Landscape Metrics as Ecological Indicators for PM₁₀ Prediction in European Cities

Abstract

Despite significant progress in recent decades, air pollution remains the leading environmental cause of premature death in Europe. Urban populations are particularly exposed to high concentrations of air pollutants, such as particulate matter smaller than 10 µm (PM₁₀). Understanding the spatiotemporal variations of PM₁₀ is essential for developing effective control strategies. This study aimed to enhance PM₁₀ prediction models by integrating landscape metrics as ecological indicators into our previous models, assessing their significance in monthly average PM₁₀ concentrations, and analyzing their correlations with PM₁₀ air pollution across European urban landscapes during heating (cold) and non-heating (warm) seasons. In our previous research, we only calculated the proportion of land uses (PLANDs), but according to our current research hypothesis, landscape metrics have a significant impact on PM₁₀ air quality. Therefore, we expanded our independent variables by incorporating landscape metrics that capture compositional heterogeneity, including the Shannon diversity index (SHDI), as well as metrics that reflect configurational heterogeneity in urban landscapes, such as the Mean Patch Area (MPA) and Shape Index (SHI). Considering data from 1216 European air quality (AQ) stations, we applied the Random Forest model using cross-validation to discover patterns and complex relationships. Climatological factors, such as monthly average temperature, wind speed, precipitation, and mean sea level air pressure, emerged as key predictors, particularly during the heating season when the impact of temperature on PM₁₀ prediction increased from 5.80% to 22.46% at 3 km. Landscape metrics, including the SHDI, MPA, and SHI, were significantly related to the monthly average PM₁₀ concentration. The SHDI was negatively correlated with PM₁₀ levels, suggesting that heterogeneous landscapes could help mitigate pollution. Our enhanced model achieved an R² of 0.58 in the 1000 m buffer zone and 0.66 in the 3000 m buffer zone, underscoring the utility of these variables in improving PM₁₀ predictions. Our findings suggest that increased urban landscape complexity, smaller patch sizes, and more fragmented land uses associated with PM₁₀ sources such as built-up areas, along with larger and more evenly distributed green spaces, can contribute to the control and reduction of PM₁₀ pollution.

1. Introduction

Air pollution is Europe’s largest environmental health risk, causing cardiovascular and respiratory diseases that impact health, reduce people’s quality of life, and cause preventable deaths. Urban air quality concerning particulate matter (PM) pollution, such as PM_2.5 and PM₁₀ (particles with a diameter of ≤10 μm) pollution, presents significant challenges to public health and the sustainability of urban development globally [1,2,3]. Particulate matter is a year-round pollutant and comprises a mixture of suspended solids and liquids in the air that are categorized based on size. The particles are produced from industrial, traffic, and geological sources (dust from roads or via chemical reactions in the atmosphere caused by released chemicals from motor vehicles or industrial sources). PM is not visible. Coarse particles (PM₁₀) are 2.5 to 10 μm in diameter. Their small size allows them to make their way to the air passages deep within the lungs where they may be deposited and result in adverse health effects [4]. PM₁₀ also causes visibility reduction. The European Environment Agency (EEA) plays a crucial role in monitoring and assessing air quality, publishing annual reports summarizing pollution levels including PM₁₀ and their effects [5]. EU directives underscore the importance of air quality modeling in managing and mitigating air pollution in European cities. Directive 2008/50/EC, a cornerstone of European air quality legislation, emphasizes the use of modeling techniques to assess and predict pollutant concentrations, particularly in areas where fixed measurements are limited. The directive recognizes the role of modeling in understanding the spatial distribution of pollutants, identifying pollution hotspots, and developing effective mitigation strategies [6]. Modeling plays a crucial role in informing policy decisions, evaluating the effectiveness of emission reduction measures, and predicting future air quality trends in European cities [7]. Accurate models for predicting PM₁₀ concentrations are essential to improving ecological management and guide urban planning efforts [1,8,9]. To understand the functioning of landscapes, landscape patterns must be considered, and indicators that address the spatial configuration of landscapes are therefore needed [10,11]. Our previous research focused on the effects of geological parameters such as soil texture, land use proportions (PLANDs), transportation network characteristics, and meteorological factors on PM₁₀ levels. We found that climatological parameters, including temperature and wind speed, were significant contributors to PM₁₀ variations [12]. Although the predictive precision of our models, as reflected in the adequate R² values, was promising, our objective was to investigate whether incorporating landscape metrics as additional ecological indicators can further enhance our PM₁₀ prediction models. We hypothesized that these metrics will provide deeper insights into the spatial characteristics of urban environments and how land use patterns influence air pollution.

