Decoding Urban Traffic Pollution: Insights on Trends, Patterns, and Meteorological Influences for Policy Action in Bucharest, Romania

Abstract

Traffic-related pollutants remain a challenging global issue, with significant policy implications. Within the European Union, Romania has the highest yearly societal cost per capita due to air pollution, which kills 29,000 Romanians every year, whereas the health and economic costs are also significant. In this context, municipal authorities in the country, particularly in high-density areas, should place a strong focus on mitigating air pollution. In particular, the capital city, Bucharest, ranks among the most congested cities in the world while registering the highest pollution index in Romania, with traffic pollution responsible for two-thirds of its air pollution. Consequently, studies that assess and model pollution trends are paramount to inform local policy-making processes and assist pollution-mitigation efforts. In this paper, a generalized additive modeling (GAM) framework is employed to model hourly concentrations of nitrogen dioxide (NO₂), i.e., a relevant traffic-pollution proxy, at a busy urban traffic location in central Bucharest, Romania. All models are developed on a wide, fine-granularity dataset spanning January 2017–December 2022 and include extensive meteorological covariates. Model robustness is assured by switching between the generalized additive model (GAM) framework and the generalized additive mixed model (GAMM) framework when the residual autoregressive process needs to be specifically acknowledged. Results indicate that trend GAMs explain a large amount of the hourly variation in traffic pollution. Furthermore, meteorological factors contribute to increasing the models’ explanation power, with wind direction, relative humidity, and the interaction between wind speed and the atmospheric pressure emerging as important mitigators for NO₂ concentrations in Bucharest. The results of this study can be valuable in assisting local authorities to take proactive measures for traffic pollution control in the capital city of Romania.

1. Introduction

Pollution is the leading environmental cause of disease and premature death in the world today, and no country is immune [1]. The World Health Organization defines air pollution as “the presence of one or more contaminants in the atmosphere, such as dust, fumes, gas, mist, odour, smoke or vapor, in quantities and duration that can be injurious to human health” and identifies several air pollutants as leading to the strongest public health concern: carbon monoxide (CO), particulate matter (PM), ozone (O₃), nitrogen dioxide (NO₂), and sulfur dioxide (SO₂) [2]. Air pollutants belong to two main types based on composition—gaseous compounds and particulate matter—while another taxonomy considers them as primary (directly emitted, such as CO or SO₂) and secondary (formed by chemical and physical reactions in the atmosphere, such as ozone, sulfates, NO₂, and PM) [3].

The State of Global Air report revealed that in 2021 the number of deaths caused by air pollution was at 8.1 million people worldwide, of which more than 700,000 were deaths of children under five years old. This places it second among causes of death, after high blood pressure and smoking, but before tobacco and diet [4]. Moreover, outdoor air pollution caused nearly twice as many deaths in 2019 than in 1990 (4.51 million vs. 2.25 million), but indoor air pollution caused half as many deaths (4.36 million vs. 2.31 million) (GBD, 2019). This means that the nature of pollution has changed over time [1], and while most countries have made progress in terms of indoor air pollution [5], outdoor air pollution’s impact on mortality has been exacerbated by an ageing and growing population [6,7].

The pervasive impact of pollution on human health is well documented in the literature, resulting in the now widely recognized contribution of outdoor air pollution to respiratory and cardiovascular diseases [8], kidney diseases [9], autoimmune diseases [10,11], allergic diseases [12,13], obesity [14], and cognitive functions [15]. A range of respiratory and cardiovascular diseases, as well as malignant neoplasms, that are associated with NO₂ and NO_x traffic-related emissions, are also mentioned [16]. For ozone, another considerable pollutant, its concentrations, when high enough, can have negative effects such as eye and nose irritation, respiratory disease emergencies, and lung function impairment [17]. Worryingly, the literature reports that pollution has a negative impact on children’s development, particularly in areas most vulnerable to pollution sources such as fossil fuel combustion and traffic [18,19]. Unfortunately, most countries’ populations are exposed to levels of air pollution above WHO guidelines, with notable exceptions including the United States, Canada, Australia, Portugal, Finland, Sweden, Norway, France, the United Kingdom, Switzerland, Iceland, Ireland, Brazil, Uruguay, Sri Lanka, and Japan [20]. More recently, ref. [21] estimated that 97.6% and 47.5% of Europeans, respectively, were exposed to PM_2.5 concentration levels above the new annual WHO and EU threshold levels, respectively.

Besides health costs, other social costs (loss of working days and reduced life expectancy) need to be considered when assessing the overall effects of air pollution. Ref. [22] highlighted in their systematic review and meta-analysis that air pollution causes significant losses in productivity and social well-being, particularly through reduced life expectancy, as well as increased school absenteeism. These, in turn, are reflected in significant economic losses that are additional to the direct healthcare expenses for treating pollution-related illnesses [23,24]. From this perspective, a CE Delft report calculated that the damage cost per capita of diesel emissions in 2018 ranged between 3004 euro (Bucharest, Romania) and 382 euro (Santa Cruz de Tenerife, Spain) in a sample of 432 European cities, while the share of damage costs in GDP varied between 9.9% (Ruse, Bulgaria) and 1.3% (Kuopio, Finland) [25].

Vehicle emissions, which have become a widespread source of pollution in higher-income countries, contribute significantly to air quality problems, resulting in photochemical pollution, or “smog”, which has a considerable influence on major urbanizations and megacities [26]. According to existing studies, motor vehicle emissions account for around a quarter of worldwide energy-related greenhouse gas emissions and cause significant air pollution, primarily in urban areas [27]. However, research on traffic-related air pollution (TRAP) and its impact on health has lagged behind that on ambient air pollution, owing to problems in quantifying highly variable individual exposure from these ubiquitous yet highly localized sources of air pollution [28].

While urban air quality encompasses a broad suite of pollutants from diverse sources, including residential heating, industrial activity, and long-range atmospheric transport, the present analysis focuses specifically on traffic-related air pollution. To operationalize this focus, we center the investigation on nitrogen dioxide (NO₂), a pollutant widely regarded as a robust tracer of vehicular emissions in European cities. By targeting an air quality monitoring site located at a major urban traffic artery in central Bucharest, the study is designed to capture the patterns and drivers of NO₂ attributable primarily to road traffic, while acknowledging that other urban sources may contribute under certain meteorological conditions.

The above considerations are strong motivators for the current research endeavor. This paper thus aims to develop generalized additive models (GAMs) to assess trends in the hourly concentration of NO₂ at a central traffic location in the capital city of Romania, i.e., Bucharest. These models are highly used in the literature to reveal the intertwined effects of traffic emissions and meteorological variability [29]. Moreover, their flexibility allows for the incorporation of non-linear relationships between predictor variables and air pollutant concentrations, thus making them very suitable for seizing the intricated dynamics of air quality systems in urban locations [30].

The main goals of the research are to investigate recent trends and patterns of this pollutant, to develop models able to explain most of the variation in its concentrations, and to explore the factors affecting its trends over the 2017–2022 timespan by using a bootstrap technique to estimate trend uncertainties. Moreover, the paper aims to reveal how changes in meteorology have affected the NO₂ concentration trends in the highest-density area of Romania.

In reaching the set goals, this study contributes to the existing literature in the following ways: (i) most importantly, it provides important climate and health policy information for an under-investigated EU capital city, Bucharest, which still struggles in the climate combat arena; (ii) it provides robust findings on the nonlinear effects of several meteorological factors on traffic pollution by switching between the GAM and GAMM setting based on model residual diagnostics and also detects fine nuances of influencing relationships by estimating unadjusted, single-factor adjusted, interaction effects models, and multifactor models, thus further assisting the issuers of policy with valuable information on the environmental conditions that can promote traffic-pollution mitigation; (iii) and it employs a battery of exploratory tools that complement the model estimations for uncovering the trends and patterns of traffic pollution in Bucharest, including the time plot, time variation, the trend heat, the calendar plot functions, and bivariate polar plots. Ultimately, by revealing new findings on trends, patterns, and impact factors for NO₂ concentrations for a vulnerable EU member, this research is equally important for local, national, and EU policymakers.

