An XGBoost-based cellular automata for modeling PM₂.₅ concentration using ground and satellite data Authors: Sajad Farokhi, Jamshid Maleki, Ehsan Foroutan Published: 2025 ABSTRACT AbstractAir pollution is one of the major environmental issues in Iran, driven by industrial greenhouse gas emissions, vehicle exhaust, power plants, and natural sources such as dust storms. Among various pollutants, PM₂.₅ is especially concerning due to its high prevalence in urban areas. Governments have utilized air quality monitoring stations to measure PM₂.₅ concentration levels to implement policies aimed at controlling this pollutant. Fixed monitoring stations, which operate for regulatory purposes, are typically located in urban areas where a large portion of the population resides. However, the measurements obtained from these stations are limited due to their insufficient number, uneven distribution, and inadequate spatial coverage, which prevent continuous pollutant estimation over large areas and the generation of comprehensive risk maps. To address these limitations, this study introduces a novel hybrid modeling framework for modeling the spatiotemporal distribution of PM₂.₅ across Iran. The proposed approach integrates Cellular Automata with the XGBoost algorithm, leveraging both satellite-based products and ground observations. Within this framework, XGBoost generates continuous transition potentials for each neighborhood window, enabling dynamic representation of spatial variations in PM₂.₅ levels. The resulting model achieved an adjusted R² of 0.88, demonstrating strong performance in capturing regional and temporal variability. A key strength of the method is its ability to localize the influence of predictor variables across different regions, allowing the development of region-specific models within each neighborhood window and improving both interpretability and predictive accuracy.Keywords: PM₂.₅ concentration, Cellular automata, Spatiotemporal modeling, Regional air pollution, Ground-based data, Satellite dataSubject terms: Climate sciences, Environmental sciences Introduction Air pollution is a pressing global challenge, particularly in rapidly industrializing economies; Iran exemplifies this pattern, as an aging, high-emitting vehicle fleet increases exposure and risk1,2. Since industrialization began in the 1970 s, ambient pollutant levels have risen steadily, and over the past two decades, concentrations have reached hazardous levels in several major cities, including Tehran, Mashhad, Tabriz, Isfahan, Ahvaz, Arak, and Karaj. A 2016 World Bank assessment estimated the annual economic damage from air pollution in Iran at roughly US$13 billion—about 2.2% of GDP3. Among regulated pollutants, fine particulate matter ≤ 2.5 μm in aerodynamic diameter (PM₂.₅) warrants particular attention because it penetrates deep into the lungs and is strongly associated with cardiopulmonary morbidity and mortality4. Accordingly, modeling PM₂.₅ concentrations is critical for targeting mitigation and evaluating policy in Iran. The main PM₂.₅ emission sources in Iran include urban traffic, industrial facilities (e.g., refineries and power plants), residential fuel combustion, and transboundary dust storms entering the country from western and southeastern borders1. Studies and national assessments indicate that many urban areas routinely record PM₂.₅ concentrations and Air Quality Index (AQI) values that exceed health-based guidelines, underscoring substantial population exposure and public-health risks1. According to1, in the 2019 revision of Iran’s National Ambient Air Quality Standards (NAAQS), the annual permissible limit for PM₂.₅ is 12 µg m⁻³; however, many metropolitan areas frequently exceed this threshold throughout the year. Moreover, the national air quality monitoring network, consisting of approximately 160–200 active stations in recent years, exhibits strong spatial imbalance, with most stations concentrated in large urban centers while vast rural, mountainous and border regions remain unmonitored1,2. In this context, satellite remote sensing combined with data-fusion and advanced machine-learning models provide a practical route to generate spatially continuous PM₂.₅ fields and to mitigate the biases introduced by uneven monitoring coverage2. Given the limited coverage of fixed monitoring stations, air pollution modeling is indispensable for estimating pollutant concentrations across entire urban and regional domains. To date, three principal approaches have been employed in air pollution modeling: deterministic models, statistical models, and machine learning–based non-deterministic models. Deterministic models rely on classical mathematics, using aerodynamic theories and physical-chemical processes to model pollutant dispersion and emission3. However, the practical application of these models is often restricted due to their reliance on numerous highly accurate parameters, the need for large input datasets, computational complexity, and extended modeling timeframes4. Moreover, the accuracy of deterministic models is strongly influenced by the spatial extent of the study area and the quality of the input data. Statistical models and machine learning models are among the most widely used approaches in air pollution modeling5–8. Statistical models typically operate without considering the complex physical-chemical reactions occurring in the atmosphere and are based solely on data-derived relationships9. However, these models often assume linear relationships, whereas most real-world relationships are non-linear. Machine learning models have been extensively used to explore non-linear and complex relationships between observations and predictive variables5,10–14. Nevertheless, machine learning, particularly deep learning models, requires large volumes of labeled data and relatively long training