Inland O₃ Production Due to Nitrogen Dioxide Transport Downwind a Coastal Urban Area: A Neural Network Assessment

Abstract

The tropospheric production of O₃ is complex, depending on nitrogen oxides (NO_x = NO + NO₂), volatile organic compounds (VOCs), and solar radiation. We present a case study showing that the O₃ concentration is higher in a rural area, 14 km downwind from a coastal town in Central Italy, compared with the urban environment. The hypothesis is that the O₃ measured inland results from the photochemical processes occuring in air masses originating at the urban site, which is richer in NO_x emissions, during their transport inland.To demonstrate this hypothesis, a feed forward neural network (FFNN) is used to model the O₃ measured at the rural site, comparing the modeled O₃ and the measured O₃ in different scenarios, which include both input parameters related to local O₃ production by photochemistry and input parameters associated with regional transport of O₃ precursors. The simulation results show that the local NO_x concentration is not a good input to model the observed O₃ (R = 0.17); on the contrary including the wind speed and direction as input of the FFNN model, the modelled O₃ is well correlated with that measured O₃ (R = 0.82).

1. Introduction

In recent decades, urban air quality has become a central focus of environmental research, policy making, and public health advocacy. Among the myriad of pollutants affecting the urban atmospheres, O₃ stands out due to its adverse effects on respiratory health, agricultural productivity, and ecosystem integrity [1,2]. The formation of the tropospheric O₃ involves complex processes: the stratosphere–troposphere air masses exchange [3] and photochemical reactions. These reactions involve the oxidation of carbon monoxide (CO), methane (CH₄), and volatile organic compounds (VOCs), catalyzed by nitrogen oxides (NO_x), which is the sum of nitric oxide (NO) and nitrogen dioxide (NO₂), as well as meteorological parameters such as temperature [4]. In urban areas, daily variations in O₃ concentration and its precursors are correlated to rush hours (which cause an increase in road traffic emissions) and to the diurnal patterns of solar radiation and other weather factors affecting the efficiency of photochemical reactions. Besides this, O₃ and NO_x concentrations are also influenced by the planetary boundary layer variations and advective transport processes, which is affected by the dominant wind direction and regional orography [5]. The O₃ precursors, such as organo nitrates, can be transported over long distances due to their longer lifetime, becoming possible sources of O₃ in a region far from their production sites [6,7]. It is useful to monitor the air quality and meteorological features with more detail in urban area and background countrysides to evaluate the influence of urban emission sources on the surrounding rural environment, where the local emissions of O₃ precursors are negligible [8]. In North-Eastern India, the yearly O₃ and NO_x concentrations have been monitored in three urban sites (Gauhati, Tezpur, and Aizwal), not affected by industrial emissions. It has been found that surface O₃ is not entirely due to local production and that during pre-monsoon and winter seasons, the transport of air masses from industrial areas is the cause of the higher pollutions level registered in the three monitoring sites [9]. Studies conducted in the Los Angeles Basin, a region notorious for its smog and air quality challenges, have provided critical insights into these dynamics. Notably, research by [10,11,12] has shed light on the intricate balance of chemical precursors in O₃ formation and the pivotal role of temperature in modulating this balance. The Los Angeles Basin serves as an emblematic case study for understanding O₃ trends and the efficacy of air quality management strategies. Despite significant reductions in NO_x and VOC emissions due to stringent regulatory measures, O₃ levels in the basin remain stubbornly high, often exceeding national ambient air quality standards. The decline in NO_x observed over the last decade in Los Angeles, which has not been followed by a reduction in O₃, has also been observed in several other urban areas worldwide, including the urban area of Pescara, Central Italy—the object of the present study [13]. The persistent O₃ exceedances highlight the non-linear relationship between O₃ formation and its precursors, underscoring the importance of a nuanced approach to emission control. Furthermore, studies underscore the impact of temperature, with warmer conditions exacerbating O₃ formation through enhanced photochemical reactions and increased emissions of temperature-dependent VOCs. In this work, we used a combination of the observations of O₃ and NO_x, and meteorological parameters in two sites—one in the urban area and the other inland and downwind of the coastal town—and a neural network technique to assess the origin of elevated O₃ concentration in the rural area compared with that in the urban site. The complex orography of the area and the emission mix due to domestic and industrial activities and those due to the presence of an international airport in the urban area, make this study interesting for strategies to mitigate O₃ pollution and protect public health in a different geographical context.