While this analysis centers on a major urban traffic site in Bucharest, the findings are most directly relevant to high-density metropolitan settings characterized by intense vehicular flows and complex microclimates. We acknowledge that pollutant dynamics and the meteorological sensitivities identified here may differ from those observed in smaller cities or rural areas across Romania. Accordingly, the empirical insights and policy considerations presented should be interpreted in the context of dense urban environments, and further research is encouraged to extend this modeling approach to a broader range of Romanian localities.

To guide the reader, the remainder of this paper is organized as follows. Section 2 outlines the most relevant results from the existing literature, and Section 3 describes the data sources, variables, and statistical methods used in the analysis, including detailed preprocessing and modeling steps. Section 4 presents the main results, including the estimation outcomes and diagnostic assessments for all model specifications. Section 5 discusses the findings in light of the previous literature, highlighting policy implications and limitations. The final section concludes and outlines avenues for future research.

2. Literature Review

Combustion produces gaseous nitrogen oxides, which form aerosols; therefore existing research has extensively evidenced that motor vehicle emissions are the primary source of NO₂ in outdoor air, serving as a good indicator of urban traffic-related air pollution [31]. The health impact of outdoor NO₂ emissions stems from its causative interaction with secondary pollutants such as ozone and particles, as well as its poor solubility, which allows it to harm lungs and cause airway inflammation, increasing the risk of respiratory infections [32,33]. Moreover, there is increasing evidence that NO₂, a prominent proxy for traffic-related pollution, has health effects that are independent from those of other common pollutants like ozone and particulate matter [34,35].

Studies using NO₂ as a proxy for TRAP are not many, as researchers preferred to focus on carbon emissions or particulate matter (usually PM_2.5) in their analyses. At the same time, effective data on NO₂ emissions caused by traffic and congestions are not easily available. Scholars used simulation modelling to estimate scenarios on the link between NO₂ concentrations generated by freeway and arterial traffic and health risks considering different traffic volumes during rush hour periods in the United States [36]. Their findings show that health risks from congestion are potentially significant and that additional traffic can substantially increase risks, depending on the type of road and other factors. Other authors [37] estimated the annual global number of new pediatric asthma cases caused by NO₂ exposure and found that approximately 4 million new pediatric asthma cases worldwide are linked to NO₂ pollution, with 64% in metropolitan areas. The exposure to traffic-generated NO₂ in the historic center of Macao was investigated by [38], who concluded that around 14% of the modelled NO₂ concentrations from the monitored 5965 receptor points are higher than the air quality standards for scenic spots in China. Ref. [39] assessed the impact of TRAP proxied by nitrogen oxides (NO₂ and NOx) on asthmatic persons quality of life in asthmatic persons within the Swedish Ga2len (Global Allergy and Asthma European Network). They concluded that asthmatics’ quality of life is not affected by moderate exposure to traffic pollution, but these findings need to be cautiously interpreted given that Sweden had in 2013 one of the lowest globally shares of death attributed to air pollution (1.2%, which declined to 0.81% in 2019). A recent systematic review on the impact of outdoor pollution and extreme temperatures on asthma-related outcomes concluded that short-term exposure to NO₂, along PM_2.5 and PM₁₀, has the potential to increase the risk of hospital admissions related to asthma, as well as visits to emergency departments in hospitals [40].

In comparison, the situation is worse in less environmentally conscious countries, where exposure to air pollution and TRAP exceeds WHO standards. Unfortunately, Romania is one of these countries, and its capital city, Bucharest, along with several other urban agglomerations, is a major source of traffic-related pollution. One of the rare studies that related to Romanian cities in terms of traffic air pollution investigated the temporal distribution of NO₂ in Galati and Braila using data collected by stationary monitoring stations in 2016 and 2017 [41]. The study discovered that NO₂ emissions from traffic sources contribute more to urban pollution than NO₂ emissions from industry sources. Furthermore, while emissions from industrial sources decreased between 2016 and 2017, emissions from traffic sources increased, which stresses the importance of traffic for air pollution in these cities. Moreover, Galati and Braila are smaller agglomerations of approximately 400–450,000 inhabitants jointly; therefore one expects to see more aggravating conditions in bigger cities. Ref. [42] examined temporal and geographical air pollution using data from monitoring stations and meteorological information at a circulating junction in Iasi, Romania’s fourth largest city by population (estimated at roughly 410,000 in 2021 by local authorities). They acknowledged that traffic was the primary cause of air pollution in the area, but contended that the number of vehicles, as well as traffic delays and congestion, must be examined together to appropriately determine the degree of pollution that threatens the adjacent population. Furthermore, local authorities were encouraged to build heavy-traffic deviations and effective traffic-light systems, as well as to increase green spaces around the city and monitor the health of the population living in the most vulnerable locations. Ref. [43] assessed the concentration levels of several contaminants in six city areas using data from the Environmental Protection Agency of Bucharest’s network of monitoring stations between 2005 and 2007. They discovered that traffic and the distribution of thermal power plants were the primary drivers of air pollution, particularly in the winter, indicating that meteorological circumstances have a strong influence on pollution. On the other hand, summer pollution seemed to be mostly caused by traffic and the re-suspension of road dust. Ref. [44] offered an assessment of the driving factors behind air pollution in five Romanian locations, of which three urban agglomerations (Bucharest, Cluj Napoca, and Iasi) were based on data collected from AirBase v.8 database of the European Environment Agency. Her findings show that pollutants concentration levels in the monitored sites were driven by temperature, wind speed, humidity, and the atmospheric pressure, while in the case of Bucharest the order was wind speed, temperature, atmospheric pressure, and humidity. Nevertheless, these results point to the intricated relations between traffic and climate and are warning signals that ignoring these relations in elaborating environmentally directed measures and policies is mistaken.

The COVID-19 pandemic’s lockdowns provided an excellent opportunity to assess the impact of restricted anthropogenic activities on air pollution in urban centers and thus to confirm the extensive influence of vehicle movement on air pollution. In the same vein, ref. [45] explored NO₂ concentration during lockdown in the Romanian city of Galati and concluded that its decrease was relatively small but may be attributed to local traffic reduction as part of the restrictions imposed during the pandemic by the Romanian authorities. After controlling for weather changes, ref. [46] claimed that lockdowns lowered NO₂ and particulate matter concentrations by 60% and 31%, respectively, in all 34 nations investigated. They assessed the dynamics of contaminants using satellite data, and quantified air pollution anomalies significant to public health using ground-level observations from over 10,000 air quality stations. Similarly, ref. [47] evaluated the link between traffic flows and nitrogen oxide and particulate matter (PM₁₀) concentrations in the city of Padova, Italy (211,000 inhabitants), during the shutdown in 2020. Their findings indicate that vehicle flows have a considerable influence on nitrogen oxide concentrations but not on PM₁₀, leading to the conclusion that traffic restriction measures appear to be successful for improving air quality only in terms of nitrogen oxide reductions. Ref. [48] also found that NO₂ reductions around British schools because of lockdown were significant, a result confirmed for schools in Wroclaw, Poland [49].

