This research was generally conducted in six general steps (Figure 1). In the first step, six air pollutants, including PM_2.5, PM₁₀, SO₂, NO₂, O₃, and CO, were monitored over 10 years. During this phase, data from monitoring stations were collected to capture the concentrations of these pollutants. A comprehensive study area map was also created, incorporating 11 spatial criteria known to influence air pollutant levels. These spatial criteria included altitude, humidity, distance to industrial areas, NDVI (normalized difference vegetation index), population density, rainfall, distance to the street, temperature, traffic volume, wind direction, and wind speed. In the next step, the researchers aimed to determine the presence of multicollinearity among the spatial criteria. The multicollinearity test assessed the degree of correlation between the various parameters. This analysis was crucial in identifying redundant or highly correlated variables, allowing for eliminating or consolidating such factors to avoid multicollinearity issues in subsequent modeling steps. To understand the importance of the spatial criteria with air pollutant concentrations, the Geodetector method was employed. This method assessed the contributions and significance of each spatial criterion to the overall air pollution levels. It helped prioritize influential factors and provided insights into the relative importance of various spatial parameters. The researchers combined the CNN algorithm and the AE technique in the modeling phase. By integrating these two methods, the researchers could leverage the strengths of both approaches. The encoded data was then fed into the CNN, which effectively learned the spatial relationships and patterns associated with the concentrations of the six air pollutants. This fusion approach enhanced the accuracy and precision of the modeling process. The trained CNN-AE model generated risk maps for the six air pollutants. To generate risk maps, the predicted values obtained for each pixel in the study area were assigned to the center of each pixel. Then, using raster to point analysis, the risk map was created. In the next step, the natural breaks classification method was utilized to categorize the risk classes for each pollutant. These risk maps provided spatial representations of the pollutant concentrations across the study area. The results and risk maps obtained from the modeling process were evaluated and compared in the final step. Evaluation metrics such as mean absolute error (MAE), coefficient of determination (R ²), root mean square error (RMSE), and the area under the curve (AUC) of the receiver operating characteristic (ROC) were employed to assess the accuracy and performance of the models.

Tehran city (the capital of Iran) is located at 35^° 36′to 35^° 44′north latitude and 51^° 17′to 51^° 33′east longitude and an altitude of 1032–1832 m above sea level. The city of Tehran as the most populous city in Iran (9,039,000) in the last 2 decades due to reasons such as the unsustainable development of industrialization and urbanization, the ever-increasing growth of the transport fleet, and the emission of their pollutants, ineffective national environmental air quality standards, and dust storms. The Middle East has faced severe air pollution, especially (PM₁₀, PM_2.5, O₃, NO₂, SO₂, and CO) (Yousefian et al., 2020). In general, 20% of Iran’s energy is consumed in Tehran. The mountain ranges surrounding the city of Tehran stop the flow of humid wind to the capital, so in winter, the cold weather and lack of wind cause the polluted air to be trapped inside the city (Naddafi et al., 2012). Urban space structures are deeply connected with the urban transportation system (Rodrigue, 2020). The statistics of Tehran indicate the high rate of land consumption in this city, which has caused a high growth in the area and size of the town, which has caused an increase in the distance and the amount of transportation (private cars and public transportation) to carry out administrative and educational activities, and entertainment, in Tehran. Figure 2 displays the geographic location of the study area in Tehran province, Iran, highlighting the air quality control monitoring stations and meteorological stations.