times. Moreover, deep learning models exhibit a high dependency on the network architecture designed for neural layers, and the black-box nature of these hidden layers adds to their complexity. To overcome these limitations, cellular automata (CA) models can be employed to model changes in pollutant concentrations influenced by multiple sources and factors15–19. CA models can simulate complex processes using simple mathematical operations20,21. The CA models are discrete, grid-based computational models composed of individual cells, each of which updates its state over time based on a set of transition rules and the states of neighboring cells. These components enable CA to capture spatial and temporal dynamics in complex systems such as air pollution dispersion. However, the development of an efficient CA model requires sufficiently accurate and reliable data, which fixed monitoring stations alone cannot provide due to their limited spatial coverage. Therefore, accurate and spatially detailed PM₂.₅ concentration measurements at regional or urban scales are essential. To address this gap, satellite remotely sensed data have been widely used to retrieve continuous, spatially resolved PM₂.₅ measurements at regional and urban scales22–24, providing critical information for estimating exposure levels25. In a recent study, a pixel-level PM₂.₅ projection method was employed using Landsat imagery, monitoring data, and historical PM₂.₅ grids16. Tang et al. (2023) developed a pixel-level PM₂.₅ projection method that integrated Landsat imagery, monitoring data, and historical PM₂.₅ grids. Applied to the Yangtze River Delta, the approach combined CA-simulated land-use changes with a spatial lag model to estimate future PM₂.₅ distributions, revealing fine-scale spatiotemporal trends that can inform sustainable policy planning16. In another study, a hybrid model that integrates land-use regression and CA to simulate real‑time NO₂ dispersion across urban environments by accounting for on‑road emissions, meteorological, and morphological factors17. Their model achieves an external R² of 0.57 and successfully replicates NO₂ reductions observed during the first COVID‑19 lockdown. Integrating data from fixed monitoring stations with ground-based and satellite data can potentially enhance the development of CA models for continuous regional-scale PM₂.₅ concentration modeling. CA models have demonstrated robust performance in spatiotemporal modeling applications, such as land use change and urban development monitoring19,26, air quality simulation and analysis17,18,27, forest fire monitoring28,29, and more. However, limited studies have focused on modeling PM₂.₅ concentrations using integrated ground, satellite, and reanalysis data to achieve countrywide fields in Iran, a prerequisite for robust policy evaluation2. While the model proposed in2 demonstrated acceptable performance in capturing large-scale spatial variations, its ability to represent localized dynamics and neighborhood-level interactions remained limited. To overcome these limitations, the present study introduces a hybrid framework that integrates CA with the XGBoost algorithm. This combination enables dynamic spatial transition modeling and improves predictive accuracy, providing a more realistic representation of PM₂.₅ spatiotemporal variability across Iran. The main objective of this study is to develop a hybrid modeling framework that integrates XGBoost with CA to simulate the spatiotemporal distribution of PM₂.₅ concentrations across Iran. The study primarily focuses on methodological innovation rather than policy assessment, aiming to enhance prediction accuracy by incorporating both spatial neighborhood interactions and nonlinear relationships among multiple predictor variables derived from ground-based and satellite data. This hybrid XGBoost–CA approach is designed to localize the effects of independent variables, capture spatial heterogeneity, and provide a reliable, high-resolution PM₂.₅ estimation model suitable for large-scale environmental analysis. Materials and methods Study area Iran, with a population exceeding 87 million and a total area of approximately 1,648,000 square kilometers, is located in Western Asia between latitudes 25° and 40° North and longitudes 44° and 63° East. Geographically, Iran is situated in a region highly susceptible to dust storms. These dust storms, often originating from neighboring countries such as Iraq and Saudi Arabia to the west, and Afghanistan and Pakistan to the east, carry fine particles and pollutants that significantly deteriorate air quality. Despite Iran’s unique situation of being surrounded by some of the most polluted countries in the world, the fixed air quality monitoring stations established for regulatory purposes are limited in number and concentrated mainly in urban areas. As a result, the PM₂.₅ data they provide suffer from sparse coverage, uneven spatial distribution, and restricted geographic reach, making them inadequate for capturing large-scale pollution dynamics or producing reliable exposure assessments. In this study, ground-based observations were obtained from 162 air quality monitoring stations distributed across Iran, operated by the Department of Environment (DOE) of Iran30. However, their distribution is uneven, with a stronger concentration in urban areas compared to rural or remote regions. The spatial distribution of these stations used in the analysis is shown in Fig. 1. Detailed information about the PM₂.₅ monitoring stations is available through the DOE’s web portal (https://aqms.doe.ir). Geographic locations of ground-based monitoring stations measuring PM₂.