Figure 1 shows the map of the sampling sites.

Before the measurement campaigns took place, a prospective visit was carried

out on the road sections during which 13 sites were identified for monitoring including 9 hot spots and 4 points in the background areas. Table 1 presents the list of ten (13) measurement sites, their location and their characteristics (heavy traffic lane paved or not).

Air quality measurements were taken using autonomous electronic devices called “Air Quality Monitor”, powered by a lithium battery or by connection to the mains, via a transformer. The device is calibrated for 4 hours at first start-up in ambient air. Once switched on, the measuring device automatically determines the concentrations of carbon dioxide CO₂, Total Organic Compounds (TVOC), methanal (CH₂O), particulate matter with diameters of 2.5 µm and 10 µm denoted PM2, 5 and PM₁₀, temperature in degrees Fahrenheit (˚F = 1˚C × 1.8 + 32) and the hygrometry (Relative humidity H%) of the ambient air. Recordings were carried out for an entire day at each measurement site in 30-minute intervals. Measurements generally start at 7 a.m. and end at 6 p.m. with 24 measurements per site. In Cases of delay, the measurement is extended beyond 6 p.m. to compensate for the delay. Units of measurement are parts per million air particles (ppm) or micrograms per cubic meter. These different parameters provide information, alone or combined, on the quality of ambient air.

The processing methods cover all the operations applied to the data to bring them as close as possible to the nominal terrain. These are descriptive statistics techniques which make it possible to achieve inferential statistics and spatial type processing.

Source: This study.

In this research [10] , multiple comparison method was used. The Wilcoxon test was used to determine the existence of a significant difference in the temporal evolution of the concentrations of the pollutants measured, at the different sampling points. Concerning the geometric transformation, it consisted of registering all the georeferenced data in the same projection system. To maintain legitimacy, all spatial data were included in the official cartographic repository of Benin: WGS 1984, Transverse Mercator Universe Projection (UTM), Zone 31 North. The Moran statistic was used to measure the levels of spatial autocorrelation of the variables studied and to test the significance.

The Moran statistic [11] is the spatial autocorrelation index used. It allows you to measure the level of spatial autocorrelation of a variable and to test its significance. It is equal to the ratio of the covariance between contiguous observations, defined by the spatial interaction matrix, to the total variance of the sample. The index has values between-1 (indicating perfect dispersion) to 1 (perfect correlation). A zero value means that the spatial distribution of the variable studied is perfectly random in the territory. Negative (positive) values of the index indicate negative (positive) spatial autocorrelation. An interpretation of the Moran index is given as follows:

· In the presence of spatial autocorrelation, we observe that the value of a variable for one observation is linked to the values of this same variable for neighboring observations.

· Spatial autocorrelation is positive when similar values of the variable being studied cluster together geographically.

· Spatial autocorrelation is negative when dissimilar values of the variable to be studied are grouped together geographically: close places are more different than distant places. This type of situation is generally found in the presence of spatial competition.

· In the absence of spatial autocorrelation, we can consider that the spatial distribution of observations is random.

Figure 2 gives an interpretation of the p-value and the z-score when processing spatial data.

In order to locate the spatial subsets, the local spatial association indices of Anselin Local Moran and Getis-Ord, Gi*, are used:

· Getis and Ord index

Getis and Ord [12] propose an indicator (Gi) making it possible to detect local spatial dependencies which do not appear in the global analysis. Thus, the Gi* from Getis-Ord makes it possible to analyze concentrations (hot pot). An interpretation of the values of this indicator is given as follows:

Gi > 0 indicates a grouping of values higher than average.

Gi < 0 indicates a grouping of values lower than the average.

· Local Anselin Moran index

Anselin [13] develops the notions introduced by Getis and Ord by defining local spatial autocorrelation indicators (Ii). These must make it possible to measure the intensity and significance of the local dependence between the value of a variable in a spatial unit and the values of this same variable in the surrounding spatial units. More precisely, these indicators make it possible to detect significant groupings of identical values around a particular location (aggregate) and to identify areas of spatial non-stationarity, which do not follow the overall process.

Ii > 0 indicates a grouping of similar values (higher or lower than average).

Ii < 0 indicates a grouping of dissimilar values (e.g. high values surrounded by low values).