Air pollutants can be categorized as either natural or anthropogenic and can be classified as primary or secondary. Primary pollutants are released directly into the atmosphere from a particular source, retaining the same composition. On the other hand, secondary pollutants are not directly released into the atmosphere and are formed in case of a reaction or interaction of primary pollutants or become another compound in the atmosphere, such as photochemical smog (Bhargav, 2020). Six pollutants, PM_2.5, PM₁₀, SO₂, NO₂, O₃, and CO, are considered for this research. Air pollution is a challenging environmental issue that endangers the health and wellbeing of people worldwide, comprising a complex blend of gaseous components and suspended particles (Özbay, 2012; Bergstra et al., 2018). Air quality in cities in developing countries has gradually deteriorated due to rapid urbanization, population growth, and industrialization (Turalıoğlu et al., 2005). The annual average concentration of air pollutants in Tehran was measured from 1 January 2012, to 1 January 2022, using data collected from 23 air quality monitoring stations located in the city. The characteristics of the six pollutants are shown in Table 1, and the trends of the data in the years 2012–2022 are shown in Figure 3. Maps related to the concentration of pollutants were prepared using kriging interpolation in ArcGIS 10.8 software with a pixel size of 30 × 30 m (Figure 4). For modeling, high-risk areas for each pollutant were converted into occurrence points (with a target value of 1) and the low-risk regions into non-occurrence points (with a target value of 0). In the following, each of these parameters will be explained.

➢ Particulate Matter (PM)

Suspended particles include large particles (PM₁₀) and fine particles (PM_2.5), associated with lung cancer and asthma. PM₁₀ can settle in the bronchi and lungs, and PM_2.5 is the most minor and most dangerous type of suspended particle and can penetrate deep into the respiratory system (Quercia et al., 2015). PM_2.5 particles are mainly caused by fuel combustion, construction dust, and vehicle exhaust, which cause dust-haze. All types of manufactured combustion and some industrial processes are among the most common human sources of PM₁₀ (Özbay, 2012).

➢ Sulfur Dioxide (SO₂)

SO₂ is mainly obtained through the combustion of fossil fuels, biomass burning, and melting ores containing sulfur (Santosa et al., 2008). It is also released through industrial activities and is considered among the harmful gases that affect human, animal, and plant life (Manisalidis et al., 2020). The release of SO₂ in industrial regions can result in serious health concerns such as respiratory irritation, bronchitis, mucus production, bronchospasm, skin redness, eye and mucous membrane damage, and deterioration of cardiovascular health (Chen et al., 2007). Moreover, the environmental consequences of SO₂ include acid rain and soil acidification (Manisalidis et al., 2020).

➢ Nitrogen Oxide (NO₂)

NO₂ is a common traffic-related pollutant that originates from automobile engines, and it is one of the most prevalent air pollutants found in urban regions (Dragomir et al., 2015; Richmond-Bryant et al., 2017). NO₂ is one of the compounds that lead to adverse effects on the environment and human health (Mavroidis and Ilia, 2012), disrupting the sense of smell, burning eyes, throat, and nose, reducing visibility and changing the color of the fabric (Chen et al., 2007).

➢ Carbon Monoxide (CO)

CO is produced due to incomplete or inefficient fuel combustion and affects the blood oxygen transfer in the body and heart (Quercia et al., 2015). CO gas emission and production sources encompass all combustion sources, including motor vehicles, power plants, waste incinerators, domestic gas boilers, and cookers (Vakkilainen, 2017; Manisalidis et al., 2020).

➢ Ozone (O₃)

O₃ is a secondary photochemical pollutant resulting from the oxidation of volatile organic compounds, including benzene, in nitrogen oxides. This colorless, pungent, and reactive gas is the primary component of smog, which is mainly attributed to automobile emissions in urban regions. The concentration of O₃ in urban areas typically increases in the morning, reaches its peak in the afternoon, and decreases at night (Yerramilli et al., 2011).

The influential factors in this research include meteorological data (rainfall, temperature, humidity, wind direction, and wind speed), altitude, NDVI, distance from street, distance from industrial areas, traffic volume, and population density (Delavar et al., 2019; Shogrkhodaei et al., 2021). Each of the mentioned factors was prepared with a 30 × 30 m pixel size in ArcGIS 10.8 software (Figure 5). In the following, each practical criterion for air pollution has been examined.