2. Materials and Methods

2.1. Site and Observations

The dataset used in this study was collected in Central Italy (Abruzzo). This region is characterized by a peculiar orography with mountainous, hilly, and coastal areas. The highest peaks of the Apennines mountain chain are located in this territory: the Gran Sasso massif (2912 m a.s.l.) in the west/north-west (about 60 km from the Adriatic Coast) and the Mount Maiella (2793 m a.s.l.) in the south-west (about 35 km from the Adriatic Coast) (Figure 1).

As a consequence, the meteorology of this region is largely affected by the local orography [13]. The coastal area of this area in Central Italy, is characterized by a humid climate with a warm summer and mild winter. Also, the NO_x and the O₃ annual trends show the expected features with anti correlated trends in summer and winter. Pescara (42.27 N, 14.15 E) is the most populated town along the Adriatic Coast, with about 300.000 inhabitants (from the Italian national statistic institute—ISTAT) in the whole metropolitan area. Moreover, the Abruzzo international Airport and one of the busiest ports of Central Italy are also located in Pescara. Details about the measurement sites, the local meteorological/climatological conditions, and air quality situation can be found elsewhere [13]. Pollutants such as O₃ and NO_x (NO₂ and NO) have been detected simultaneously in the two monitoring stations: Sacco (henceforth, Urban Site (US)) and Cepagatti (henceforth, Rural downwind Site (RS)). The US station is located along the coastline and is in the urban area close to the airport and the main industrial area. On the contrary, the second station (RS) is installed in a rural background site about 14 km from the urban area in the Pescara River Valley (Figure 1).

The data were collected in 2018, with a sampling frequency of 1 h. The whole dataset was used to compare the O₃ measured at RS in different meteorological conditions, i.e., for air masses from different wind directions. Due to gaps in the dataset when running FFNN, we selected the period between the end of April and the beginning of July, corresponding to the interval during which the simultaneous availability of the different parameters used as FFNN input was maximized.

2.2. Data Analysis

Figure 2 shows the time series of NO, NO₂, NO_x, and O₃, and the meteorological parameters recorded in May–June 2018 in both the rural and urban monitoring stations, in Figure 2A and Figure 2B respectively.

As depicted, NO_x had higher concentrations in the urban station compared with the rural site; on average, it was 22.0 ± 11.9 µg/m³ in the urban environment, whereas it was 7.4 ± 3.3 µg/m³ in the rural area. This significant difference between the urban and rural environments also applies to the maximum values of NO_x, which were more than twice as high in the urban site compared with the rural one, as well as for NO and NO₂, see

Table 1. Basilar statistics of NO, NO₂, NO_x, and O₃ measured between 1 May and 29 June 2018 at the urban monitoring station (US) and at the rural downwind station (RS).

		Urban (μgr/m3)	Rural (μgr/m3)
O3	Max	141.0	129.5
	Min	2.8	11.4
	Mean ± Std	40.4 ± 29.6	74.1 ± 24.8
NO	Max	38.2	10.1
	Min	0.0	0.0
	Mean ± Std	6.0 ± 4.4	1.3 ± 1.5
NO2	Max	62.2	33.3
	Min	0.0	0.0
	Mean ± Std	13.4 ± 8.6	5.4 ± 3.4
NOx	Max	88.8	32.1
	Min	0.0	0.0
	Mean ± Std	22.0 ± 11.9	7.4 ± 3.3

for a detailed comparison between the two environments. Interestingly, the mean O₃ concentration was higher in the inland rural station, at about 34 µg/m³ (Table 1), suggesting the possibility of O₃ production in rural areas due to photochemical processes during the transport of air masses from the urban coastal sites to the background station by regional transport.