To assess air quality, we used Table 2 which presents the guide values for assessing

Source: This study.

air quality based on the limit values of the World Health Organization (WHO) and regulations, Beninese. These values are defined by the manufacturer of the mobile air pollutant measurement monitor.

The map in Figure 3 shows the spatial variability of temperature in southern Benin.

From the analysis of Figure 3, we note that the temperatures vary from 20˚F to 95˚F with an average of 69.5˚F ± 21˚F and a %CV = 30. Therefore, the values taken by the temperature are dispersed over the study zone. The standard deviation ellipse is oriented East-West. We deduce from this the existence of a direction in the distribution of temperature values. This allows us to suspect a spatial dependence of ambient temperature values, hence the need to test spatial autocorrelation. Moran's Spatial Autocorrelation Index = 0.489992 at the 0.03 threshold. This index is different from 0, which leads us to say that the realization of the regionalized random variable is therefore dependent. It is therefore possible to predict the temperature values at one point, knowing the values observed at another point, based on the distance that separates them. This observed dependence may be the result of chance. It is therefore necessary to verify the validity of this spatial dependence on a global scale. This statistical test of independence of the regionalized variable carried out is General G which gives a Z score equal to 0.711827 and a p-value of 0.476572 greater than alpha = 0.05. This test puts the interpretation of the spatial autocorrelation test into perspective and allows us to conclude that the hypothesis of spatial dependence is not verified and is only the effect of chance.

Figure 4 shows the spatial distribution of the values taken by the relative humidity RH.

Analysis of the map in Figure 4 reveals, as in the Case of the temperature map, that the standard deviation ellipse is oriented West-East, raising suspicion of a direction in the spatial distribution of relative humidity data. which suggests a spatial dependence of the humidity data.

The spatial autocorrelation index gives a value = −0.382090 at the threshold of 0.298181 greater than alpha. Therefore, there is no autocorrelation between relative humidity values on a global scale. However, carrying out the General G test gives an index of 0.000023 read at the threshold of 0.37. It confirms the interpretation of the Moran index and allows us to conclude on the non-existence of aggregates in which local autocorrelation is expressed.

The map in Figure 5 shows the spatio-temporal variation of CO₂ levels. From

the analysis of the map data, it appears that the CO₂ contents vary between 400 and 1444 ppm with an average of 486.8 ± 184.33 ppm and a %CV = 37.86. CO₂ levels vary from one site to another. The high CO₂ emission zones are located at the major intersections of Cotonou such as the “Etoile rouge, Sica Toyota and Terrain” intersections.

The calculation of the spatial autocorrelation index carried out on the 10 collection points and the values observed on these points present the following results: Moran’s spatial autocorrelation index = −0.13; p-value = 0.91. The Moran index is close to 0. The realization of the regionalized random variable is therefore independent. This means that the CO₂ values at a collection point are not influenced by the values observed in neighboring points. It is therefore not possible to predict the CO₂ values at one point, knowing the values observed at another point, based on the distance that separates them. This observed independence may be the result of chance and only has value for the sample used for the calculation. It is therefore necessary to carry out a test to verify the validity of this spatial independence for the spatial variable in its entirety. This statistical test of independence of the regionalized variable gives a Z score equal to −0.10 at the threshold of 0.91. This threshold is below the critical threshold of 5%. We deduce from this a total absence of spatial autocorrelation over the entire variable. Given the z score of −0.10, the model does not appear to be significantly different from chance which supports the interpretation of the Moran index.

The map in Figure 6 shows the spatial distribution of PM_2.5 values.

From the analysis of the information from the map in Figure 6, we note that the standard deviation ellipse is oriented East-West leading to suspect a spatial dependence of the PM_2.5 pollution data.

The Moran index is different from 0 at the threshold of 0.012211. There is therefore a probability of less than 5% that this clustered pattern is the result of chance. To confirm or refute this trend, we searched for the General G index whose z-score is 3.06427722821, hence there is less than 1% probability that this highly clustered pattern is the result of chance. To confirm or refute this trend, we looked for the General G index whose value is 0.000049 read at the threshold of 0.002. Exploring this test at the local scale reveals the existence of aggregates in which local autocorrelation is expressed. It is therefore appropriate to identify these aggregates and look for the factors behind these concentrations of similar values at certain points. In order to locate these spatial subsets, the local spatial association indices of Anselin Local Moran and Gi* of Getis-Ord are used.

