Sector 2 is highly correlated with SO 2 concentrations observed at a site downwind (Point D, see Figure 1), as was later demonstrated (see Section 3.4).
Figure 3(a) shows the evolution for the frequencies of Sector 2 from the point of view of poi . A partic r can be se izin nding sea in the order s tum d spring.are two ma ntribution be discr ted within the es: the flu of the "day" and t nce of the "season".
ular yea en by visualer, au-g its correspo sons umm n, winter an There in co s intended to ence imina he influe seri in Figure 3(b) was built by averaging the days of Figure 3(a) corresponding to Points A and J for the two periods under study.From now on, and for practical purposes, the sequence of the analysis is mainly concentrated in Point A.

Series Trend
In order to detect possible remaining patterns in the series, LOESS (see Section 2.3) was applied to the residuals.The smoothed line in Figure 3(d) is the result of the application of this method to   2 A S R t .Although no periodic pattern can be appreciated, a decreasing trend is implied at the end of the curve.To analyze features of th ean, the der auto al variability for the De La d estuary for different variables [41, is kind-that may appear in any of the residual graphswindows of 48 data were considered; then the m first or correlation coefficient and the deviation for the mean (see Section 2.3) for each window were computed (see Table 4).Since the differences between the means of consecutive windows are, in general, smaller than the deviations for the mean, there is no evidence of a significant trend for the series.Data from residuals of Sector 2 corresponding to Point J for 1998-2003 and 1998-2009 as well as the data of Sector 1 treated in the same way revealed the absence of a decreasing or an increasing trend.This result is in accordance with studies that analyze the interannu Plata River coast an 46,47].(a) in A 1998-2003  P int J 1998-2003  Point J

Wind Direction Frequencies and Air Pollutants
Figure 5 shows hourly occurrences for Sector 2 involving spring 2005 at site J, the spring average for the period 1998-2003 at site J, the spring average for the period 1998-2003 at site A and SO 2 concentrations (ppbv) observed during spring 2005 at Point D [12].The low values observed for SO 2 were attributed to the distance between industrial sources and the site the monitoring deariation and further compare its shape with different T alues obtained (Table 6) show an te t are higher than those observed at Point J; within the boundary layer friction forces decrease with height.With the application of Equation (3) differences between sites tend to be negligible.Although meteorological observations carried out at airports are not very appropriate for air pollution considerations [48] we took into account La Plata Airports' monthly vice was allocated.SO 2 concentrations are represented in a different scale in order to better visualize its hourly v cases of Sector 2. he high MCD v example of how observed averaged industrial pollutants detected at site D may depend linearly on winds from Sector 2.
An MCD estimate correlating the average of Sector 2 at Point A and the average of Sector 2 at Point J during 1998-2003 gives 0.795, a relatively high value when compared individually to the wind directions composing the sector (i.e.ENE, E and ESE) (Table 2).The same occurs when correlating averages for Sector 1 for the same period (MCD is 0.953).This behavior allows correlating air pollutant measurements carried out wherever within the city area with Sectors 1 and 2. No hat this is not possible when individual wind direction frequencies are involved (Section 3.2).

Wind Velocities and Calms
In order to provide context for the wind direction frequency analysis we summarized information regarding averaged wind velocities and calms.Table 7 shows observed wind velocities at points A and J for both periods under study and their corresponding corrected values estimated with Equation (3).Observed values at Point A

Conclusions
The hourly analys win obse ed a he tw allowed detect the prese nte culations in the co t of nop sca wind Sea b daytime while land-breeze takes place following a weaker pattern involving S and SW winds.
The two methods employed to assess similarity between sites allowed to gain insight of wind patterns.The distance approach showed in general good similarities.On the other hand, the robust correlation approach re- between patterns at both sites.This issue should be taken into account when correlations between air pollutant concentrations and individual wind direction frequencies ered.Detected differences in hourly erns between sites are mainly attribof the daily and an of the day was found to be more pronounced than the influence of the season for both sectors and sites, but at Point A (close to the river bank) the gap is more important due to the sea-land breeze effect.
The trend analysis employing LOESS with the further computing of the mean and deviate from the mean for different time intervals proved to be a sound approach.The analysis of the residuals for Sectors 1 and 2 series showed there was no remaining pattern when subtracting the effect of the day and the season.
Correlations between wind direction frequency aver an be correlated with winds from be need to be consid wind direction patt utable to the sea-land breeze phenomenon but also differences in terrain roughness, data quality and instrument exposures are to be taken into account.
The analysis of series corresponding to Sectors 1 and 2 allows a preliminary assessment of the probability of occurrences and the characterization nual cycles of both sectors.General mean for Sectors 1 and 2 observed at both sites are very similar.Winds from any of the two sectors may occur more than 50% of the time which is very important considering the transport of air pollutants towards exposed population.The influence -. ages for Sector 1 and 2 at both points A and J are strong This implies that air pollutants measured at any site within La Plata area c Sectors 1 and/or 2 observed at sites A or J.The robust correlation coefficient revealed, as an example, the linear character of dependence between winds from Sector 2 and sulfur dioxide concentrations.
Observed and corrected wind velocities showed a general agreement when comparing sites.Wind velocities observed at the airport (reference) were found to be around two times higher than those observed at sites A and J.This difference can be attributed to differences tween urban and rural climates but also in data quality and terrain roughness.Differences in calms between sites are attributed to differences in instrument exposure, data quality and roughness.