2.3. Model Analysis

A feed-forward neural network (FFNN) was adopted to accurately predict the O₃ concentrations at the RS site, diverging from the recurrent neural network (RNN) models that are traditionally prevalent in this field [13,14]. This strategic choice was influenced by our specific prediction objectives and the rich, multidimensional dataset, rather than by the temporal nature of the data. FFNN networks are renowned for their capacity to discern complex patterns within a diverse array of data, a feature that proves useful in our context, where the aim is to integrate chemical and meteorological inputs from a range of monitoring stations [15,16,17]. The architecture of FFNN networks, with their unidirectional data flow from input to output, is exceptionally effective at identifying intricate, static patterns in data. Our FFNN model is specifically tailored to process variables such as nitrogen oxides (NO_x, NO₂, and NO) and various meteorological factors from the RS and US stations, to model the O₃ levels at the RS location. The predictive model was implemented in Python 3.10, utilizing the TensorFlow Keras library (https://www.tensorflow.org ver. 2.16) for its extensive deep learning functionalities [18]. The input neurons, as depicted in Figure 3, included NO, NO₂, NO_x, wind speed (WS), wind direction (WD) relative humidity (RH), and temperature (T) measured at the rural site. The model performance was critically evaluated using statistical metrics, including Pearson’s linear correlation index (R), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE).

Pearson’s R measures the linear association between the predicted and observed O₃ values, while RMSE quantifies the model’s prediction errors. MAPE offers insight into the model accuracy by expressing forecast errors as a percentage of the actual values. These metrics collectively guide the iterative refinement of the model to ensure the most accurate O₃ concentration predictions.

The validation process of the FFNN model involved splitting the dataset into training and testing subsets, ensuring that the model performance could be evaluated on unseen data. The training set was used to train the model, while the testing set was used to evaluate its predictive capability. The data were also standardized to improve the model performance. For each scenario, the model was trained and tested multiple times, and the best-performing model was selected based on the highest Pearson correlation coefficient. The scenario with the best performance included input parameters such as nitrogen oxides (NO, NO₂, NO_x), wind speed, wind direction, relative humidity, and temperature, achieving a high correlation coefficient (R) of 0.90. This indicates that the model performed exceptionally well in capturing the relationship between these variables and the O₃ levels. The model’s performance was also evaluated using external datasets by testing on data that were not part of the training set, ensuring that the model’s predictions were generalizable and reliable across different data samples. This comprehensive approach to validation demonstrated that the model could accurately predict O₃ concentrations in various scenarios, reinforcing its robustness and applicability in real-world settings.

Quality check and preprocessing of the dataset is crucial to provide accurate and reliable inputs to the FFNN model. As a first step, the missing or inconsistent data have been removed for the whole dataset, imputing missing values using statistical methods like mean or median where necessary. The outliers were identified using statistical techniques, such as standard deviation, and then either corrected or removed based on the domain knowledge. Finally, for temporal data, we ensured the alignment of time series data from different sources, verifying that those observations corresponded correctly over time.

3. Results and Discussion

The local O₃ concentration was due to two main mechanisms that determined it: The local photochemistry and the regional transport. We investigated the O₃ production mechanism in a rural background site, considering both the local photochemistry and the regional transport from a metropolitan area, where the urban station is installed (Table 1 and Figure 1). In order to evaluate the relationship between O₃ and NO₂ and NO, we compared the diurnal trends of these compounds and NO_x in both stations (Figure 4C).