The calculation of the Anselin and Moran Local Index is carried out on the values taken by PM_2.5 in order to identify areas of similarity. Figure 7 and Figure 8 show the maps relating to the levels of similarity/dissimilarity between neighboring points and highlight areas with homogeneous or heterogeneous values.

The map in Figure 7 shows that we have categorized where the Anselin Local

Moran Index Z score values have been categorized into three classes. The low value class (−1.71 to −0.70) expresses a dissimilarity between neighboring points. The classes of average values (−0.69 to 1.14) expressing an average similarity and high values (1.15 - 2.57) giving the strong similarities between neighboring points. Analysis of the map information in Figure 7 reveals two clusters of points in the southwest and one cluster of points in the southeast. The first cluster in the southwest is a subset of points of low similarity and medium similarity. On the other hand, the 2^nd subset brings together points of strong similarity. The last subset is composed of medium and strong similarity points. The first subset and the 3^rd subset are therefore heterogeneous zones, which implies that the PM_2.5 values observed in these places are dissimilar. The hypothesis of the existence of a structure in the spatial distribution of PM_2.5 values is therefore verified. However, it is necessary to analyze the aggregates more closely. This is what the use of the Getis-Ord GI* index aims to achieve.

In order to identify areas of high concentration (hot pot) and areas of low concentrations of PM_2.5, we used the Gi* from Getis-Ord. Map 30 shows the spatial distribution of hot pot areas. The Getis-Ord GI* Z score presents the following statistics in the dry season for the 18 observation points: the minimum Z score is −1.83, its maximum 1.70 for an average of 0.64 and a standard deviation of 1.18; the accumulation of all the values being 22.98.

From the analysis of the map in Figure 8 it appears that high exposures to PM_2.5 are concentrated in the southwest of Benin and more precisely in the city of Cotonou and its surroundings and the further one moves away from Cotonou going towards the East, PM_2.5 pollution decreases. The least polluted area of ​​the study area is the south-eastern part of Benin.

The map in Figure 9 shows the spatial distribution of PM₁₀ levels in southern Benin. From the analysis of this Figure, we see that the direction ellipse is stretched on the East/West axis. It covers 6 observation points out of 13, or 46.15% of the PM₁₀ values observed. We deduce the existence of a direction in the distribution of PM₁₀ values. This allows us to suspect a spatial dependence of PM₁₀ values, hence the need to test spatial autocorrelation.

Data analysis reveals that the Moran index is different 0 at the threshold 0.000336 andthe z-score is 3.58581379254, hence, there is less than 1% probability that this clustered pattern is the result of chance. The General G index gives 0.000065 at the threshold of 0.000061 with a z-scoreof 4.00809924822. General G’s statistic confirms that of Moran, that is to say that there is less than 1% probability that this highly grouped model is the result of chance.It is therefore appropriate to identify these aggregates and look for the factors behind these concentrations of similar values at certain points. In order to locate these spatial subsets, the local spatial association indices of Anselin Local Moran and Gi* of Getis-Ord are used. The calculation of the Anselin and Moran Local Index is

carried out on the values taken by PM₁₀ in order to identify areas of similarity. Figure 10 and Figure 11 show the levels of similarity/dissimilarity between neighboring points and highlight areas with homogeneous or heterogeneous values.

The map in Figure 10 shows three classes of Z score values of the Anselin Local Moran index. The low value class (−1.47 to −0.12) expresses a dissimilarity between neighboring points. The classes of average values (−0.11 to 1.29) expressing an average similarity and high values (1.30 - 2.54) giving the strong similarities between neighboring points. Analysis of the information on map 32 reveals two clusters of points in the southwest and one cluster of points in the southeast. The first cluster in the southwest is a subset of low similarity points. On the other hand, the 2^nd subset brings together points of strong similarity. The last subset is composed of medium and strong similarity points. This area is therefore heterogeneous, which implies that the PM₁₀ values observed in this location are dissimilar. The hypothesis of the existence of a structure in the spatial distribution of PM₁₀ values is therefore verified. This observed distribution is therefore not the result of chance. However, it is necessary to analyze the aggregates more closely. This is what the use of the Getis-Ord GI* index aims to achieve.