The integration of satellite remote sensing represents a significant advancement in air quality monitoring and prediction, offering encouraging opportunities for science-based anti-pollution measures and policies. One of the revolutionizing tools is the TROPOspheric Monitoring Instrument (TROPOMI) aboard Sentinel-5P that provides daily global NO₂ tropospheric column density measurements at high resolution, empowering the detection of individual emission sources and patterns in urban-scale pollution [50]. Recent studies show a strong correlation between TROPOMI NO₂ column observations and surface concentrations, particularly in urban settings [51,52]. Using advanced machine learning methods, ref. [53] have estimated surface NO₂ concentrations across Europe using TROPOMI data combined with meteorological variables, while [54] employed ensemble methods based on satellite-generated data for Los Angeles to demonstrate superior predictive performance for PM₂.₅ concentrations. Besides TROPOMI, other satellite-based collection tools that retrieve NO₂ available, such as the Ozone Monitoring Instrument (OMI) on the Aura satellite of NASA, incorporate significant improvements suggested by expert users and enhance data quality through improvements to the air mass factors (AMFs) employed in the retrieval algorithm [55]. For China, ref. [56] used the Berkeley High Resolution product (BEHR-HK) and meteorological outputs from the Weather Research and Forecasting (WRF) model to describe column NO₂, arguing that this approach is appropriated for places with large spatial variabilities and terrain height differences as is the case of China. Also, scholars have used daily data on NO₂ delivered by the South Korean instrument GEMS on the GK2B satellite, the first instrument in geostationary orbit that offers hourly daytime observations of NO₂ [57]. Ref. [58] similarly relied on sophisticated machine learning models such as VAR-XGBoost Kriging technique, and XGBoost to increase O₃ prediction accuracy and geographic distribution over time in China, finding that wind speed and temperature are the most critical parameters affecting O₃ pollution. In a more advanced approach, the work of [59] proposes a high-dimensional multi-objective optimization strategy for power systems that takes into account the spatial-temporal distribution of pollutants, achieving pollutant reductions and ensuring economic viability.

While satellite-based studies focus on large-scale spatial coverage, the GAM approach used in this study addresses an important research gap referring to the thorough understanding of the interactions between meteorological variables and pollution at a local urban scale, thus supporting the calibration of accurate satellite retrieval and further formulation of policy-relevant air quality measures.

3. Materials and Methods

3.1. Data Sample

The capital city and the economic, administrative, and cultural center of Romania is Bucharest (or “București” in Romanian) [60]. According to the TomTom Traffic Index for 2021, Bucharest remains one of the world’s cities with the most traffic. As such, the Romanian capital ranks sixth in Europe and eighth worldwide with an average congestion level of 50%. Estimations also show that people in the capital city of Romania spend 48% more time in traffic than is necessary. Thus, as of 2018, the amount of time lost in traffic in Bucharest is 218 h per year, or the equivalent of 27 days of work [61]. Furthermore, in 2021, Bucharest’s levels of congestion increased dramatically. The information provided by TomTom (https://www.tomtom.com/traffic-index/bucharest-traffic/; accessed on 22 October 2022) indicates that the daily average travel time increased in the city during 2021 by 5 min [62].

3.1.1. Measurement Stations

The air quality-monitoring site that provides hourly nitrogen dioxide pollutant concentration data employed in the current study is located in the center of Bucharest at a height of 30 m, latitude 44.43504, and longitude 26.098297—see Figure 1.

The study thus sourced hourly NO₂ data from the NET-RO016A air quality monitoring network in Bucharest, which has the EU-code STA-RO0070A. In turn, hourly meteorological data were obtained from the Filaret station, located in Bucharest at latitude 44.41667 and longitude 26.1.

The selected measurement station is sited at one of Bucharest’s busiest central traffic corridors, ensuring that the recorded NO₂ concentrations reflect predominantly traffic-related pollution. Although other urban activities may influence air quality, NO₂ at this location is overwhelmingly shaped by motor vehicle emissions, as supported by previous studies on European urban environments [29,63]. The analysis therefore provides insights most relevant to traffic-induced pollution, with the understanding that results are nested within the broader urban air quality context.

3.1.2. Pollutant Concentration and Meteorological Data

Hourly concentrations of NO₂ sourced from the urban traffic monitoring station in central Bucharest span the period from 1 January 2017 to 21 December 2022. Hourly data from the meteorological station (located at Otopeni, WMO code 154220-99999) include wind speed, wind direction, air temperature, atmospheric pressure, dew point, and relative humidity. The analysis period was selected to ensure a sufficiently long and recent timeframe while maximizing the availability of high-quality, complete data for both pollutant and meteorological parameters.

Initial raw datasets comprised over 51,000 hourly records for each variable. Data cleaning was performed in several steps to ensure data integrity and comparability. First, all negative and physically implausible NO₂ values (including any values greater than 500 μg/m³, and the observed minimum of –9.04 μg/m³) were excluded. After removing these values, only records with complete observations for all key meteorological covariates were retained for modeling. No imputation was performed; only complete cases were used.

As a result of this filtering process, the final analytical dataset used for statistical modeling includes 32,515 hourly records, with all variables present and within reasonable physical bounds.

Table 1. Variables details and descriptive statistics (cleaned, complete dataset).

Variable	Abbreviation	Unit	Mean	Median	Min	Max	1st Quartile	3rd Quartile
Nitrogen dioxide	NO2	µg/m3	50.42	46.81	2.82	218.05	31.66	65.69
Wind speed	ws	m/s	1.49	1.00	0.00	9.00	1.00	2.00
Wind direction	wd	Degrees	143.6	120.0	0.0	360.0	50.0	230.0
Air temperature	air_temp	°C	12.41	11.70	−15.60	40.20	4.20	20.20
Atmospheric pressure	atmos_pres	hPa	1017.2	1016.6	989.3	1044.1	1012.1	1022.2
Dew point	dew_point	°C	6.19	6.10	−17.40	24.40	0.70	12.20
Relative humidity	RH	%	70.60	73.44	12.41	100.00	53.04	90.11

has been updated to reflect the summary statistics of the cleaned and merged dataset, while

Table A1. Data cleaning and case selection steps applied to the merged air quality and meteorological dataset for Bucharest, 2017–2022.

Step	Records Remaining	Excluded at This Step	Excluded (%)
Raw merged dataset	51,203	–	–
Remove negative/outlier NO2 values	39,970	11,233	21.9
Remove incomplete meteorology	32,515	7455	14.6
Final model dataset	32,515	–	–

now includes a summary of the number and percentage of records excluded at each cleaning step.

3.2. Methods

Trends and patterns of NO₂ concentrations are first depicted by employing the time plot, time variation, calendar plot, and trend heat functions embedded in the “openair” package in R software [64]. Subsequently, bivariate polar plots are drawn to discover and assess the sources of air pollution [65,66,67]. Particularly, wind rose plots [68,69,70] are employed to reveal which meteorological conditions control high and low concentrations of NO₂ by splitting the time series of polluting concentrations into four quantiles and applying the windRose() function within the R’ “openair” package.

Subsequently, the nonlinear effects of the meteorological variables for NO₂ pollution in the capital city of Romania are explored through a robust generalized additive model (GAM) framework. GAM has become more prevalent in environmental research [71] due to its capacity to extend the generalized linear model (GLM) [72] by relaxing some of its restrictions [73]. Consequently, it is able to provide greater flexibility in analyzing complex nonlinear relationships [66,71,74,75]. In the context of urban air quality studies, and especially in high-density traffic environments such as central Bucharest, the use of GAMs has become standard practice for disentangling the intertwined effects of traffic emissions and meteorological variability (see [29,63]), further supporting their adoption in the present analysis.

Although recent developments in machine learning, including Random Forests, Gradient Boosting Machines, and deep neural networks, have been used to predict air quality and retrieve NO₂ concentrations from satellites [53,54], these methods frequently compromise interpretability for predictive accuracy. In light of the policy-relevant goal of separating the discrete and interdependent impacts of meteorological conditions on NO₂, the current study places a high priority on explainability and transparent inference. A special blend of statistical rigor and interpretability is provided by generalized additive models (GAMs), which enable flexible estimation of smooth surfaces and nonlinear relationships while yet producing precise, statistically validated effect estimates. Compared to black-box models, GAMs are particularly advantageous for environmental applications where understanding the mechanistic role of covariates—and quantifying uncertainty—is as important as predictive accuracy. Nevertheless, as demonstrated in several recent studies the predictive performance of GAMs for traffic-related NO₂ is often comparable to that of more complex machine learning algorithms when trained on moderate-sized, high-quality datasets.

A GAM can use the smooth spline function, kernel function, or regression smooth function to fit response and explanatory variables in the following manner:

$$g\left[E\left(Y\right)\right]={\beta }_0+f_1\left(X_1\right)+f_2\left(X_2\right)+f_3\left(X_3\right)+\cdots +f_n\left(X_n\right)+\epsilon ,\epsilon \left(0,{\sigma }^2\right)$$

(1)

where E(Y) stands for the expectation of the explained variable, ${\beta }_0$ for the intercept, and $\epsilon$ for the normally distributed residual; $g()$ and $f_1$() to $f_n()$ are the connection and the spline functions, respectively, whereas $X_n$ denote the explanatory variables.