Figure 1 .
Figure 1.Map covering parts of La Plata and surroundings.Measurement points are indicated with a square and other reference points with a circle.Point A: National University of Technology.Point B: city center.Point C: river bank.Point D: Center of Optical Research (CIOp).Point H: center of the rectangle close to oil refinery plants.The rectangle (dotted lines) indicates the area with high industrial activity including a shipyard and steel plants.Point J: Agrometeorology Station.Sectors 1 and 2 are shown at the low left corner of the figure.Density of population is expressed qualitatively as two, one or half triangles depending on its degree.Dashed lines embrace main populated areas.
ce the 16 directions of rview of Figure 2 shows that wind frequencies fo night between hours 0 and 8 the frequencie

Figure 2 .
Figure 2. Accumulated averaged frequencies (on hourly basis) of winds observed at Point A and J in summer (from (a1) to (a16)) and winter (from (b1) to (b16)) covering all the directions of the compass.Y axis indicates the percent of occurrences for a particular direction and hour respect to all the occurrences for that hour (adding all the accumulated frequencies for a given hour and season it merges 100%).X axis indicates the hour of the "day".first stage of the sea-breeze development that occurs during the morning hours when winds from the river start to blow towards the land.Winds flow then increasing the northerly component [ (a13) and Figure 2(a14)) while ENE (Figure 2(a16)), E (Figure 2(a1)) and ESE (Figure 2(a2)) becomes dominant until they reach a peak during the evening (around hours 20 and 21).

Figure 3 .
Figure 3. Original serie for Sector 2, daily and seasonal influence analysis for Sector 2 from the point of view of points A esiduals for Sector 2 at Point A and the corresponding non parametric smoothing of the residuals.(a) is the percent of occurrences of winds from Sector 2 respect to the occurrences for all the directions of the co the period 1998-2003 (blues line).analogous but for point J and covers the period 1998-2009 (black line); for analysis purposes this series is divided into t o periods 1998-2003 and 1998-2009.Each individual point represents the frequency of winds from sector 2 (taken from the corresponding wind rose) for a given hour (t) and for a particular season and year.Values for t are identified each 24 data and are expressed in an abbreviated way, e.g.Sum 00 H0 represents the frequency for Hour 0 of summer 2000 corresponding to Sector 2. The whole data set has 576 points for Point A (covering six years) while 1152 points for Point J (covering 12 years); (b) The Y axis represents the percent of occurrences for the average day for Sector 2 from the point of view of points A (blues line) and J for the two periods under study (black line).It was built by averaging each accumulated hour through all seasons and years; (c) The Y axis represents e percent of occurrences for the average for the seasons; (d) Residuals for th ries of (a) at Point A. The smoothed line was obtained by the application of the locally weighted regression method.Vertical lines indicate the starting of the year.(1998-2003) and J (1998-2003; 1998-2009), r 2 ( ) A S Y t mpass covering

2 C 1 Table 5 .
procedure the residuals for Point J for the period 1998-2003 and for t od 1998-2009 were obtained.The rotocol (not shown due to space constraints) was c out to Sector 1 data (Figure4).To measure the variation removed by each of the two subtracting steps the variances were computed (see Table5(a)).For example, subtracting the variance of the remaining series from the variance of the original value of 83 (147.0) is obtained.This means that 56.5% of the original variance corresponds to the influence of the day (see 20.5 and represents only 13.9% (the percent of variance caused by the influence of the season) of the total variation (see influence of the season (IOS) for Sector 2 in Table 5(b)).Finally, the variance of the residuals represents 29.6% of the original variance and constitutes the unexplained fraction of the total variance (see lained variance (UNE) for Sector 2 ble 5(b)).An overview of Table 5(b) shows that Point A has more variation due to the day than Point J and this variation is slightly more pronounced for Sec-Sector Variances and percent of variances.(a) Variance for the original series, variance for the remaining series after the subtraction of the influence of the day and variance for the residuals according to the procedure explained in Section 2.4 (i) corresponds to the sector and period of the corresponding headings of the columns.(b) % of variation attributed to the influence of the day, the season and unexplained respect to the total variance of the original series.a % of variance attributable to the influence of the day; b % of variance attributable to the influence of the season; c % of unexplained variance.

Figure 4 .
Figure 4. Analogous to Figure 3(a) this figure represent the original series for Sector 1 from the point of view of points A covering 1998-2003 (green line) and J covering 1998-2009 (black line).

Figure 5 .
Figure 5. Left Y axis refers to the % occurrences of winds from Sector 2 (on hourly basis-X axis) observed at Point A and J for spring for different periods.Right Y axis refers to SO 2 concentration hourly averages observed during a short monitoring campaign at Point D during spring 2005.

Figure 6 .
Figure 6.Calm occurrences observed at sites of the city and surroundings.
1. Then for  

Table 7 . Averaged wind velocities (km•h -1 ) observed at points A (12 m height) and J (5 m height) and their corre- sponding corrected values according to Equation (3) (Sec- tion 2.5).
[20] during the decade 2001-2010[20]available from the National Meteorological Service to provide context for our measurements.Averaged velocities for 8 direction wind roses at the Airpor measured at 10 m above t sum n autumn, 13.4 -1 in winter, 15.0 k Figure1).These values are