₅ concentrations in Iran. Research data Table 1 summarizes the datasets used in this study, providing details on feature type, variable name, unit of measurement, spatial resolution, and data source. Most of the input datasets, including ground-based observations and satellite-derived variables, were originally available at daily or 48-hour temporal resolutions. However, satellite products such as Sentinel-5P and MAIAC AOD frequently suffer from missing observations due to cloud contamination and low-quality retrievals. To ensure spatial completeness and temporal consistency among all variables, the datasets were aggregated into monthly averages and resampled to a spatial resolution of 5 km. This temporal aggregation effectively mitigated data gaps in cloudy regions, reduced random noise, and provided continuous, gap-free fields suitable for large-scale spatiotemporal modeling of PM₂.₅ concentrations. While this harmonization was necessary due to the native resolution and availability of the satellite-based and global PM₂.₅ datasets, it inherently limits the model’s ability to represent short-term temporal variability and fine-scale spatial heterogeneity. As a result, the proposed XGBoost–CA model operates at a regional scale rather than resolving local pollution dynamics at sub-kilometer or daily intervals. Table 1Summary of targets and features used for Spatiotemporal modeling of PM₂.₅ in Iran.CategoryVariableUnitSpatial ResolutionData SourceTargetsGround-basedPM₂.₅µg/m³Point BaseNational Air Quality Monitoring System30Satellite DataPM₂.₅ global modelµg/m³5000 mThe ground-level PM₂.₅ global/regional model31FeaturesSatellite DataAOD-1000 mMODIS Satellite DataSatellite DataNDVI, AI, NO2, SO2, CO, O3, HCHOmol/m²3500 × 5500 mSentinel-5p Satellite DataSatellite DataDEMMeter30 mSRTM Administrative DataMeteorological DataAir TemperatureKelvin1000 mFLDAS-Central Asia Meteorological ModelHumiditykg/kg1000 mPressurePa1000 mRainfallkg/m²1000 mDew Point°C10,000 mMeteorological DataWind Speedm/s10,000 mFLDAS-Global Meteorological ModelGround-basedMain Roads DensityMeter5000 mOpenStreetMapsGround-basedDistance to Power PlantsMeter5000 mGeographical DataGround-basedPopulation DensityPeople/KM25000 mThe 2016 National Population and Housing CensusGround-basedLongitude, LatitudeDecimal DegreesPoint BaseGeographic Coordinates Summary of targets and features used for Spatiotemporal modeling of PM₂.₅ in Iran. This study utilizes three categories of data at the national level in Iran: ground-based, satellite, and meteorological data. Each category is described in detail in the following sections. Ground-based data Ground-based data related to PM₂.₅ concentrations were obtained from the National Air Quality Monitoring System, managed by the Department of Environment (DOE) of Iran30. PM₂.₅ concentration data, measured in micrograms per cubic meter (µg/m³), were recorded by ground monitoring stations and made available on the system as three-hour averages. As shown in Fig. 1, the distribution of these ground-based monitoring stations is uneven, with most stations concentrated in urban centers. This uneven spatial distribution limits the ability to provide comprehensive and continuous air pollution assessments across the country. Furthermore, the ground-level PM₂.₅ global/regional model31, which is a chemical transport model based on GEOS-Chem, estimates ground-level PM₂.₅ concentrations. The data obtained from this model were downloaded as monthly averages with a spatial resolution of 1 km for the entire study area throughout 2021. Figure 2 presents the PM₂.₅ concentrations estimated by this model for January 2021. Global PM₂.₅ model (January 2021). Given the spatial limitations and uneven distribution of air quality monitoring stations across Iran, the PM₂.₅ measurements obtained from these stations only cover limited areas within their reach. This limitation results in interpolation processes based solely on these measurements lacking acceptable accuracy. To address this challenge, the global PM₂.₅ model data were integrated with the PM₂.₅ measurements from the monitoring stations using the following empirical Eq. 1. In the above equation, r is equal to five (in kilometers) and represents the radius of influence, while d refers to the distance from air quality monitoring stations. Also, geographical data, including the distance to main roads, distance to power plants, population density, as well as latitude and longitude, were extracted using QGIS software based on Euclidean distance and kernel density methods. These data were obtained at a spatial resolution of 5 km (Fig. 3). Spatial distribution of roads density, population density, and distance to power plants across Iran. Satellite data A set of satellite-derived variables was used as independent predictors in the modeling framework (Fig. 4). Aerosol Optical Depth (AOD) was obtained from the MAIAC product, while trace gas measurements—including nitrogen dioxide (NO₂), absorbing aerosol index (AI), sulfur dioxide (SO₂), ozone (O₃), carbon monoxide (CO), and formaldehyde (HCHO)—were retrieved from the TROPOMI instrument onboard the Sentinel-5P satellite using the Google Earth Engine (GEE) platform. To ensure data quality, Sentinel-5P observations with cloud cover greater than 30% were excluded prior to generating monthly composite values. Spatial distribution of Sentinel-5P Products across Iran (January 2021). The Digital Elevation Model (DEM) extracted from the Shuttle Radar Topography Mission (SRTM), with a spatial resolution of 30 m for the entire territory of Iran, from the United States Geological Survey (USGS) website (Fig. 5). Digital elevation model (DEM) of Iran. Meteorological data The Famine Early Warning Systems Network Land Data Assimilation System (FLDAS) aims to generate high-quality outputs of land surface states and fluxes by utilizing observational datasets, advanced modeling techniques, and data simulation methods to support decision-making for FEWS NET32. FLDAS comprises three main modeling systems: (1) FLDAS-Global, (2) FLDAS-Central Asia, and (3) FLDAS-Forecast. FLDAS-Global operates twice a month (including an initial and a final run), producing monthly outputs at a spatial resolution of 0.1 degrees over a quasi-global domain33. FLDAS-Central Asia runs in near real-time (with a one-day delay) over the Central Asia domain, generating outputs with a spatial resolution of 0.01 degrees34. FLDAS-Forecast runs in Africa and the Middle East, generating monthly forecasts at a spatial resolution of 0.25 degrees for up to five months35,36. FLDAS comprises three main modeling systems: (1) FLDAS-Global, (2) FLDAS-Central Asia, and (3) FLDAS-Forecast. FLDAS-Global operates twice a month (including an initial and a final run), producing monthly outputs at a spatial resolution of 0.1 degrees over a quasi-global domain33. FLDAS-Central Asia runs in near real-time (with a one-day delay) over the Central Asia domain, generating outputs with a spatial resolution of 0.01 degrees34. FLDAS-Forecast runs in Africa and the Middle East, generating monthly forecasts at a spatial resolution of 0.25 degrees for up to five months35,36. In this study, two modeling systems were utilized to obtain monthly data throughout the year 2021. Ground surface air temperature, humidity, pressure, and precipitation data were obtained from the FLDAS-Central Asia system (Fig. 6), while dew point and wind speed data were retrieved from the FLDAS-Global system (Fig. 7). Spatial distribution of meteorological variables extracted from the FLDAS–Central Asia system (January 2021). Spatial distribution of dew point and wind speed extracted from the FLDAS–Global system (January 2021). Proposed method This section begins with a brief explanation of the basic concepts of CA and XGBoost, which serve as the foundations of the proposed model. Following that, the structure and execution process of the integrated XGBoost-CA model are described, including its input components and implementation flowchart. CA basic concepts Cellular Automata (CA) are discrete dynamic systems composed of a grid of cells, each holding a specific state at any given time. The state of each cell evolves through a set of transition rules that consider the states of neighboring cells. CA models are particularly effective for simulating spatial and spatiotemporal phenomena due to their local interaction mechanisms and parallel updating structure. The essential components of a CA model include: Cell Space: A two-dimensional grid that represents the geographical or spatial domain of the system. Cell Space: A two-dimensional grid that represents the geographical or spatial domain of the system. Neighborhood Configuration: Defines which surrounding cells influence the central cell. In this study, a 5 × 5 extended Moore neighborhood was employed, which includes 24 neighboring cells. Neighborhood Configuration: Defines which surrounding cells influence the central cell. In this study, a 5 × 5 extended Moore neighborhood was employed, which includes 24 neighboring cells. Transition Rules: Deterministic or probabilistic rules that update each cell’s state based on its current state and those of its neighbors. Transition Rules: Deterministic or probabilistic rules that update each cell’s state based on its current state and those of its neighbors. Boundary Conditions: Define how edge cells are handled. Custom boundary conditions were implemented to exclude missing data and to limit predictions within Iran’s national borders. Boundary Conditions: Define how edge cells are handled. Custom boundary conditions were implemented to exclude missing data and to limit predictions within Iran’s national borders. Stochastic Factor: A random factor was incorporated into the model to account for uncertainties and unknown disturbances in the PM₂.₅ estimation process. Stochastic Factor: A random factor was incorporated into the model to account for uncertainties and unknown disturbances in the PM₂.₅ estimation process. Overall, CA is well-suited for modeling environmental phenomena such as air pollution due to its ability to simulate local interactions across space and time. XGBoost basic concepts XGBoost (Extreme Gradient Boosting) is a high-performance machine learning algorithm based on the gradient boosting framework. It builds an ensemble of decision trees sequentially, where each new tree attempts to correct the prediction errors made by the previous trees. Key features of XGBoost include: Gradient Boosting Mechanism: Trees are added one at a time, each minimizing a defined loss function (e.g., mean squared error) by following the gradient of that loss. Gradient Boosting Mechanism: Trees are added one at a time, each minimizing a defined loss function (e.g., mean squared error) by following the gradient of that loss. Regularization: Incorporates both L1 (Lasso) and L2 (Ridge) regularization terms to prevent overfitting and improve model generalization. Regularization: Incorporates both L1 (Lasso) and L2 (Ridge) regularization terms to prevent overfitting and improve model generalization. Efficient Handling of Missing Data: XGBoost can natively process missing values and determine optimal splits accordingly. Efficient Handling of Missing Data: XGBoost can natively process missing values and determine optimal splits accordingly. Tree Pruning: Uses a depth-first strategy with a pruning mechanism based on a minimum loss reduction threshold to avoid over-complex trees. Tree Pruning: Uses a depth-first strategy with a pruning mechanism based on a minimum loss reduction threshold to avoid over-complex trees. Parallel and Distributed Computing: Designed for scalability and speed through parallel processing and optimization. Parallel and Distributed Computing: Designed for scalability and speed through parallel processing and optimization. In the context of the proposed model, XGBoost acts as a spatial rule engine that learns weighted relationships between independent variables and PM₂.