Even if the mutual interactions in the troposphere between O₃, NO₂, and NO are complex and other chemical and deposition processes should be considered, from Figure 4A, we can identify that, remarkably, in the rural background station, the concentration of O₃ is, on average, higher than the O₃ levels detected in the urban station. This can be explained as follows: Considering that the regional transport could move air masses, richer in O₃ and its precursors (i.e., NO_x), from the densely populated coastal area (US) to the inland hilly areas (RS). The relatively high NO concentration, measured even during the night at the US station (Figure 4B), explains the lower O₃ measured in this site (Figure 4A) that, by reacting with NO, produces nocturnal NO₂ (Figure 4D). The higher NO concentration in the US can be explained by considering the position of this station, which is installed in a heavy traffic and industrial area of the town. On the other hand, at the background RS station, the NO (Figure 4B) shows a low concentration with a small peak at around 08:00 a.m., corresponding to the morning rush hour. The NO_x is about five times lower than those measured in the US station, with a typical diurnal trend with the daytime and late afternoon peaks related to vehicle traffic emissions (Figure 4C). Because of the low NO concentration in RS, the O₃ titration between midnight and 7 AM is significantly less efficient in RS than in US, and the O₃ concentration in this interval is about double than those sampled in US. At the same time, the local NO₂ and NO at RS is too low to explain the local O₃ level, suggesting that its origin should be due to its precursors transport from the US area to the background site (RS). These findings are confirmed by analysing the wind speed measured in the US and RS stations and the relationship between the O₃ measured at the rural downwind site and the wind direction (Figure 5).

Because of the wind speed of the air masses reported at the US (Figure 5A), it is possible to estimate that the NO_x produced in the US station by traffic and industrial emissions travelled for about 3 to 7 h to reach the rural station. Despite the chemistry of NO_x and O₃ being complex, the quick NO conversion to NO₂ and the resulting O₃ production is a well-known mechanism for explaining O₃ production during the transport process of its precursors, i.e., NO₂, which can cause a higher concentration of O₃ in the rural region, e.g., lacking local sources of O₃ precursors, downwind of an urban/industrial area. The time to transport air masses from the US to RS is in line with the NO_x lifetime expected in the troposphere. Figure 5B) shows the relation between the O₃ measured at the RS station and the NO₂ to NO_x ratio measured at the urban area, where the main emissions of O₃ precursors could trigger the high O₃ concentration sampled in the rural site with very low local emission of NO. As expected, when the NO₂/NO_x ratio is low in the US, meaning that the NO emission is more significant (as also shown in Figure 4B,D), the O₃ in the RS is higher, confirming the possible mechanism of O₃ production at RS from NO–NO₂ photochemistry during the air masses transport from US to RS. Figure 5C shows the polar plot of the O₃ measured at the RS station. The highest values of O₃ occurs when the winds blow from North-East, i.e., from the urban site with higher NO_x. On the contrary, with winds coming from West (inland mountains), the O₃ concentration is much lower, confirming that in inland regions, the origin of O₃ is not related to local emissions, but due to transport from coastal areas with higher concentrations of NO_x and O₃. Figure 6 shows an intercomparison between the O₃ measured at the RS during the period investigated in this case study and the O₃ measured at the same site during the 2018 selecting only air masses from the West, i.e., from the hill and mountainous areas of this region characterized by cleaner air. It is evident from Figure 6 that the O₃ at the RS is strongly affected by the origin of the air masses and, in detail, by the NO and NO₂ emitted at the US, which is located east of the RS site.

To confirm these results, an FF neural network was employed to identify the mechanisms based on the high O₃ level measured at the rural downwind site. Different scenarios were simulated (

Table 2. Results of the several FFNN model scenarios run to predict the O₃ measured at the rural site (RS). The input parameters used in the different scenarios have been listed, as well as the main statistical parameters to determine the goodness of the model results. Finally, the slope of the scatter plot between the measured and modelled O₃ is also reported.

Scenario	INPUT PARAMETERS	R	RMSE	MAPE	SLOPE
1	NOx, NO, NO2, WS, WD, T, HR	0.90	11.53	0.14	0.77
2	NOx, NO, NO2, WS, WD	0.81	15.18	0.19	0.63
3	NOx, NO, NO2	0.17	25.23	0.36	0.02
4	WS, WD	0.81	15.24	0.19	0.60

) in order to identify whether the O₃ concentration measured at the rural site was more important than the local chemistry, i.e., the local emission of NO and NO₂, or the O₃ production during the regional transport from the urban site where the O₃ precursor concentration, i.e., NO and NO₂, was significantly higher than the one measured at the RS. The O₃ at the rural site was modeled using different parameters as the FFNN model inputs, depending on the simulation scenarios (Table 2).