• Meteorology data

Air pollution has two natural and human causes, natural causes include volcanic eruptions and severe drought, and human activities include motor vehicle emissions, industry, and the burning of agricultural lands and forests, which cause the release of various types of pollutants with multiple characteristics and effects (Sakti et al., 2023). Pollutants in the atmosphere can be dispersed or diluted under various meteorological conditions, such as rainfall, air temperature, and wind speed (Özbay, 2012). Meteorological data, such as wind speed, wind direction, precipitation, temperature, and humidity, were collected from the National Meteorological Organization. The data were obtained from 12 stations and represented the annual average between 2012 and 2022. The kriging interpolation technique was used in ArcGIS 10.8 to create these maps, with a pixel size of 30 * 30 m. The following discusses the impact of meteorological parameters on air pollutants.

➢ Rainfall

Rainfall is one of the main factors of meteorological conditions that affect air quality and has a specific inhibitory effect on air pollutants (Guo and Jiang, 2020; Shukla et al., 2008). Rainfall can affect the concentration of air pollutants by removing gaseous pollution and deposition of suspended particles through atmospheric chemical processes (Kayes et al., 2019).

➢ Temperature

Temperature plays a crucial role in urban air quality, directly influencing gas properties, heterogeneous chemical reaction rates, and the gas-to-particle partitioning process (Aw and Kleeman, 2003). Sunlight and high temperatures stimulate chemical reactions in pollutants and increase smog. The effect of temperature on air pollutants is such that an increase in temperature increases the dispersion and decreases the concentration of contaminants (Shogrkhodaei et al., 2021).

➢ Humidity

Humidity, as one of the meteorological parameters, plays an essential role in air pollutants (their concentration and dispersion) in the urban environment (Endeshaw and Endeshaw, 2020). Most pollutants negatively correlate with relative humidity, so the amount of air pollutants decreases with the increase in humidity (Kayes et al., 2019).

➢ Wind speed and direction

Air quality is affected by wind speed. One case is that wind speed reduces the concentration of pollutants and dilutes them (in areas with higher concentrations). Another issue is that wind speed leads to the entry of contaminants from further distances and increases the concentration of pollutants in an area with a lower concentration (Oleniacz et al., 2016).

• NDVI

Urban vegetation impacts air quality by affecting the sedimentation and dispersion of pollutants (Janhäll, 2015). Urban trees and vegetation are considered an ecosystem regulating service in removing air pollutants (Setälä et al., 2013). The NDVI is a primary indicator of the physiological properties of land vegetation. The NDVI standard was prepared using Landsat-8 satellite images in Google Earth Engine (GEE) (https://earthengine.google.com/) as an annual average from 2012 to 2022. NDVI index was calculated using Eq. 1. NDVI = NIR   –   Red NIR + Red (1)

In this equation, the symbol NIR denotes the reflectance in the near-infrared band, and the symbol Red represents the reflectance in the red band. By taking the difference and sum of these reflectance values, the NDVI equation normalizes the values and produces an index that ranges from −1 to +1.

•Altitude

Air pollution is affected by the change in altitude, so the increase in altitude causes an increase in sunlight and causes the problem of photochemical smog (U.S. EPA, 1978). This research prepared the height through a digital elevation model (DEM) with a pixel size of 30 × 30 m through the SRTM (Shuttle Radar Topography Mission) image in GEE platform.

• Distance to industrial areas

Industrial sources located inside and close to city borders are among the influential primary factors of urban air pollution (Hosseini and Shahbazi, 2016). Heavy industry causes the release of many dangerous pollutants in the air that affect health (Bergstra et al., 2018). Industrial areas were extracted through land use layers of industrial areas with a scale of 1:10,000. Subsequently, the aforementioned criterion was transformed into a raster map with a pixel resolution of 30 × 30 m by employing the Euclidean distance function in ArcGIS 10.8.

•Distance to street

Motor vehicles produce more air pollutants than any other human activity. Motor vehicle emissions from roads can be considered as a mobile line source with an emission rate per unit of road length (Oji and Adamu, 2020). Therefore, the distance from the measurement location to the roads affects the air quality near the streets (Dragomir et al., 2015). The data relating to the roads of Tehran was extracted through the open street map (OSM) (https://www.openstreetmap.org) with a scale of 1:100,000 in 2022. Subsequently, the mentioned criterion was transformed into a raster map with a pixel resolution of 30 × 30 m by utilizing the Euclidean distance function within the ArcGIS 10.8 software.