According to the ’pattern and process’ paradigm (how landscape structure affects ecological processes) [13,14], the pattern of the landscape structure is both a determinant and an indicator of the ecological and environmental processes that occur in a given landscape [15,16,17,18,19,20,21,22,23]. According to [24], the characterization of the landscape structure should assess both the spatial proportion (composition) of the patch types and the spatial arrangement (configuration) of the land use and land cover (LULC) patches (Figure 1).

Different landscape metrics have diverse effects on atmospheric particulate matter concentrations at different scales [25,26]. Metrics such as the SHDI, Shape Index (SHI), and Mean Patch Area (MPA) have been widely used to assess the complexity and heterogeneity of urban landscape structures, factors that can affect the dispersion and concentration of air pollutants [25,26,27,28,29,30,31,32]. The SHDI is a key measure of landscape complexity, which reflects the diversity and abundance of different types of land use within a given area. A more diverse and heterogeneous landscape can influence the movement and accumulation of PM₁₀ by altering airflow patterns and pollutant dispersion [29,33,34,35,36].

Furthermore, the SHI and MPA provide insight into the dominance and scale of land use types and the complexity of the landscape that influence how pollutants are distributed throughout the landscape [29,31]. The contrast indexes, specifically the CI₁ metric, reflect the interaction between land use types and air pollutant dispersion, with more extensive edges potentially enhancing pollutant trapping or dispersion in certain zones with different land use metrics.

In our previous findings, we identified correlations between different PLANDs and monthly average PM₁₀ concentrations, categorizing land use types into three groups: those with a positive correlation (Group 1), a negative correlation (Group 2), and changeable correlation (Group 3) with PM₁₀ levels [37]. Building on these results, we chose specific landscape metrics for each group to explore their potential as indicators in our model. For Groups 1 and 2, we selected metrics such as contrast indexes and AREA_MN, given their relevance in areas where certain land use types positively or negatively influence PM₁₀ concentrations. For Group 3, our objective was to investigate whether landscape patterns without strong direct correlations could still play an indirect role in PM₁₀ dispersion through their interactions with adjacent land uses.

Recent advances in machine learning (ML) techniques offer powerful tools for analyzing large and complex datasets, allowing for the capture of complex relationships between ecological landscape indicators and air pollution levels [38,39,40,41,42,43]. The Random Forest (RF) algorithm, in particular, is well-suited for this type of ecological modeling due to its ability to handle high-dimensional data and complex non-linear interactions [44,45,46,47,48]. Our study used RF to analyze a dataset comprising 26 independent variables, including climatological parameters, geological factors, and landscape metrics, based on more than 3000 data pairs. The capacity of RF to assess the importance of variables is crucial to identifying which landscape metrics are key predictors of PM₁₀ concentrations. RF has been shown to be effective in similar environmental studies, demonstrating robustness and high accuracy [49,50,51].

The primary objectives of this study were to improve the accuracy of PM₁₀ concentration prediction models by incorporating landscape metrics as landscape ecological indicators, assess the importance of these metrics in relation to PM₁₀ pollution, and explore the correlations between landscape metrics and PM₁₀ concentrations in different seasonal periods, such as the heating and cooling seasons. By achieving these goals, our aim was to advance the understanding of urban ecology and contribute to sustainable air quality management practices in European cities.

2. Study Area and Methods

2.1. Data Sources

The 2018 monthly average PM₁₀ concentration data (µg/m³) were obtained from the European AQ Portal, which covers 1216 air quality (AQ) monitoring stations (1039 urban and 177 suburban) throughout Europe [52]. The landscape indicators surrounding the PM₁₀ monitoring stations were analyzed using the 2018 Copernicus Land Monitoring Urban Atlas Service of the European Union. This polygon-based LULC dataset, with a minimum mapping unit of 0.25 hectares, was published by the EEA under the Copernicus program on 16 April 2020, and updated on 16 July 2021 [53]. Based on previous studies, the aggregated Urban Atlas land use categories were applied [12,37]. Based on LUCAS topsoil data, the USDA soil texture class dataset was used to assess the physical properties of the European topsoil. It includes soil texture categories such as clay, silty clay, sandy clay loam, and more [54]. This dataset was published in 2015 by the European Soil Data Center (ESDAC) under the European Commission and Joint Research Center. Monthly data for 2018 were sourced from the Climate Data Store (CDS) to analyze meteorological conditions, such as the average monthly 10 m wind speed (m s⁻¹) (the horizontal speed of the wind, or movement of air, at a height of ten meters above the surface of the Earth), MSL (mean sea level) pressure (Pa) (the pressure (force per unit area) of the atmosphere on the surface of the Earth, adjusted to the height of the mean sea level), 2 m average temperature (the temperature of air 2 m above the surface of land, the sea, or inland waters), and total precipitation (mm) (the accumulated liquid and frozen water, comprising rainfall and snowfall, on the Earth’s surface) [55].