In the context of generalized additive models (GAMs), the “smooth term” is a nonparametric function—typically represented as a penalized regression spline—used to flexibly model the relationship between an explanatory variable and the response. Formally, for a given predictor $\left(X_j\right)$, the smooth function $f_j\left(X_j\right) \mathrm{can}$ be written as:

$$f_j\left(X_j\right)={{\displaystyle \sum }}_{k=1}^{K_j}{\alpha }_{jk}{B}_{jk}\left(X_j\right)$$

(2)

where $B_{jk}\left(X_j\right)$ are basis functions (e.g., thin plate splines or cubic regression splines), ${\alpha }_{jk}$ are coefficients estimated from the data, and $K_j$ is the basis dimension chosen for the smoother. The complexity of each smooth term is controlled by penalizing excessive “wiggliness” in the fitted function, with the degree of smoothing (i.e., effective degrees of freedom) determined by generalized cross-validation (GCV) or another information criterion:

$$\mathrm{Penalty}={\lambda }_j{\int }_{a_j}^{b_j}{\left[f_j^{″}\left(x\right)\right]}^2 dx$$

(3)

where $a_j$ and $b_j$ are the minimum and maximum observed values of the predictor x, and higher values of ${\lambda }_j$ lead to smoother estimated functions. In this study, the optimal degree of smoothness for each term is automatically selected via GCV during model fitting.

In estimation, the GAM seeks to minimize the penalized sum of squared residuals, balancing data fidelity and smoothness of the estimated functions. The optimization problem is formulated as:

$$\underset{f_1,\dots ,f_p}{min}{{\displaystyle \sum }}_{i=1}^n{\left[y_i-{\beta }_0-{{\displaystyle \sum }}_{j=1}^p{f}_j\left(x_{ij}\right)\right]}^2+{{\displaystyle \sum }}_{j=1}^p{\lambda }_j{{\displaystyle \int }}_{a_j}^{b_j}{\left[f_j^{″}\left(x\right)\right]}^{2}dx$$

(4)

subject to each $f_j$ being twice-differentiable on [a_j,b_j], and where the penalty parameter ${\lambda }_j$ controls the degree of smoothness for each term, and the second derivative ${\left[f_j^{″}\left(x\right)\right]}^2$ quantifies the function’s curvature. The optimal values of ${\lambda }_j$ are determined by generalized cross-validation (GCV) during model fitting.

In this study, the hourly NO₂ concentration is the response variable, whereas meteorological factors depicted in Table 1 are the explanatory variables. Moreover, the trends and seasonal variations in NO₂ concentrations are included in all estimated GAMs. The gam() function in R’s “mgcv” package is called to perform estimations [76]. As per [65,66], the log connection function was employed to connect explanatory and response variables. Furthermore, in all model estimations, fitting is accomplished by employing the thin plate smoothing spline function, whereas smoothing parameters are produced through generalized cross-validation (GCV), which provides the non-trivial advantage of computing efficiency. Moreover, an extra penalty is included to allow coefficients to tend toward zero with the argument select = TRUE set within the gam() function.

First, the GAM that does not account for any meteorological covariate, called the unadjusted GAM, has been estimated. On the other hand, models that account for one or more MET factors are called adjusted GAMs. Consequently, the unadjusted model that assesses the trend and seasonal components is given by:

$$log\left({NO}_2\right)={\beta }_0+s\left(trend\right)+s\left(month\right)+\epsilon$$

(5)

where the trend component is estimated as:

$$trend=year+\left(month-1\right)/12$$

(6)

Subsequently, meteorological factors are introduced into GAMs one at a time, thus producing the single-factor (i.e., MET factor) adjusted GAMs, such that:

$$log\left({NO}_2\right)={\beta }_0+s\left(trend\right)+s\left(month\right)+s\left(ME{T}_i\right)+\epsilon$$

(7)

where i is sequentially the wind speed, wind direction, atmospheric pressure, air temperature, dew point, and relative humidity.

In cases where serial correlation in residuals is suspected, a generalized additive mixed model (GAMM) is fitted to the data. The GAMM extends the GAM by introducing a correlation structure on the residuals, typically modeled as an autoregressive process of order one (AR(1)). The general form of the GAMM applied in this study is:

$$log{\left({NO}_2\right)}_t={\beta }_0+s\left({trend}_t\right)+s\left({month}_t\right)+{{\displaystyle \sum }}_j{f}_j\left(X_{j,t}\right)+u_t$$

(8)

$$u_t=\mathsf{\rho }{u}_{t-1}+{\epsilon }_t,\hspace{1em}{\epsilon }_t\sim N\left(0,{\sigma }^2\right)$$

where $u_t$ denotes the error term following an AR(1) structure with autocorrelation parameter ρ, $f_j$ represents smooth functions for the meteorological predictors or their interactions, and ${\epsilon }_t$ is a white noise error term. This structure is implemented using the corAR1 argument within the gamm() function from R’s mgcv package, allowing the model to capture and adjust for serial dependence in the monthly averaged residuals.

The rationale for employing a GAMM lies in its capacity to account for any residual temporal autocorrelation not explained by the primary covariates and smooth components. The inclusion of this structure allows for more accurate estimation of parameter uncertainty when autocorrelation is present [76].

Of note, the variance inflation factor (VIF) of environmental factors was calculated to estimate multivariable multicollinearity, where higher multicollinearity is linked with a greater VIF. In turn, if a predictor variable is uncorrelated with other predictors, its VIF is 1 [77]. Estimation results (i.e., via the vif() function in R’s “car” package) indicated high VIF values for two of the MET variables, i.e., air temperature and dew point, suggesting that collinearity would be an issue in model estimations. After deleting the two factors, all VIF values decreased and passed the significance test.

In a further step, interaction effects between MET variables were also considered. Particularly, all interaction terms involving the meteorological factor with the highest impact on NO₂ concentration (i.e., detected within the single-factor framework) are considered one by one in interaction-effects GAM estimations.

Model selection at each step was guided by multiple statistical criteria. Candidate models were compared using the Akaike Information Criterion (AIC) and Generalized Cross-Validation (GCV) scores, with preference given to specifications yielding the lowest values. In parallel, adjusted R² and the proportion of deviance explained were used to assess explanatory power. The final, optimal model was chosen as the specification achieving the best balance of parsimony and fit, as indicated by the lowest AIC and GCV, alongside the highest adjusted R² and deviance explained. This comprehensive approach ensured that the selected model was both statistically robust and substantively meaningful.

The basis dimension (k) for each smooth term in the GAMs was selected based on both substantive considerations and methodological guidance [76]. For each model, k was set to a value comfortably above the anticipated effective degrees of freedom, balancing flexibility and parsimony. Following estimation, model adequacy and potential undersmoothing were checked via the gam.check() function in R. The diagnostics confirmed that, for all models reported, the effective degrees of freedom (edf) remained below the specified k for every smooth, and the k-index values and their p-values indicated no evidence of undersmoothing or model misspecification (see Appendix A for representative output). This approach ensures robust inference while minimizing the risk of overfitting.

For robustness checks, this study also estimates generalized additive mixed models (GAMMs) [78,79], an extension of generalized additive models incorporating random effects that are extensively employed to model correlated and clustered responses [80]. In terms of integrating smooths, GAMMs allow the same flexibility as GAMs. A significant aspect of GAMMs is that each term in the model describes a partial effect, i.e., the impact of that particular term when all the remaining terms in the model are held constant [81]. This approach also fulfills a secondary purpose by correcting any potential residual correlation in the GAM framework. Thus, to assure the robustness of results, all models are also specified in the form of GAMMs via the gamm() function in R’s ‘mgcv’ package.