• Traffic volume

Urban air pollution is primarily caused by traffic emissions (Guarnieri and Balmes, 2014). Monitoring data about pollution near roads shows hot pollution spots in high-traffic areas (Samet, 2007). Traffic congestion increases vehicle emissions, decreases air quality, and increases air pollutants, including CO, CO₂, nitrogen oxides (NO_x), and PM, which cause complications such as death to drivers, passengers, and people who live near the main roads (Zhang and Batterman, 2013). Data on the traffic volume in Tehran city were obtained from the Tehran Traffic Control Company. The data represent the average traffic volume between 2015 and 2020.

• Population density

The expansion of urban areas and population growth has a significant adverse effect on ambient air quality (Kumar et al., 2016), as the rise in population is linked to an increase in the number of vehicles (traffic density) and industrial and commercial operations (Shogrkhodaei et al., 2021). This factor was obtained based on the data from Iran Statistics Center in 2017.

A multicollinearity test was performed to assess the presence of multicollinearity among the independent variables utilized in the geographic modeling and risk mapping of the six air contaminants. The results of the test, presented in Table 2, indicate the levels of VIF for each independent variable. From the results, it can be observed that none of the independent variables have VIF values exceeding the threshold limit of 10. This indicates no severe multicollinearity among the independent variables, suggesting that they can be considered individually and collectively in the spatial modeling and risk mapping analysis. The VIF values, ranging from 1.07 to 6.80, indicate that the independent variables have relatively low to moderate levels of correlation with each other. This suggests that the variables provide unique information and do not excessively duplicate each other’s predictive power.

The GeoDetector method was employed to determine the importance of different parameters on air pollutants (Figure 6). The analysis revealed distinct findings for each pollutant. CO, temperature, wind speed, and wind direction emerged as the most significant parameters. Variations notably influenced the levels of CO in these factors. In the case of O₃, humidity, precipitation, and altitude were identified as the primary criteria affecting its concentration. Altitude plays a crucial role in the formation and distribution of ozone in the atmosphere.

Conversely, for PM₁₀, altitude, wind direction, and wind speed were deemed the most influential parameters. These factors influenced the dispersion and transport of PM₁₀ particles. Regarding NO₂, altitude, rainfall, and wind direction were found to have the most significant impact. Altitude affected the vertical distribution of NO₂, while rainfall and wind direction influenced its dispersion and movement. Similarly, for PM_2.5, altitude, rainfall, and temperature were identified as the key parameters. Altitude affected the vertical distribution of PM_2.5 particles, while rainfall and temperature were crucial in their formation and dispersion. Lastly, for SO₂, temperature, wind direction, and altitude were determined as the most important parameters. Temperature played a role in the chemical reactions involving SO₂, while wind direction and altitude affected its transport and dispersion. In general, altitude, wind direction, wind speed, rainfall, and temperature parameters had the most significant effect on pollutants in the study area.

The AE, comprising encoder and decoder layers, is a pre-training step to learn a compact and efficient representation of the input data. The CNN, however, is designed to capture spatial patterns and dependencies within the pollutant data. The input data includes various spatial features such as altitude, humidity, distance to industrial areas, NDVI, population density, rainfall, distance to the street, temperature, traffic volume, wind direction, and wind speed. The Autoencoder’s encoded features serve as input to the CNN, which then extracts spatial features. The model’s weights, biases, learning rates, regularization techniques, and dropout rates are randomly initialized and updated during the training process using the Adam optimizer. A loss function is utilized to measure the difference between the predicted pollutant concentrations and the actual measurements to assess the model’s performance. Common regression loss functions, such as mean squared error (MSE), are commonly used. The results of the loss functions for all pollutants, as shown in Figure 7, indicate the convergence and effectiveness of the integrated CNN-AE model. The loss function values for the training and test data decrease throughout training, demonstrating the model’s ability to learn and capture the underlying patterns in the pollutant data. The decreasing trend of the loss function values suggests the model successfully minimizes the discrepancy between the predicted pollutant concentrations and the actual measurements during training. This indicates that the model is learning to make accurate predictions and is effectively capturing the complex relationships within the data. The decreasing loss function values in the training and test data support the notion that the integrated CNN-AE model successfully learns and generalizes to unseen data, highlighting its ability to capture the spatial patterns and dependencies of the air pollutants.