2.2. Spatial Data Analysis

2.2.1. Calculation of Landscape Metrics

Landscape patterns quantified using landscape metrics are useful for predicting and understanding PM₁₀ concentrations. Landscape metrics can provide information on the composition and configuration of the landscape (Figure 1).

Previous studies have shown that landscape metrics can provide a good explanation for PM concentrations, but effectiveness is affected by the spatial scale of the analysis [25,29,56]. Due to our previous studies [12,37,57], we applied different buffer zones with a radius of 1 km and 3 km centered on the 1039 European air quality monitoring stations in urban landscapes. Buffer zones with less than 80% land use coverage were excluded to focus on zones where a predominant land use type is present, improving the reliability of the correlation between land use and PM₁₀ concentrations. In areas with at least 80% coverage, the dominant land use provided a clearer and more consistent influence on air quality, enhancing the predictive precision of PM₁₀ concentration modeling.

In our analysis, to calculate the landscape metrics relevant to PM₁₀ prediction modeling (

Table 1. Brief description of the landscape metric indicators used for the analysis of urban landscape structure.

Landscape Index	Definition
Shannon Diversity Index (SHDI)	An index based on the relative area of each landscape type and the total number of LULC types. It is more sensitive to rare patch types than Simpson’s diversity index.
Proportion of LULC Categories (PLANDs)	This index reflects the percentage of the total area of a certain type of LULC patch in the entire landscape area, determining the basis for judging dominant landscape elements.
Mean Patch Area (MPA)	The average area of patches in the landscape of each type. Calculated for each LULC group (PM10 sources, barriers, and changeable).
Shape Index (SHI)	Describes the complexity of the shape of LULC patches within a landscape. It compares the perimeter of a patch with the perimeter of a standard shape (usually a square or circle) with the same area, thus giving insight into how irregular or fragmented a patch is. Calculated for each LULC group (PM10 sources, barriers, and changeable).
Contrast Class Edge (CCE)	Calculated as a percentage of the edge length of “PM10 source” LULC polygons shared with “PM10 barrier” LULC polygons.
Contrast Index 1 (CI1)	The edge length of the LULC polygons’ PM10 source showed a positive correlation with the concentration of PM10 in each buffer zone.
Contrast Index 2 (CI2)	The edge length of the LULC polygons of the ’PM10 barrier’ showed a negative correlation with the PM10 concentration divided by the area of each buffer zone.

), we used two main tools: the V-LATE extension in ArcMap and the ZonalMetrics tools in ArcGIS Pro 3.2.2. The SHDI was calculated for each buffer area to assess the diversity of land cover types surrounding the air quality monitoring stations. For class-level analysis, the MPA and SHI were determined for each of the three groups: barriers, sources, and changeable land uses. Land use polygons with centroids located within designated buffer zones were specifically selected to calculate the SHI. This strategy allowed us to include complete patches representing the landscape structure within each buffer, ensuring a more accurate connection between shape characteristics and PM₁₀ concentrations. By focusing only on polygons with centroids inside the buffer, we minimized issues associated with fragmented patches along the buffer edges. Contrast indices such as the Contrast Class Edge (CCE), CI₁, and CI₂ were calculated specifically between land uses that are sources or barriers of PM₁₀, allowing us to evaluate differences in PM₁₀ concentrations associated with the respective land use categories. This targeted approach provided information on the interactions between various landscape structures and air quality.

Three types of land cover were formed in the Urban Atlas according to the correlation of their area with PM₁₀ values. Among the LULC types of the Urban Atlas, some increased (sources) and others decreased (barriers) urban air PM₁₀ concentrations during all months of the year. The third so-called ‘temporary changeable’ group of Urban Atlas LULC categories was made up of LULC types that showed opposite relationship with PM₁₀ values during the heating and cooling periods (

Table 2. Land use groups categorized by their impact on monthly average PM₁₀ concentrations, as identified in our previous study [37].

Effect on PM10 Concentration	Urban Atlas LULC Categories Within the 1000 m Buffer Zone	Urban Atlas LULC Categories Within the 3000 m Buffer Zone
Source of PM10 Pollution	Vacant Lands Urban Parks Arable Lands Built-Up Areas	Railways Mine, Dump, and Construction Sites Vacant Lands Arable Lands Developing Areas
Barrier to PM10 Pollution	Railways Forests	Industrial Units Urban Parks Grasslands Forests Water
Changeable Effect on PM10 Pollution	Industrial Units Roads Grasslands Water	Roads

) [37].