Model adequacy and selection between GAM and GAMM are guided by detailed diagnostic analysis. For each model specification, we assess residuals for approximate normality, heteroscedasticity, and temporal autocorrelation using Q-Q plots, residuals vs. fitted plots, histograms, and autocorrelation function (ACF) plots. Where significant autocorrelation remains in the residuals, the GAMM framework is preferred; otherwise, the GAM is adopted for its interpretability and parsimony. All diagnostic results are documented in Figure A1.

To further guard against multicollinearity, variance inflation factors (VIFs) were calculated for all meteorological predictors using the vif() function in R’s “car” package. Variables with VIF substantially greater than 5 were excluded from the final multifactor model.

A comprehensive schematic summarizing the sequential methodological steps and analytical framework employed in this study is presented in Figure 2.

4. Results

4.1. Trends and Patterns of NO₂ Concentrations

Figure 3a shows the evolution of monthly NO₂ concentrations in Bucharest over the analysis period spanning 2017–2022. The air quality in Bucharest has registered an overall improvement between 2017 and 2022, as NO₂ levels show a decreasing trend that has intensified during the most recent period. Figure 3b offers a clear depiction of the 5-year trend in NO₂ concentrations in Romania’s capital by normalizing the data by first estimating annual means and subsequently indexing it to 100 at the beginning of 2017. The graphical representation in Figure 3b thus shows that concentrations of NO₂ have decreased by approximately 40% over the last 5 years. The two figures also show the beneficial impact of COVID-19 restrictions imposed in 2020, illustrated by the drop in NO₂ levels below 30 µg/m³ in the spring and general NO₂ levels below the ones in 2019.

Figure 4 presents the temporal variation of NO₂ concentrations in central Bucharest, focusing on monthly and weekday patterns. The results reveal a distinct annual cycle: NO₂ levels are lowest in late spring (April–May) and December, while concentrations rise sharply during autumn, peaking in October. This seasonal dynamic likely reflects the interplay between photochemical activity, meteorological conditions, and changes in emissions patterns associated with residential heating and traffic. Weekly variation is also evident, with consistently elevated NO₂ concentrations during weekdays and a pronounced decline over the weekend, highlighting the strong influence of anthropogenic activity, particularly traffic emissions, on local air quality.

Figure 5 further reinforces the analysis and reflects the variation in NO₂ concentrations by year, season, and daylight versus nighttime hours. It confirms that the NO₂ concentrations are decreasing over the most recent years and indicates that the highest concentrations are recorded in autumn, during afternoon hours.

Moreover, Figure 6 provides useful information on the meteorological conditions that control high and low concentrations of NO₂ in central Bucharest. A distinct wind rose is drawn for each quantile of measured NO₂ concentrations, showing that relatively strong southwesterly winds dominate the lowest concentrations of NO₂, whereas the highest NO₂ concentrations are dominated by northeasterly winds.

4.2. GAM Estimation Results

4.2.1. Unadjusted NO₂ GAM Results

To establish a baseline for temporal variation, a generalized additive model (GAM) was first estimated including only the long-term trend and the monthly seasonal component. The model achieves an adjusted R² of 0.57, explaining 64% of the deviance in log-transformed monthly mean NO₂ concentrations (

Table 2. Unadjusted generalized additive model (GAM) results for monthly mean NO₂.

Term	edf	Ref.df	F	p-Value
s(trend)	4.93	7	10.44	<0.001
s(month)	4.66	7	1.41	0.068
Adj. R2: 0.57	Deviance explained: 64%

).

Figure 7a confirms that NO₂ decreased during 2017–2022 in central Bucharest, whereas Figure 7b shows that there is seasonal variation in NO₂ levels, with the highest concentrations recorded in September–October and the lowest in April–May. Moreover, the impact of the COVID-19 pandemic outbreak on NO₂ levels in the first half of 2020 is also highlighted, although pollutant concentrations have returned and even surpassed their pre-pandemic level by the end of 2020. However, the nitrogen dioxide concentrations did register a significant decrease over the most recent year. A closer look at the uncertainty intervals depicted in Figure 7a reveals that the drop in NO₂ during the analysis period has been statistically significant at the 95% level of confidence.

4.2.2. Single-Factor Adjusted NO₂ GAM Results

To assess the contribution of individual meteorological covariates, each variable was incorporated into the GAM alongside trend and seasonality. Wind speed and relative humidity each yielded strong improvements in fit, with adjusted R² values of 0.66 and 0.73, respectively (

Table 3. GAM results for log(NO₂) with single meteorological factors (controlling for trend and month).

Model	edf (Covariate)	F (Covariate)	Adj. R2	Deviance Explained	GCV
Wind speed	6.41	3.25	0.662	73.6%	0.0282
Wind direction	2.15	0.84	0.521	55.4%	0.0335
Air temp	2.48	0.62	0.576	64.9%	0.0333
Atm. pressure	4.51	1.92	0.668	75.3%	0.0289
Dew point	3.08	1.13	0.597	66.5%	0.0314
Rel. humidity	5.55	4.82	0.731	79.1%	0.0226

). Both factors display pronounced nonlinear effects on NO₂.

Next, Figure 8 depicts the fitted components of the wind speed-adjusted NO₂ GAM model in central Bucharest and confirms the evolution of the trend and seasonal components, when the most relevant MET covariate is accounted for. The left panel reflects the long-term trend in log-transformed NO₂ levels from 2017 to 2022, revealing a pronounced decline, particularly in the latter half of the period. This downward trend likely reflects both regional emission reduction efforts and the transient impact of pandemic-related mobility restrictions, with the effect statistically significant throughout. The middle panel demonstrates marked seasonality, with elevated NO₂ concentrations in late autumn and early winter and lower values during spring and summer, in line with known patterns of atmospheric dispersion and seasonal emission variability in temperate urban environments. The right panel highlights the role of wind speed, evidencing a robust inverse relationship: increased wind speed is consistently associated with lower NO₂ concentrations, indicating effective dilution and removal of traffic-related pollution during episodes of stronger atmospheric mixing. The confidence intervals (shaded regions) underscore the statistical robustness of these estimated effects, whereas these findings emphasize the necessity of accounting for both temporal dynamics and meteorological factors in modeling urban air quality.

4.2.3. GAM Models with Interaction Effects

To further elucidate the complex relationships between wind speed and meteorological covariates in modulating NO₂ concentrations, a suite of interaction-effect GAMs was estimated. These models employ tensor product smooths (te) to capture both the nonlinear and interactive influences of wind speed and each meteorological factor.

Table 4. Interaction-effects generalized additive model (GAM) estimation results.

Interaction	edf (te)	F-Value	Deviance Explained (%)	Adjusted R2	GCV
Wind speed × Wind direction	7.45	1.641	81.1	0.737	0.02390
Wind speed × Air temperature	6.03	1.071	81.0	0.729	0.02515
Wind speed × Atmospheric pressure	12.83	2.449	88.4	0.803	0.02162
Wind speed × Dew point	16.06	3.981	90.2	0.834	0.01831
Wind speed × Relative humidity	7.71	3.930	86.6	0.807	0.01823

presents the results for each bivariate specification.

Model fits indicate substantial improvements in explanatory power over single-factor models, particularly for interactions with atmospheric pressure, dew point, and relative humidity. Notably, the wind speed × dew point model produced the best fit, with an adjusted R² of 0.834 and 90.2% of deviance explained. This indicates a particularly strong joint effect of wind speed and dew point on NO₂ variability in central Bucharest. Interactions with atmospheric pressure and relative humidity also resulted in high model performance, confirming the importance of meteorological interplay in urban air quality.

All interaction terms are statistically significant, with p-values well below 0.001, except for some slight differences in edf and F-value reflecting the nonlinearity and complexity of each relationship.

The wind speed × dew point interaction, in particular, highlights the complex role of atmospheric moisture in NO₂ dispersion, likely reflecting both dilution and condensation processes. From a meteorological perspective, low wind speeds limit both horizontal and vertical pollutant dispersion, leading to local NO₂ accumulation. Dew point, representing absolute atmospheric moisture, exerts an additional effect: at moderate dew points, the stability and limited turbulent mixing may promote NO₂ persistence. As dew point increases, enhanced atmospheric moisture facilitates wet scavenging and promotes aqueous-phase chemical reactions that efficiently remove or transform NO₂ into more soluble products [82]. This explains why NO₂ concentrations decline with increasing dew point and wind speed, reflecting the combined action of dispersion and chemical removal mechanisms.