₅ concentrations within the CA framework. XGBoost-CA model description The proposed method is a continuous CA model that utilizes the XGBoost ensemble learning algorithm as the core mechanism for estimating the spatial transfer potential of PM₂.₅ concentrations. This integrated model is referred to as the XGBoost-CA model. A key advantage of this approach is its ability to develop region-specific sub-models that address the heterogeneous behavior of independent variables across different areas of Iran. This helps capture the complex spatial dynamics and variability of PM₂.₅ concentrations. Figure 8 illustrates the implementation process of the XGBoost-CA model, as well as a comparative linear regression-based CA (LR-CA) model. Flowchart of the steps for executing and implementing the XGBoost-CA and LR-CA models. In general, the inputs of the proposed model are divided into three categories: dependent variables, independent variables, and CA variables. These three categories of inputs were prepared through a series of processes to be entered into the model, and a brief description of each is provided below. The dependent variable of this model corresponds to the measured PM₂.₅ data with a three-hour average obtained from the National Air Quality Monitoring System of the Department of Environment. After performing data engineering steps, the data were prepared as monthly averages throughout the year 2021 to serve as the dependent variable for the model. The independent variables of the proposed model were obtained from relevant sources. After processing steps such as masking cloud-covered values in some cases and interpolating missing data, they were integrated at a spatial resolution of 5 km and represented as monthly averages for the year 2021. Before finalizing the independent variables, a feature selection and multicollinearity analysis were performed to improve model efficiency and avoid redundant predictors. First, the pairwise Pearson correlation coefficients among all variables were computed, and those with strong correlations (|r| > 0.85) were excluded from the final feature set. Next, the Variance Inflation Factor (VIF) was calculated for the remaining predictors, and all variables with VIF values greater than 5 were removed. After this two-step screening process, 18 out of the initial 21 independent variables were retained for model development, ensuring that only statistically independent and relevant predictors were used. Overall, the monthly values of the dependent and independent variables were organized into a four-dimensional matrix of size 330 × 432 × 12 × 19. The first two dimensions represent the size of the cellular grid used, the third dimension corresponds to the months of the year, and the fourth dimension includes the dependent variable as the first entry, followed by the eighteen independent variables. The CA variables, which are also known as the fundamental parameters of the CA model, include neighborhood configuration, boundary conditions, a stochastic factor, and CA rules. The neighborhood used in this study is a five-by-five extended Moore neighborhood, allowing for the localization of the modeling process by employing 24 neighboring cells to estimate the value of the central cell. The boundary conditions were custom-designed so that cells were considered from the top-left corner to the bottom-right, excluding missing data in the estimation process and using the boundaries of Iran as a limiting factor to prevent estimations beyond national borders. To account for uncertainties and unknown disturbances in the PM₂.₅ concentration estimation process, a stochastic factor generating random values between zero and one was applied. The rule used in this study was a weighted averaging rule defined by XGBoost within the specified neighborhood. According to this rule, the eight cells within a Moore neighborhood of radius R = 1 were assigned a weight of 1.0, while the remaining cells within the neighborhood—excluding those with missing values—were assigned a lower weight of 0.9 during the averaging process used to estimate the value of the central cell. Evaluation criteria To evaluate the predictive performance of the proposed XGBoost-CA model, four widely accepted statistical metrics were employed: the coefficient of determination (R²), the root mean square error (RMSE), the mean absolute error (MAE), and the mean bias error (MBE). These metrics provide complementary insights into the model’s explanatory power and accuracy in estimating PM₂.₅ concentrations. Coefficient of determination (R²) The R² metric quantifies the proportion of the variance in the observed data that is explained by the model. It is defined as Eq. (1) Where yi is the observed PM₂.₅ concentration, is the predicted concentration, is the mean of the observed values, and n is the number of observations. An R² value of 1 indicates perfect agreement between predictions and observations, while a value of 0 implies that the model fails to explain any of the observed variance. Higher R² values thus correspond to better model performance in capturing the variability in PM₂.₅ concentrations. Root mean square error (RMSE) The RMSE measures the average magnitude of prediction errors, offering an absolute evaluation of model accuracy. It is computed as Eq. (2) This metric expresses the standard deviation of the residuals (prediction errors), reflecting how far predictions deviate from actual values on average. Lower RMSE values indicate higher predictive precision. Mean absolute error (MAE) The MAE provides a straightforward measure of the average absolute deviation between predicted and observed values. It is defined as Eq. (3) MAE is less sensitive to large errors than RMSE, offering a complementary perspective on overall model accuracy. Lower MAE values indicate better predictive performance. Mean bias error (MBE) The MBE evaluates the average bias of the predictions, indicating whether the model tends to overestimate or underestimate observations. It is computed as Eq. (4) Positive MBE values suggest that the model overestimates PM₂.