Additionally, metrics such as MAE, RMSE, and R ² are calculated to assess the accuracy and predictive power of the model (Table 3; Table 4). For the pollutant PM_2.5, the CNN model exhibited reasonably good performance, achieving an R2 of 0.889. The corresponding RMSE and MAE values were 0.166 and 0.046, respectively. However, the CNN-AE model surpassed the CNN model’s performance, demonstrating an improved R ² of 0.969. Moreover, the RMSE and MAE values for the CNN-AE model were 0.087 and 0.157, respectively, indicating better accuracy and precision in predicting PM_2.5 concentrations.

Regarding the pollutant PM₁₀, both models performed exceptionally well. The CNN model achieved an impressive R ² of 0.972, suggesting that the model could explain approximately 97.2% of the PM₁₀ concentration variance. Additionally, the CNN model exhibited low RMSE and MAE values of 0.082 and 0.053, respectively. The CNN-AE model further enhanced the prediction accuracy, yielding an even higher R ² of 0.98. The RMSE and MAE values for the CNN-AE model were 0.0701 and 0.045, respectively, indicating a significant improvement over the CNN model. For the pollutant SO₂, both the CNN and CNN-AE models demonstrated commendable performance. The CNN model achieved an R ² of 0.951, suggesting that the model could explain approximately 95.1% of the SO₂ concentration variability. The corresponding RMSE and MAE values were 0.11 and 0.075, respectively. The CNN-AE model showed similar performance, with an R2 of 0.955, indicating a comparable ability to explain the variability in SO₂ concentrations. The RMSE and MAE values for the CNN-AE model were 0.105 and 0.067, respectively, demonstrating their effectiveness in predicting SO₂ levels.

Regarding the pollutant NO₂, both models exhibited solid predictive capabilities. The CNN model achieved an R ² of 0.972, indicating that the model could explain approximately 97.2% of the NO₂ concentration variability. The RMSE and MAE values were 0.083 and 0.054, respectively, suggesting accurate predictions. The CNN-AE model outperformed the CNN model, attaining an exceptional R ² of 0.994. The RMSE and MAE values for the CNN-AE model were significantly lower at 0.038 and 0.032, respectively, indicating superior precision and accuracy in predicting NO₂ concentrations. For the pollutant O₃, both models demonstrated satisfactory performance. The CNN model achieved an R ² of 0.949, suggesting that the model could explain approximately 94.9% of the O₃ concentration variability. The RMSE and MAE values were 0.112 and 0.08, respectively. The CNN-AE model improved the prediction accuracy with an R ² of 0.96.

Regarding the CO pollutant, the CNN model demonstrated a high level of performance, as indicated by an R ² value of 0.952. This suggests that the model’s predictions account for around 95.2% of the variability in CO concentrations. The RMSE and MAE values for CO were calculated as 0.109 and 0.073, respectively. Notably, the CNN-AE further enhanced the accuracy of CO predictions. The CNN-AE model achieved an improved R ² value of 0.978, indicating that the model captured approximately 97.8% of the CO concentration variability. The corresponding RMSE and MAE values were calculated as 0.073 and 0.044, respectively.

Moving on to the test data, the CNN exhibited moderate performance with R ² values ranging from 0.57 to 0.715 for six pollutants. The RMSE values ranged from 0.265 to 0.324, indicating some difference between the predicted and actual values. The MAE values ranged from 0.162 to 0.233, representing the average absolute difference between predicted and actual values. In contrast, the CNN-AE improved performance on the test data compared to the CNN. It achieved higher R ² values ranging from 0.681 to 0.829, indicating a better fit. The lower RMSE values, ranging from 0.205 to 0.281, suggested more accurate predictions. The MAE values ranged from 0.106 to 0.185, indicating a more negligible average absolute difference between predicted and actual values compared to the CNN.