The wind speed × atmospheric pressure and wind speed × relative humidity models also capture strong meteorological modulation effects, as evidenced by their high deviance explained. The wind speed × atmospheric pressure surface demonstrates that stagnant, high-pressure systems with low wind speed can trap pollutants near the surface, exacerbating NO₂ accumulation, whereas increased winds or deviations from average pressure aid dispersion and vertical mixing. The interaction between wind speed and relative humidity reveals that the removal of NO₂ is maximized at high RH and wind speed, consistent with the enhanced wet deposition and turbulent mixing found in humid, windy conditions.

Figure 9 illustrates the estimated bivariate smooth surface for wind speed and dew point. High NO₂ concentrations are found at low wind speeds and moderate dew point values, whereas increases in either variable are associated with declining pollutant levels, consistent with greater atmospheric dispersion and mixing. The non-linear effect of relative humidity on NO₂, seen as an inverted-U shape, reflects the dual role of moisture in urban air chemistry. At moderate RH, photochemical reactions may favor NO₂ formation or limit removal, but as RH exceeds ~80%, efficient wet scavenging and aqueous-phase reactions (e.g., conversion to nitric acid) dominate, sharply reducing NO₂ concentrations. This phenomenon is well-documented in the atmospheric chemistry literature and is typical of humid urban environments.

4.2.4. Multi-Factor Adjusted NO₂ GAM Results

In light of previous findings, the multi-factor models were constructed to evaluate the concurrent contribution of various factors on NO₂ concentration. In the multi-factor setting, the degree of fitting between the influencing factors and NO₂ concentration was improved (adjusted R² of 0.805 and explained 87.2% of the deviance).

Table 5. Multifactor NO₂ generalized additive model (GAM) estimation results.

Term	edf	F	p-Value
Long-term trend (year)	6.57	19.58	<0.001
Seasonality (month)	4.77	2.25	0.004
Wind speed × Atmospheric pressure interaction	2.45	1.02	0.010
Wind direction	0.71	0.17	0.138
Relative humidity	5.79	4.63	<0.001
Adj. R2 = 0.805	Deviance explained = 87.2%

reports estimation results for the best-fit multifactor model including included trend, seasonality, the interaction of wind speed and atmospheric pressure, wind direction, and relative humidity. All major covariates (except wind direction) were highly significant and nonlinear.

The results of the multifactor GAM, visualized in Figure 10, offer a nuanced depiction of how both temporal and meteorological drivers influence ambient NO₂ concentrations at the urban traffic monitoring site in Bucharest.

The long-term trend reveals a pronounced non-linear decline in NO₂ over the study period, with statistically significant reductions observed particularly after 2020 and a tendency toward stabilization in the most recent year. The seasonal effect highlights distinct intra-annual variability, characterized by minimum concentrations in late spring and a secondary peak in autumn—patterns consistent with local meteorological conditions and emission cycles.

The interaction surface between wind speed and atmospheric pressure demonstrates that the highest NO₂ concentrations are encountered under conditions of low wind speed and mid-range atmospheric pressure. Conversely, pollutant levels decline at higher wind speeds or when pressure deviates from its typical range, underscoring the importance of atmospheric dispersion and vertical mixing in modulating air quality.

The smooth effect of wind direction suggests a modest but discernible influence on NO₂, with the lowest concentrations associated with prevailing northwesterly winds. This likely reflects the microclimatic setting of the monitoring site and the distribution of local emission sources upwind.

Finally, the effect of relative humidity is strongly non-linear: NO₂ concentrations increase gradually up to about 80% RH, above which they decline sharply. This pattern is suggestive of enhanced atmospheric removal processes or chemical transformations occurring under very humid conditions.

4.3. Robustness Checks and Model Diagnostics

To ensure the robustness of the main findings, the preferred multi-factor GAM specification was extended to a generalized additive mixed model (GAMM) with an AR(1) correlation structure [73,76].

Table 6. Multifactor NO₂ generalized additive mixed model (GAMM) estimation results.

Term	edf	Ref.df	F-Value	p-Value
Long-term trend (year)	1.00	1.00	6.95	0.011
Seasonality (month)	4.59	7.00	2.58	0.003
Wind speed × Atmospheric pressure interaction	3.00	3.00	4.71	0.006
Wind direction	1.00	1.00	0.22	0.642
Relative humidity	1.86	1.86	3.51	0.072
Adjusted R2	0.48

Note: edf = effective degrees of freedom.

summarizes the GAMM estimation results for the primary model including the trend, seasonality, the bivariate wind speed–atmospheric pressure interaction, wind direction, and relative humidity.

Of mention, variance inflation factors (VIFs) were calculated for all model predictors, and all values were below 5, confirming the absence of problematic multicollinearity.

The GAMM results reaffirm the primary meteorological drivers of NO₂ concentrations. The wind speed–atmospheric pressure interaction and the seasonal component remain statistically significant, while the trend is modeled as linear in the presence of autocorrelation adjustment. The overall fit, reflected by an adjusted R² of 0.48, is lower than in the unconstrained GAM, as expected when explicitly modeling serial correlation in the residuals.

Regression diagnostics for both the multi-factor GAM and the corresponding GAMM are presented in Figure A1. Residual plots, Q-Q plots, histograms, and autocorrelation function (ACF) plots confirm that residuals are reasonably symmetric, approximately normal, and not subject to problematic autocorrelation. These findings support the adequacy of the models for inference and suggest that the main results are not unduly influenced by violations of regression assumptions.

Of note, the distinction between the GAM and GAMM frameworks is particularly important when evaluating model fit and the explanatory power of covariates. In our case, the primary multifactor GAM model achieves an adjusted R² of 0.81 and explains 87.2% of deviance. However, when the same model is estimated as a GAMM—including an AR(1) autocorrelation structure on the residuals—the adjusted R² falls to 0.48. This is not an indication that the model has become misspecified, but rather a consequence of the GAMM partitioning a substantial portion of variance into the autocorrelation structure, rather than the predictors themselves [76].

In practice, the appropriate modeling strategy depends on the structure of the residuals. For our data, residual diagnostics for the multifactor GAM (see Figure A1) confirm the absence of significant autocorrelation and demonstrate approximate normality. This is due in part to aggregation to monthly means and the inclusion of flexible smooths for long-term trend and seasonality, which absorb much of the temporal dependence. In such settings, the GAM provides a parsimonious and interpretable summary of the data-generating process.

Where strong residual autocorrelation persists, however, the use of a GAMM is warranted to ensure valid inference; in such cases, one should expect the explanatory power of the covariates to decrease accordingly, as some variance is attributed to the autocorrelation process itself.

5. Discussion

5.1. Air Pollution Trends in Romania

Air pollution is a major public health concern that must be addressed by Romanian local and national authorities. Specifically, Romania is ranked seventh in the European Union in terms of premature deaths attributed to air pollution and carries the highest annual social cost per capita, with estimations indicating that each Romanian loses €1.810 because of air pollution. In particular, as of 2021, Bucharest recorded the highest pollution index in Romania, with an index of 75 [83]. Moreover, Bucharest is the most congested EU city and consistently ranks among the most congested cities in the world, with a high share of aged diesel cars in the vehicle fleet [84], which makes traffic pollution responsible for two-thirds of air pollution in the Romanian capital. Concurrently, nitrogen oxides (NOx = NO₂ + NO) constitute one of the largest contributors to air quality degradation in urban/industrialized areas, while anthropogenic NOx emissions, mostly from fuel burning, account for approximately 65% of total global NOx emissions. The majority of NO₂ non-compliance cases worldwide are encountered in urban areas, owing to traffic-related NOx emissions from diesel vehicles [85].