₅ concentrations on average, while negative values indicate underestimation. An MBE close to zero reflects unbiased predictions. Together, R², RMSE, MAE, and MBE provide a robust framework for assessing and comparing the predictive performance of the XGBoost-CA model. Results In the following section, the findings and results of the study are presented exclusively based on the test and cross-validation data, which were not involved in the model training process during any month. The results are provided in the form of figures, tables, and charts. Hyperparameter tuning of the model In this study, a Grid Search algorithm with Cross-Validation (GS-CV) was employed to determine the optimal values for tuning the hyperparameters of the XGBoost model. The main hyperparameters tuned in this study are briefly described below. n_estimators defines the number of boosting trees in the ensemble. eta (learning rate) controls how much each tree contributes to the model, where smaller values yield smoother convergence. max_depth specifies the maximum depth of trees, regulating model complexity. min_child_weight determines the minimum sum of instance weights required in a child node, helping to prevent overfitting. subsample and colsample_bytree denote the fractions of samples and features randomly used for each tree, which improve generalization. Finally, lambda and alpha are regularization parameters that control model complexity and enhance stability. The cross-validation of hyperparameters was conducted using a five-fold approach, and the final results obtained from this process are presented in Table 2. In addition, to evaluate the model’s predictive performance over time, a temporal cross-validation scheme was employed. Specifically, a leave-one-month-out strategy was implemented, where the model was trained on data from eleven months and validated on the remaining month in a rotating manner. This design ensured temporal independence between training and testing samples and provided a robust assessment of the model’s generalization capability across different time periods. Table 2Results of the GS-CV algorithm in estimating the optimized Hyperparameters.Parameter NameData TypeValueDescriptionn_estimatorsint110Number of boosting treesetafloat0.1Learning rate of the modelsubsampleint1Subsample ratio of the training instancemax_depthint16Maximum depth of each treelambdafloat0.2L2 regularization termmin_child_weightint20Minimum sum of instance weight (hessian) needed in a childalphafloat1L1 regularization termcolsample_bytreefloat1Subsample ratio of columns when constructing each tree Results of the GS-CV algorithm in estimating the optimized Hyperparameters. Model results and evaluation To evaluate the results obtained from the proposed model, linear regression (LR), XGBoost, and a combination of LR and XGBoost methods with CA were employed. Overall, these methods were trained and tested crosswise on a monthly basis throughout the year 2021. The models were trained using data from 11 months of the year and tested using the remaining month. The results of the models in estimating monthly PM₂.₅ concentrations are presented in Table 3. These results are based on the statistical metrics of the adjusted coefficient of determination (R²-Adjusted) and root mean square error (RMSE). Table 3Results of the proposed model with comparison of other models.ModelMonthLRXGBoostLR-CAXGBoost-CARMSER²RMSER²RMSER²RMSER²January9.610.425.900.777.570.614.740.85February7.320.534.070.856.950.543.900.86March7.140.584.010.866.740.583.880.90April7.240.573.530.894.990.783.350.91May6.880.613.490.905.730.703.420.91June7.240.553.520.875.430.753.590.88July8.770.535.250.848.230.555.090.89August7.990.524.040.865.070.783.550.90September8.470.474.250.866.630.833.740.89October8.450.454.670.836.640.663.910.88November8.470.484.640.858.280.753.490.89December8.430.444.740.846.300.704.000.88 Results of the proposed model with comparison of other models. Table 3 shows that XGBoost-based models outperform the alternatives across the reported metrics. Consistent with these results, Fig. 9 shows that the proposed model further improves on the baseline XGBoost and achieves lower prediction errors and a stronger fit on the independent test set. Comparing the monthly performance of XGBoost and XGBoost-CA models by RMSE. The CA model has demonstrated remarkable performance by leveraging its capability to localize the behavior of independent variables, significantly increasing accuracy and reducing error. The effectiveness of the CA model’s performance is particularly evident when it transforms the LR model into an efficient and competitive model comparable to powerful machine learning models in modeling PM₂.₅ concentration at the national level in Iran (Table 3). Comparing the overall performance of models To evaluate the performance of the four models, the Adjusted R², RMSE, MAE and MBE metrics were calculated in all months. As illustrated in Fig. 10, the XGBoost-CA model consistently achieved the highest accuracy and stability among all models, demonstrating its superior capability in modeling monthly PM₂.₅ concentrations. As shown in Fig. 10, the MBE values estimated for the four models are approximately + 1.17, + 0.59, −0.37, and − 0.32 µg/m³ for the LR, LR–CA, XGBoost, and XGBoost–CA models, respectively. These results indicate that the proposed XGBoost–CA model exhibits nearly unbiased performance, while the baseline LR model shows a tendency toward overestimation. Overall performance of models based on R²-Adjusted, RMSE, MAE, and MBE. As shown in the monthly and overall performance of models, the proposed XGBoost-CA model achieved the best performance in modeling PM₂.