The selection of specific parameters for each scenario in Table 2 was guided by a scientific understanding of the mechanism behind the high O₃ concentration in the RS compared with the one measured at the US. In order to have a reference for the model performance, we included all of the parameters available to simulate the O₃ (Scenario 1). To exclude the local production of O₃ due to temperature and RH, strictly related to solar radiation, we ran the model including only the chemical compounds and the wind speed and direction (Scenario 2). To exclude the local production of O₃ due to the local NO and NO₂ emissions, we ran an FFNN including only these measurements (Scenario 3). Finally, to prove our hypothesis that the O₃ at the RS was the result of the transport of NO and NO₂ from the metropolitan area, we ran the FFNN model only considering the wind speed and direction as the inputs (Scenario 4). All of the statistical parameters for determining the goodness of the model results between Scenario 2 and Scenario 4 were similar, demonstrating that the O₃ at the RS can be explained mainly by its precursors transport from the US.

As expected, the best simulation is the one corresponding to the first scenario (Figure 7), in which NO_x, NO, NO₂, WS, WD, T, and HR were used as the inputs for the FFNN model (R = 0.90).

Furthermore, to understand the causes behind the higher O₃ measured at the rural site, it is interesting to observe that, when only the NO, NO₂, and NO_x (Scenario 3) measured at Rthe S are used as the inputs, the FFNN model is not able to reproduce the O₃ both in term of absolute value and diurnal trend (Figure 8), with a correlation coefficient of R = 0.17.

On the other hand, if these inputs also included the wind speed and direction (Scenario 2), the correlation coefficient significantly increased, reaching 0.81 (Figure 9).

This result strongly suggests that wind speed and direction are the most relevant inputs to model the observed O₃. It should be observed that in this scenario, despite the temperature, a well-known proxy to model O₃, was not included as an input parameter, the correlation coefficient between the model and measured O₃ at the RS station was very high and comparable to the one obtained for the first scenario. This suggests that transport is a possible explanation for the higher concentration of O₃ inland compared with the one measured in the urban area. Finally, the FFNN was run using only the wind speed and direction as the input (Scenario 4, Figure 10); the correlation coefficient between the measured and modelled O₃ was about 0.81, close to the one obtained when including NO, NO₂, and NO_x (Scenario 3).

This confirms that the high O₃ level measured at the rural downwind site was not related to the local production of O₃ precursor species (NO and NO₂) and their local photochemistry, but it was mainly due to the photochemical processes that took place in the air masses emitted at the urban site during the downwind transport. It is interesting to observe that the model underestimated the measurements especially between 12:00 and 17:00 p.m., i.e., at the maximum O₃ concentration. This could be related to some missing chemical compounds relevant to O₃ formation, such as volatile organic compound (VOC) oxidation products or organo nitrates (i.e., peroxy acetyl nitrate (PAN)) that could be dissociated back into NO₂, which were not included in the model as they had not been measured.

4. Conclusions

Higher O₃ concentrations weremeasured at an inland rural site and were compared with the nearest coastal urban site in Central Italy. We found that the NO_x directly emitted in the urban environment during transport resulted in higher levels of O₃ produded about 14 km downwind from the NO_x source areas. An FFNN model was employed to identify the mechanism to explain O₃ production in the rural environment. It was found that the local NO_x, which was significantly lower than the one measured in urban area, was totally irrelevant for modelling the local O₃ at the rural downwind station (RS). On the other hand, including the wind speed and direction as the input for the model, which were proxies of the regional transport, the correlation between the measured and modelled O₃ in the rural site increased significantly (from R = 0.07 using only NO, NO₂, and NO_x to R = 0.82 by also adding the wind speed and direction). These results are another piece of evidence that the surface O₃ formation is rather complicated especially in complex orography where the area downwind of an urban coastal area can be strongly affected by emissions from far away with implications on policy to control and mitigate air quality. The two main limitations of this analysis were the length of the analyzed dataset, which only focused on 2018, and the absence of VOC measurements in both the rural and urban monitoring stations. As a consequence, our work will need further research and field campaigns to analyse the impact of VOCs and organo nitrates to fully explain the mechanisms behind the O₃ chemistry in this complex mountain–hill–coast region.