In terms of pollution, Romania remains a key source of worry in the European Union. According to the European Environment Agency’s (EEA) country fact sheet on air pollution, the percentage of urban population exposed to NO₂ concentrations over EU criteria was 2.5% in 2019, up from 0% in 2016, while exposure to PM₁₀ exceeded EU requirements by 11%, down from 58.2 in 2015. However, Romania has significantly improved its air pollution indicators: in 2022, the percentage of population exposed to exposed to NO₂ concentrations above EU standards dropped to 0.4%, while the exposure to PM₁₀ exceeded fully met EU standards [86]. A simple examination of the European Air Quality Index reveals the overall poor quality of the air across Romanian monitoring points, particularly in Bucharest—for example, on 3 November 2022, eight of the ten stations in Bucharest, the country’s capital, indicated poor air quality, with only two indicating moderate air quality (https://www.eea.europa.eu/themes/air/air-quality-index/index; accessed on 23 November 2022). In comparison, all three stations in Ljubljana, Croatia’s capital city, and just one in twelve in Prague, Czech Republic’s capital city, indicated poor air quality, but no other capital city in the European Union matched this unfavorable situation.

The important influence of traffic on local air quality is supported by recent data from the Bucharest Environmental Protection Agency (Agenţia pentru Protecția Mediului București, APMB) [87]. From 20 November to 3 December 2024, APMB carried out a targeted air monitoring program that used a mobile laboratory to measure the amounts of pollutants at important high-traffic access sites into Bucharest, such as the A1, A2, and A3 highway entries. The study found that, while most pollutants remained below hourly and daily legal thresholds, concentrations of PM₁₀ consistently exceeded daily limits at all monitoring sites, and NO₂ hourly maxima approached or surpassed 80 μg/m³—levels comparable to those reported by the permanent National Air Quality Monitoring Network. These results demonstrate the continued occurrence of sporadic air quality exceedances, especially in locations with significant traffic volumes during rush hours. The official study emphasizes that the recorded amounts closely follow the conditions of weather dispersion and traffic intensity, particularly for NO₂ and particle matter. Despite the campaign’s short duration, its findings support the idea that urban transportation still has a significant impact on Bucharest’s air quality, supporting previous conclusions drawn from larger datasets and models.

Moreover, the latest CE Delft report for the European Public Health Alliance (EPHA) places, sadly, Romania in first place based on the annual social cost per capita caused by air pollution (1810 euro). However, the cost for people living in Bucharest is estimated at 3004 euro, followed by 1710 euro for Brasov and 1643 euro for Timisoara, the other two major cities in the country. Moreover, the deaths of 14,723 Romanians were attributed in 2019 to outdoor air pollution, 6 and 6.5 times higher than by indoor air pollution and drug use, respectively. But these numbers tell only part of the story. The same report also shows that no Romanian city has introduced low emission zones or congestion charging (considered among the most efficient air pollution combating measures—see, in this respect, [88,89]). Moreover, the vehicle park is old (half of Romanian cars are older than 15 years) and municipalities have not rushed to replace diesel buses used in public transportation by electric ones. Therefore, EPHA sees air pollution as a critical public health issue that needs to be addressed sooner rather than later through preventive measures instead of reactive actions.

Fortunately, there is some fresh air on the horizon, as Romania is a member of the European Union, the world’s leading region in the fight against climate change and the first major economic area to establish a legally binding framework for all member states (27 by the end of 2021) aimed at meeting the Paris Agreement pledges. In this context, extensive research is needed to assist local and national authorities in implementing the most efficient and effective measures to mitigate pollution and meet the mandatory EU targets.

5.2. Model Findings

This study has applied a GAM framework backed by GAMM specifications to explore the relationships between NO₂ concentration and its influencing factors in Bucharest, the capital city of Romania. Results indicate that NO₂ concentration in Bucharest had significant trend and seasonal components, as well as noteworthy correlations with meteorological factors. These findings confirm previous conclusions reached by [43,44] for Romanian cities, but also the outcomes of the research conducted by [90,91,92].

Our research shows that unadjusted GAMs can be constructed to explain a considerable amount of the hourly variation in the concentration of NO₂ at an urban traffic site and confirms that there is clear evidence of a downward trend in NO₂ concentrations in Bucharest over 2017–2022 revealed by both the adjusted and unadjusted models. Current results reflect the effect of local measures aimed at reducing traffic pollution that have been implemented in Bucharest during the analysis period. This is further confirmed by the dramatic decrease in NO₂ concentrations detected during the first wave of the COVID-19 pandemic due to the containment measures implemented in Romania, which caused sharp reductions in traffic. Thus, in 2020 the municipality established the “Oxygen vignette”, a special fee on vehicles operating in the capital city of Romania that do not meet Euro 5 emission criteria. The European emission standards refer to pollution resulting from the use of new vehicles sold in the European Union, EEA countries, and the UK, as well as ships in EU waters. These standards have progressively restricted the pollutants generated by vehicles when in use since 1992. The Euro 5 standard has been issued in 2009. Currently, the most restrictive standard is Euro 6, and Euro 7 will enter into force in 2026 [93]. As part of this measure, the city center, where the air quality monitoring station is located, has been designated as part of an air quality action area (ZACA). In turn, this could have impacted the air quality in central Bucharest and thus explained the decreasing trend detected in the estimations. The calendar plots (not shown) also confirm this finding, highlighting days where maximum levels registered for NO₂ concentrations in central Bucharest exceed 150 μg m⁻³ (very high) and 100 μg m⁻³ (high) in 2017 and in 2021 (the last full year of data), respectively. The plots attest that the number of very high NO₂ pollution days decreased from 23 in 2017 to 9 in 2021, whereas the number of high NO₂ pollution days decreased from 122 to 79 over the five-year period. However, it should also be acknowledged that traffic pollution continues to plague the capital of Romania, with important health and societal costs [94].

Moreover, estimations consistently report that wind speed has a significant negative impact on NO₂ concentrations. Consequently, we find that at high wind speed levels, pollutant emissions are more easily dispersed, which confirms previous findings that report a lower level of ozone and its precursors when wind speed is high [95,96,97]. On the other hand, these results deviate from those of [73,77], which in turn report a positive correlation between wind speed and pollutant emissions concentrations, a discrepancy that can be explained by the characteristics of the geographical location of the measurement station that obstruct plume dilution and instead produce recirculation (i.e., a street canyon at Marylebone Road in central London and a river valley basin in Lanzhou, respectively).

Other important findings reveal that higher air pressure is conducive to the removal of nitrogen dioxide. In this respect, our findings are in line with those of [66,77,98], which also encounter a negative link between air pressure and environmental pollutants. As [99] show, a low-pressure environment contributes to the accumulation of air pollutants, which is in line with current results. In turn, as per [100], a higher atmospheric pressure fosters the dispersion of air pollutants. Moreover, a nonlinear negative relationship between relative humidity and NO₂ concentrations has been detected, confirming previous studies (among others, [66,101,102]. The removal of NO₂ as the relative humidity increases can be explained by the higher water vapor saturation that removes pollutants and decreases pollutant concentration levels [77].

While recent years have seen the growing use of machine learning algorithms such as random forests, boosted trees, and neural networks for NO₂ estimation and air quality forecasting (see [29,77,103]) these models often emphasize predictive accuracy at the expense of interpretability and transparency. On the other hand, the GAM framework, as used in this work, allows for the simultaneous modeling of seasonality, nonlinearity, and meteorological interactions in a way that is still directly interpretable and available for inference that is relevant to policy. Our results are in line with new research showing that GAMs often offer model fits comparable to more sophisticated machine learning techniques for datasets with comparable quality and temporal resolution, while providing significantly more insight into the interactive and mechanistic roles of meteorological drivers. The justification for preferring GAM-based models in urban air quality evaluation is highlighted by this balance of adaptability, interpretability, and strong statistical inference.