₅ concentration compared to the other three models. This is evidenced by the highest values in the statistical metric of R²-Adjusted, as well as the lowest error values according to the RMSE criterion across all months of the year 2021. Comparing CA-based models in a Spatial context The spatial performance of the LR-CA and XGBoost-CA models across the territory of Iran based on the R²-Adjusted index is presented in Fig. 11. Spatial performance of the LR-CA and XGBoost-CA models across Iran based on the R²-Adjusted index. According to Fig. 11, the proposed XGBoost-CA model has demonstrated significantly strong performance in estimating PM₂.₅ concentrations at the national level in Iran. Only in certain areas of the northern strip of the country and a few limited regions in the western part did the model estimate PM₂.₅ concentrations with lower accuracy. Overall, the XGBoost-CA model shows a clear difference in modeling accuracy compared to the LR-CA model. To further assess the spatial reliability of the CA-based models, Fig. 12 presents a comparison between the observed and predicted PM₂.₅ concentrations. The predicted spatial patterns generated by the XGBoost–CA models closely align with the observed PM₂.₅ distribution, capturing major pollution hotspots such as the metropolitan areas of Tehran, Isfahan, and Ahvaz. The proposed XGBoost-CA model shows improved spatial smoothness and reduced local discontinuities compared with their LR-CA, indicating the positive effect of incorporating non-linear behaviors of XGBoost through the CA mechanism. This spatial consistency highlights the robustness of the hybrid modeling approach for mapping regional-scale air pollution patterns. Spatial maps comparing the observed and predicted PM₂.₅ concentrations for the CA-based models. Scatter plots of CA-based models The performance of these two models is comparatively presented on a monthly basis for the last four-month periods of 2021 in Fig. 13. As observed across all presented months in Fig. 13, the scatter plots of the XGBoost-CA model demonstrate a high correlation between the actual (measured) values and the predicted values. Additionally, the data points in all presented months exhibit significantly lower bias compared to the LR-CA model. Scatter plot showing the relationship between the actual PM₂.₅ and the predicted values from the LR-CA and XGBoost-CA models for the last four months of 2021 (September, October, November, and December). Discussion and conclusion Given the unique conditions of Iran regarding the widespread presence of particulate matter (PM) emission sources—such as consecutive droughts, excessive use of fossil fuels, the distribution of factories and thermal power plants across the country, and its specific geographical location surrounded by the most polluted countries in the region in terms of particulate matter emissions—the need for air pollution modeling in this region has become essential. Moreover, air pollution modeling, due to its specific nature, requires models that accurately capture the spatiotemporal behavior of independent variables. This study introduces two novel models with acceptable and distinct performances, named LR-CA and XGBoost-CA. These models aim to generate high-accuracy PM₂.₅ concentration maps over a wide and regional area covering the entire territory of Iran. The core of the proposed models is based on the CA framework. By localizing the behavior of LR and extreme gradient boosting (XGBoost) models, the CA model addresses the spatial heterogeneity of independent variables influencing PM₂.₅ concentration changes as the dependent variable. Consequently, the proposed models exhibit stronger performance in modeling PM₂.₅ concentration variations. Additionally, the CA model utilizes a stochastic factor, which introduces random values at the local level within each neighborhood window. This approach mitigates overfitting to a reasonable extent and ensures high accuracy and precision in predictions. Essentially, the LR and XGBoost models serve as generators of the potential transformation or transition for the CA. Spatial heterogeneity in regional air pollution modeling is a crucial factor that has received limited attention. To deal with spatial heterogeneity, CA is an appropriate model. On the other hand, XGBoost was selected as an appropriate choice for generating continuous air pollution transition potential maps for the CA model. This selection was based on XGBoost’s efficiency in handling large datasets and its ability to model nonlinear behaviors of the phenomena. As a result, the proposed model, by incorporating this capability into the air pollution modeling process, addresses a critical need on a large regional scale, marking the primary innovation of this study. To model monthly PM₂.₅ concentrations across Iran, four different models were used for performance comparison, utilizing available ground-based and satellite data. Among these, the primary proposed model, XGBoost-CA, consistently outperformed the other three models in all months. The XGBoost-CA demonstrated the highest predictive performance among all benchmark models, achieving an adjusted R² of 0.89, RMSE of 3.88 µg/m³, MAE of 2.33 µg/m³, and an MBE of −0.32 µg/m³ across the test folds. Additionally, the LR-CA model showed significant improvements in accuracy and performance compared to the base LR model. These results clearly indicate that incorporating spatially localized transition potentials within the CA structure substantially enhances model accuracy compared to standalone machine learning approaches. The consistency of these improvements across all months further highlights the robustness and stability of the proposed method. The results of this study are consistent with previous research that employed machine learning techniques to model PM₂.