In order to identify areas of high concentration (hot pot) and areas of low concentrations of PM₁₀, we used the Gi* from Getis-Ord. The map in Figure 11 presents the spatial distribution of hot pot zones obtained from the classification of Z GI* scores into three classes at the natural Jenks threshold and spatialization.

From the analysis of Figure 11, we note that the area of high pollution by PM₁₀ is the city of Cotonou and Abomey-Calavi.

The TCOV in the study area ranges from 0.01 to 6.91 mg/m³ with an average of 0.35 ± 0.65 mg/m³ and %CV = 181.38. From these remarks we can conclude that the values are dispersed around the average. The spatial distribution of TCOV contents is presented in Figure 12. From the analysis of this Figure, it appears that the standard deviation ellipse is oriented in the East-West axis, from where we deduce the existence of a direction in the distribution of TCOV values. This allows us to suspect a spatial dependence of the values of Total Volatile Organic Pollutants (TCOV), hence the need to test the spatial autocorrelation of the levels of this parameter.

From the analysis of the data, it appears that the Moran index is close to 0 at the threshold of 0.787762. So we deduce a total absence of spatial autocorrelation of TCOV contents. This total absence of autocorrelation may be the result of chance or due to the size of the sample used. To remove the doubt, we calculated the General G index which gives 0.000031 at the threshold of 0.110337 greater than alpha = 0.05. This statistic definitively confirms the trend observed with Moran’s statics and we can safely say that there is a total absence of autocorrelation

in TCOV contents.

HCHO contamination levels vary from 0.005 to 3 mg/m³ with an average of 0.054 ± 0.17.62 mg/m³ and a %CV = 320.90. The spatial distribution of aldehyde contents in the study area is presented by the map in Figure 13. From the analysis of this Figure, it appears that the standard deviation ellipse is stretched in the East-West direction. Then, the existence of a direction in the distribution of HCHO values can be considered and in turn suspect a spatial dependence of HCHO values, hence the need to test the spatial autocorrelation of the contents of this parameter.

From the analysis of the data, we note that the Moran index is close to O at the threshold of 0.218310 hence the absence of spatial autocorrelation of HCHO contents. This absence of spatial autocorrelation may be the result of chance or linked to the size of the sample used for the calculations. To remove the ambiguity, we calculated the General G index which gives 0.000031 and a p-value of 0.372789 greater than alpha. This statistic definitively removes the doubt and we can safely affirm that there is no spatial autocorrelation between the data observed for HCHO.

From the analysis of the information recorded on the map in Figure 14, we can note that the AQI values vary from 03 to 243 with an average of 39 ± 31.16 and a coefficient of variation of 85.06. The standard deviation ellipse follows an East-West direction which suggests an orientation in the spatial distribution of the values taken by the Air Quality Index (AQI) and by extension suspect an autocorrelation in the distribution of the values of the AQI.

Analysis of the data shows that the Moran index is close to 0 at the threshold 0.068299 greater than alpha = 0.05 therefore, there is no spatial autocorrelation between the values taken by the AQI. This absence of spatial autocorrelation may be the result of chance and/or linked to the size of the sample used for the calculations, hence the need to verify this observation by calculating the General G index which gives 0.000060 at the threshold of 0.003085. This statistic puts the finding of the Moran index into perspective and highlights the existence at the local level of an autocorrelation between the values taken by the AQI.It is therefore appropriate to identify these aggregates and look for the factors behind these concentrations of similar values at certain points. In order to locate these spatial subsets, the local spatial association indices of Anselin Local Moran and Gi* of Getis-Ord are used. The calculation of the Anselin and Moran Local Index is carried out on the values taken by the Air Quality Index (AQI) in order to identify areas of similarity. Figure 15 and Figure 16 show the levels of similarity/dissimilarity between neighboring points and highlight areas with homogeneous or heterogeneous values.