In summary, integrating the AE with the CNN algorithm showed promising results in air quality management. The CNN and CNN-AE models exhibited strong performance in the training phase, with the CNN-AE model consistently outperforming the CNN. Although there was a slight decrease in performance during the testing phase, the CNN-AE model maintained its superiority over CNN. Figure 8 shows the fitting diagram of the training and test data on the target data.

Using the trained model, the CNN-AE model estimated pollutant concentrations for each location in the study area. These estimated concentrations were then assigned risk levels to different regions based on classification criteria. The risk levels could be categorized as very low, low, moderate, high, and very high, representing varying degrees of pollution severity (Figure 9). The risk maps were generated by overlaying the estimated pollutant concentrations onto a geographical map of the study area. Each region was color-coded according to the assigned risk level, providing an intuitive visualization of the pollution hotspots and areas of concern, and according to the risk maps generated from the CNN-AE model, the southwest and northeast regions exhibited higher risk levels for CO pollution. Concerning O₃ pollution, elevated risk levels were observed in the north, east, and west areas. The risk of NO₂ pollution was particularly pronounced in the north and central regions. In the case of SO₂ pollution, the risk was concentrated in the south and northeast areas. PM₁₀ pollution posed a higher risk in the west and southwest regions, while PM_2.5 pollution was more prominent in the southern part.

Several evaluation metrics were employed to assess the effectiveness of the risk maps generated by the CNN-AE method, including the ROC curve, AUC, and Youden index J. These metrics were used to analyze the performance of the risk maps in terms of their ability to accurately discriminate between different risk levels. The evaluation results, as presented in Table 5 and Figure 10. For NO₂, an AUC of 0.964 was obtained, indicating a high level of discrimination between different risk levels. The Youden index J was 0.8936, further confirming the model’s ability to identify the optimal threshold for risk classification. The Standard Error was 0.0235, and the 95% Confidence Interval ranged from 0.903 to 0.991, indicating high precision in the risk map. The z statistic value was 19.72, and the significance level was p < 0.0001, demonstrating the statistical significance of the results.

Similarly, for PM₁₀, an AUC of 0.95 was achieved, indicating good discriminatory power. The Youden index J was 0.8276, highlighting the model’s effectiveness in identifying risk thresholds. The Standard Error was 0.0175, and the 95% Confidence Interval ranged from 0.907 to 0.977, indicating a high confidence level in the risk map. The z-statistic value was 25.764, and the significance level was p < 0.0001, further confirming the statistical significance of the findings. The performance of the CNN-AE algorithm was also evaluated for CO, PM_2.5, O₃, and SO₂. The AUC values for CO, PM_2.5, and O₃ were 0.896, 0.878, and 0.877, respectively, demonstrating moderate to high discriminatory power. The Youden index values were 0.75, 0.7368, and 0.7292, indicating the model’s ability to identify suitable risk thresholds. The Standard Errors were 0.0321, 0.0298, and 0.0391, respectively, showing the precision of the risk maps. The 95% Confidence Intervals ranged from 0.827 to 0.944 for CO, 0.815 to 0.926 for PM_2.5, and 0.794 to 0.935 for O₃, further strengthening the reliability of the risk estimates. The z-statistic values were 12.332, 12.69, and 9.632, respectively, and the significance levels were p < 0.0001 for all three pollutants, underscoring the statistical significance of the observed results. Lastly, for SO₂, an AUC of 0.811 was obtained, indicating an acceptable level of discrimination between risk levels. The Youden index J was 0.6308, suggesting the model’s capability to identify appropriate risk thresholds. The Standard Error was 0.0425, and the 95% Confidence Interval ranged from 0.733 to 0.874, providing a reliable estimate of the risk map. The z statistic value was 7.317, and the significance level was p < 0.0001, affirming the statistical significance of the results. Integrating the AE with the CNN algorithm proved effective in spatial modeling and risk mapping of the six air pollutants. The high AUC values, significant Youden index values, narrow confidence intervals, and low p-values indicate the model’s ability to discriminate between different levels of pollutant risk and its statistical reliability. These results contribute to our understanding of the spatial distribution and potential.