₅ concentrations; however, the proposed XGBoost-CA model offers significant improvements in both spatial resolution and predictive accuracy2,37. For example, a prior study that used a random-forest model reported an R² of 0.65 at the monthly scale across large, spatially heterogeneous regions37. In contrast, the integration of CA enabled our model to capture localized dynamics, while XGBoost provided a robust mechanism for handling complex nonlinear interactions among predictors. With an adjusted R² of 0.88, our hybrid model outperformed comparable studies and demonstrated enhanced capability in producing spatially continuous PM₂.₅ concentration estimates2. This performance gap underscores the value of combining spatial modeling frameworks with advanced ensemble learning algorithms in air quality assessment. This study provides several key insights that can inform environmental policy and urban air quality management in Iran. First, the demonstrated ability of the XGBoost-CA model to produce localized predictions underscores the importance of region-specific strategies for pollution control, rather than relying on uniform national policies. Second, the integration of remotely sensed satellite data with machine learning algorithms enables the generation of continuous pollution maps in areas where ground monitoring is sparse or unavailable, providing valuable support for early warning systems and public health planning. Lastly, the spatial modeling approach adopted here can be extended to other environmental variables or pollutants such as NO₂ or O₃, supporting a more comprehensive air quality management framework. Policymakers and environmental managers are encouraged to invest in both data infrastructure and computational modeling capacity to leverage such tools for targeted, evidence-based interventions. Future research should aim to address several key limitations identified in the current study. One major limitation is the spatial bias introduced by the concentration of air quality monitoring stations in urban areas, which restricts the representativeness of ground-truth data in rural or remote regions. This urban-centric sampling structure can introduce spatial bias into the training dataset, potentially reducing the model’s ability to fully characterize pollution dynamics in regions with limited or no ground observations. Although the integration of satellite-derived predictors and the global/regional PM₂.₅ model partially alleviates this imbalance by providing spatially continuous inputs, uncertainties may still propagate into the XGBoost-CA predictions in areas with low station density. Therefore, inferred concentrations in such regions should be interpreted with caution, and future work would benefit from expanding the monitoring network or incorporating additional independent validation sources to improve spatial representativeness and reduce this source of uncertainty. Another methodological limitation of this study concerns the temporal and spatial resolution of the datasets used for model development. Because both the fused ground–satellite PM₂.₅ product and several meteorological and geophysical predictors were only available as monthly composites at a 5 km spatial resolution, the XGBoost-CA model operates at a regional scale rather than capturing fine-scale temporal fluctuations or sub-kilometer spatial heterogeneity. As a result, rapid pollution dynamics associated with short-term meteorological events, localized emission spikes, or topographically driven gradients may not be fully reflected in the model outputs. Future studies could benefit from integrating higher-resolution datasets, such as daily PM₂.₅ retrievals or 1 km satellite products, to enhance the model’s capability for capturing local-scale variations. Moreover, incorporating meteorological forecast data and land-use dynamics may further improve model generalizability and allow for near-real-time prediction capabilities. Furthermore, an examination of the spatial performance maps (Fig. 11) reveals that the model exhibits relatively lower accuracy in certain regions of northern and western Iran. These areas are characterized by complex topography, strong local meteorological gradients, and sparse ground-based monitoring coverage, all of which can reduce the reliability of spatially continuous predictions. In the northern regions, proximity to the Alborz mountain range and frequent inversion events may introduce localized variability that is not fully captured by the monthly 5 km input datasets. Similarly, parts of western Iran experience intermittent dust intrusions and rapidly shifting synoptic conditions, contributing to higher prediction uncertainties. Acknowledging these region-specific limitations helps prevent overgeneralization and highlights the need for additional high-resolution observations to further enhance spatial model performance. These enhancements would significantly strengthen the model’s utility for public health applications and environmental policy planning. Author contributions S.F. developed the methodology, performed data preprocessing, and conducted the modeling experiments. J.M. conceived and supervised the study, contributed to the design of the research framework, and revised the manuscript. E.F. collected and processed satellite and ground-based data, and prepared the figures and tables. S.F. and J.M. wrote the main manuscript text. All authors reviewed and approved the final version of the manuscript. Data availability The data that support the findings of this study are available from the corresponding author, Dr. Jamshid Maleki, upon reasonable request. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement The data that support the findings of this study are available from the corresponding author, Dr. Jamshid Maleki, upon reasonable request.