The map in Figure 15 shows three classes of Z score values of the Anselin Local Moran index. The low value class (−1.69 to −1.41) expresses a dissimilarity between neighboring points. The classes of medium values (−0.40 to 0.58) expressing an average similarity and high values (0.59 to 1.89) giving the strong similarities between neighboring points. Analysis of the information from map 15 reveals two clusters of points overall in the study area. The first cluster is a subset of points of low similarity and medium similarity. This area located in the southwest is therefore heterogeneous, which implies that the Air Quality Index (AQI) values observed in this area are dissimilar. On the other hand, the 2^nd subset located in the southeast brings together points of strong similarity. The hypothesis of the existence of a structure in the spatial distribution of AQI values is therefore verified. There are concentration aggregates of strong and weak values as well as an orientation in the spatial distribution of AQI values. This observed distribution is therefore not the result of chance. However, it is necessary to analyze the aggregates more closely. This is what the use of the Getis-Ord GI* index aims to achieve.

In order to identify areas of high concentration (hot pot) and areas of low AQI values, we used Getis-Ord’s Gi*. Map III.14 presents the spatial distribution of hot pot zones obtained from the classification of Z GI* scores into three classes at the natural Jenks threshold and spatialization.

From the analysis of the information on the map in Figure 16, it appears that the air quality index decreases from West to East, indicating that air quality is more degraded in the West than to the East.

In this research, Principal Component Analysis was used to identify the different existing links between air quality monitoring parameters on the one hand and between air quality monitoring parameters on the other hand, air and tracking points. The correlation matrix in Table 3 reveals how the monitored parameters are correlated.

From the analysis of the information in Table 3, it appears that CO₂ is positively correlated with relative humidity at the 5% threshold. Thus, in conditions of high relative humidity, exposure to CO₂ becomes more significant with its corollaries of tearing headaches and suffocation in humans. This absence of links between the CO₂ levels and the levels of other ambient air quality monitoring parameters in southern Benin leads to the conclusion that the sources of CO₂ input into the ambient air at the measurement sites are not not exclusively road traffic. TCOV are significantly correlated with PM₁₀, temperature and the Air Quality Index at the 1% threshold and with PM_2.5 at the 5% threshold. In other words, the unburnt products coming out of chimneys and/or exhaust pipes impact the ambient air temperature due to their very high emission temperature. Likewise, these volatile organic compounds can be emitted in particulate form and/or adsorbed on inorganic particles which they coat. Aldehydes are not

*Significant correlation at the 5% threshold; **Significant correlation at the 1% threshold.

correlated with any of our ambient air quality monitoring parameters. This observation leads us to assume that aldehydes and TVOCs do not have the same emission sources. In other words, the aldehydes detected in the study presence, unlike TCOV which have a pyrolytic source, would come from other non-pyrolytic sources such as forests, industrial oxygenated organic solvents and others. Table 4 presents the weights of the axes in the interpretation of the results of the Principal Component Analysis (PCA).

From the analysis of the data in this table, it appears that the first axis alone allows more than 33% of the information to be interpreted, while axis 2 only allows access to 17.02% of the information. The cumulative percentage gives for the first two axes a cumulative percentage of more than 50% interpretation of the analyzed data. By using the three axes we have access to more than 63% of the information compared to 75.7% using the first four axes, which is more than sufficient in the Case of our study. Figures 17-19 present the graphical representations of the factor maps.

From the analysis of the information in Figure 17, we note that the AQI, PM_2.5 and PM₁₀ are well represented on axis 1, while CO₂, relative humidity and temperature are carried by the axis 2. On this axis the temperature is diametrically opposite to relative humidity and CO₂.At the level of Figure 17, as at the level of Figure 18, axis 3 carries the temperature and TCOV.

In Figure 19, we note that the aldehydes are mainly carried by axis 4.

Figure 20 shows the graphical representation of the factorial map F1-F2 and F1-F3.

The analysis of Figure 20 shows two groups of points, namely those located to the left of the ordinate axis and grouped around the abscissa axis. These are the points Aglangandan, Zogbo, Athiémé, Bopa, Ouiga characterized by a good Air Quality Index. All other points located to the right of the y-axis and grouped around the x-axis are characterized by moderate to severe Ambient Air Quality Indices. These are Sèmè, Ekpè, Tankpè, Togoudon Tori-Bossito, Carrefour Etoile rouge, Carrefour Toyota and crossroads after friendship stadium. At these last points, the quality of ambient air is marked by high levels of particulate matter, especially PM_2.5, TCOV, sometimes HCHO and CO₂. Particulate matter and TVOCs are two classes of air pollutants most harmful to human health.