Overall, the current research acknowledges that the cause of NO₂ pollution in the capital city of Romania is multifaceted, with various concurrent influencing factors and interaction effects. Consequently, Bucharest’s NO₂ pollution is mainly affected by the trend and seasonal components and the interaction of meteorological factors. In particular, aside from the trend and month, the interaction effect of wind speed and air pressure showed the most significant effect on variations in NO₂ concentrations. The multifactor model showed the best goodness-of-fit and the highest total deviance explained for Bucharest NO₂ concentration, producing a 10% increase in the adjusted R squared as compared to the unadjusted GAM model. Consequently, the study reports that meteorological factors are important driving factors for changes in air quality, supporting the findings of [66,103,104,105,106], among others. Not least, the current study also contributes to confirming a non-trivial strength of the GAM approach that makes it particularly suitable for environmental research, which consists of its ability to model the interaction between influencing factors as a bivariate surface [73].

5.3. Policy Implications

In light of these findings, Romanian authorities should implement a thorough policy framework that integrates traffic management with atmospheric science results, with measures that can be enhanced during periods of high wind speed and increased atmospheric pressure, and more interventions during stagnant atmospheric conditions characterized by low wind speeds and high humidity levels. Drawing from successful European experiences [107,108,109], Romania should prioritize measures to encourage personal vehicle renewal and reduce the share of diesel cars in the vehicle fleet, which could result in an immediate reduction in traffic-related pollution in the capital city and other urban areas. This approach has been demonstrated by the positive impact of Madrid’s low-emission zone (LEZ) which is fully implemented as of January 2025 and covers the entire municipal area [88]. In the same vein, a scenario analysis of eight European cities revealed that urban NO₂ mainly originates from diesel vehicles, with around 15% reduction achievable from passenger diesel cars, 13% from trucks, and 6% from vans [85]. This can be accomplished through programs that provide subsidies for the renewal of old cars and vehicles for both individuals and businesses, as well as by incentivizing people to use public transportation instead of their own cars. The implementation of low-emission zones in major urban areas becomes essential, building on the successful experiences of Germany, United Kingdom or Italy [107,108]. Considering our findings, these zones should be contingent on local meteorological patterns, including higher restrictions during autumn months and afternoon hours when NO₂ concentrations peak. Furthermore, the meteorological insights regarding the interaction between wind speed and atmospheric pressure should form the basis for designing emergency air quality protocols that activate traffic restrictions when adverse weather conditions exist. Additionally, authorities should put in place air quality plans that explicitly protect health, focusing on traffic planning and management. The EU, that Romania is part of, is a successful story of coordinated policy approaches that led to significant reductions in transport emissions [86,110,111].

In an era of rising and volatile oil and gas prices, public buses, trams, subways, and trains are not only less expensive than driving a car, but they can also help reduce congestion and air pollution. Furthermore, investments in electric buses are required to combat pollution, as are smart traffic management systems that reduce congestion. Examples from Spain and Germany that have drastically reduced the cost of public transportation may be pursued; while they may appear expensive at first glance, they will significantly save public money in terms of health costs associated with pollution. Similarly, local and central governments may encourage ride sharing for daily commuting through special programs and advertising the overall positive impact on everyone’s life. Also, smart traffic management systems may incorporate real-time meteorological data to optimize traffic flows, maximizing the natural air cleaning effect of favorable wind conditions and atmospheric pressure that our study identified.

In addition, Romanian authorities should adopt legal frameworks aligned with the EU’s revised Ambient Air Quality Directive, entered into force in 2024, which sets significantly higher binding air quality standards for 2030 compared to previous limits, that aim at improving air quality and reducing the harmful effects of pollution on human health and the environment. At the same time, national regulations should permit local authorities to impose green taxes and traffic restrictions, as well as pedestrian zones, combined with sustainable urban planning measures.

Another approach is to encourage citizens to ride bicycles by providing special routes and emphasizing the overall health benefits of cycling. Going even further, urban gardening initiatives and the use of green roofs may successfully reduce CO₂ emissions while increasing social interaction, efficient space use, and food security.

5.4. Limitations and Future Research

As with most empirical investigations, this study is not without limitations. The results presented here pertain to a single, high-density urban monitoring site in central Bucharest. While this location is emblematic of traffic-related air quality challenges found in many major European capitals, caution is warranted when generalizing these findings to cities, towns, or peri-urban areas with different urban morphology, population density, emission profiles, or meteorological regimes. Air quality dynamics—and, by extension, the most effective policy interventions—may diverge substantially in other Romanian regions, especially those characterized by lower traffic intensity, more dispersed built environments, or unique climatological conditions. It follows that differentiated, regionally tailored air quality management strategies are likely to be needed to ensure policy effectiveness on a national scale.

Methodologically, the use of a flexible generalized additive modeling (GAM) framework provides a powerful means of disentangling the nonlinear effects of meteorological and temporal covariates on NO₂ concentrations. Nonetheless, the estimated smooth functions and associated confidence bands are inevitably shaped by both the structure of the underlying data and modeling choices such as the selection of smoothing parameters, basis dimension (k), and the inclusion or exclusion of correlated predictors. Parametric uncertainty in the coefficients and structural uncertainty in the form and stability of the smooth functions remain intrinsic to all spline-based models, particularly in settings where covariates may be interrelated or important sources of variability remain unobserved. For example, while the penalty term in GAMs helps guard against overfitting, model results still reflect the data at hand and may not capture structural breaks, rare meteorological events, or evolving emission patterns outside the analysis window.

Moreover, while our analysis provides robust estimates of average associations and their uncertainty, the study does not explicitly decompose the contribution of individual emission sources, nor does it explore higher-order interactions among more than two meteorological drivers. Both the parametric and nonparametric components of our model are thus subject to the usual caveats of observational research, including potential residual confounding and the risk of extrapolation beyond the observed data range.

These limitations highlight the need for future research that (i) applies the current modeling framework across a wider range of Romanian cities and rural locations, (ii) leverages additional source attribution data where available, and (iii) investigates the robustness of the smooth effect estimates under alternative model specifications, such as Bayesian approaches or machine learning-based ensemble models. In particular, further studies could apply the GAM methodology employed in this paper to large-scale NO₂ data retrieval in other metropolitan areas. This requires datasets such as satellite-derived NO₂ columns from TRO-POMI/Sentinel-5P, ground-based monitoring networks from European Environment Agency stations, and ERA5 high-resolution global reanalysis data from the European Centre for Medium-Range Weather Forecasts (ECMWF). Moreover, using machine learning techniques together with GAM can provide real-time NO₂ forecasts for European cities, while including socio-economic indicators and emission data can support the assessment of policy impact in areas of interest.

Only through such extensions can we move toward a more comprehensive, regionally nuanced understanding of urban air pollution dynamics and generate actionable insights for policymakers tasked with improving air quality across diverse spatial contexts.

6. Conclusions

This paper documents trends in NO₂ concentrations as well as the influence of relevant meteorological variables on NO₂ pollution in the capital city of Romania, Bucharest. Overall, this study contributes to the extant literature by using fine-granularity data and bringing new evidence related to trends and patterns of traffic-related pollution in an under-investigated EU capital city, by assessing the multifaceted nonlinear effects of various meteorological factors on NO₂ concentrations, including the wind speed, wind direction, atmospheric pressure, and the relative humidity, and by employing a multi-model approach to assure the robustness of findings (GAM and GAMM).

The main findings of the study can be summarized as follows: (i) the highest NO₂ concentrations in central Bucharest are registered in the autumn, in working days, and during the afternoon hours; (ii) wind speed has a significant negative impact on NO₂ concentrations, as pollutant emissions are more easily dispersed at high wind speed levels; (iii) atmospheric pressure has a significant negative impact on NO₂ concentrations, as a higher atmospheric pressure fosters the dispersion of NO₂; (iv) relative humidity has a significant negative impact on NO₂ concentrations, as higher water vapor saturation removes pollutants; (v) the interaction effect of wind speed and air pressure has the most significant negative impact on NO₂ concentrations, surpassing the additive individual effects.

These results provide important insights for both researchers and policymakers aiming to better understand and address urban air pollution in rapidly changing European environments. By demonstrating the value of high-resolution, temporally detailed analysis and the use of flexible statistical models, this study highlights the necessity of context-specific approaches in environmental monitoring and policy formulation.