The configuration of urban landscapes was characterized by averages of landscape metrics describing the MPA and complexity (SHI) of the LULC patches belonging to the three LULC groups above, and the length of their edges divided by the areas of the different buffer zones (CCE, CI₁, and CI₂). The compositional heterogeneity of urban landscapes (buffer areas around PM₁₀ monitoring stations) was characterized by the spatial PLAND of each Urban Atlas LULC category (PLAND) and the SHDI index of the heterogeneity of the LULC pattern.

2.2.2. Calculation of Soil and Meteorological Factors

Other independent variables such as soil texture, and meteorological variables—including wind speed (m/s), mean sea level pressure (Pa), total precipitation (mm), and mean temperature (°C)—were calculated or extracted from urban landscape 1000 m and 3000 m buffer zones, following prior findings on scale sensitivity [36]. The spatial analysis was performed using Arc Map 10.6.1, QGIS 3.22, and ArcGIS Pro 3.2.2.

2.3. Statistical Analyzes

2.3.1. Random Forest Modeling (RF)

We chose 18.3 °C as the threshold for the monthly average temperature in Europe to distinguish between heating and cooling periods for buildings, following the guidelines of previous studies [58,59,60,61,62]. The urban landscape dataset was divided accordingly. When the monthly mean temperature was below 18.3 °C, it was classified as the heating period and when it exceeded 18.3 °C, it was classified as the cooling period [12].

In this study, we applied a Random Forest regression model using the Databricks platform to predict PM₁₀ concentrations, incorporating a variety of ecological landscape indicators (26 independent variables). Random Forest was selected for its robustness in handling large datasets, its ability to capture complex relationships, and its built-in feature importance mechanism. This algorithm excels in scenarios with complex and nonlinear interactions between variables, which is particularly important in environmental studies where the interdependency between ecological indicators in the landscape can be complex [44]. Furthermore, Random Forest models tend to be less prone to overfitting, making them highly suitable for ecological research that involves diverse and complex datasets [63].

Cross-validation was used to enhance the reliability of the model by tuning the hyperparameters and evaluating the performance of different subsets of the data, ensuring that the model generalized well to unseen samples. This approach helps prevent overfitting while optimizing the predictive power of the model [64]. The Databricks platform, a cloud-based environment optimized for big data analytics, was leveraged for its seamless integration with Apache Spark, which allowed the use of distributed computing to process our large datasets efficiently. Apache Spark is an open-source distributed computing system designed for big data processing. It is known for its speed due to its in-memory data processing, scalability between computer clusters, and support for various tasks such as data streaming and machine learning [65]. This integration significantly reduces the computational time and improves scalability, which is critical in studies that require the analysis of large ecological datasets. Databricks also offers an interactive and collaborative environment that supports Python-based libraries such as Scikit-Learn, making it an ideal choice for iterative model building and hyperparameter tuning [66].

2.3.2. Model Development

We used a 70/30 split of training/test datasets, following a similar approach to that used by [12] to ensure consistency in model development and validation. This split allowed us to evaluate model performance on both the training and test datasets, providing a balanced assessment of the model’s predictive power. The Random Forest model was implemented using the Scikit-Learn library in Python (version 3.9.19), a well-established tool for machine learning due to its flexibility and reliability [66]. Scikit-Learn is used in ecological modeling and environmental studies, providing reliable and efficient solutions for handling large datasets and complex models [67], as well as supporting high-performance machine learning techniques for classification, regression, and clustering tasks [68]. Random Forest is a machine learning technique that is particularly well-suited to handling multicollinearity, a condition where independent variables in a dataset are highly correlated. Unlike linear regression models, which can suffer from inflated standard errors and misleading statistical inferences due to multicollinearity, Random Forests are less affected because of their tree-based structure and random feature selection process during training. Random Forests build multiple decision trees, each using a random subset of features, which reduces the influence of correlated predictors on the model’s performance [44,69]. To optimize the model, we performed 5-fold cross-validation, which is a robust method for tuning hyperparameters like ‘max_depth’ and ‘n_estimators’. Cross-validation ensures that the model avoids overfitting and generalizes well to unseen data [70]. After selecting the best hyperparameters, we evaluated the performance of the model using the coefficient of determination (R²) for both the training and test sets. Strong agreement between the R² scores of these sets indicated minimal overfitting, demonstrating the strong ability of the Random Forest model to generalize to new data [44]. This consistent performance confirmed the robustness of the model in predicting PM₁₀ concentrations.

2.3.3. Model Evaluation

We evaluated the accuracy of all models using key performance metrics, including the coefficient of determination (R²), RMSE (Root Mean Square Error), and MAE (Mean Absolute Error). The high R² values in both the training and the test datasets further validated the precision and reliability of the Random Forest model in predicting PM₁₀ concentrations. Additionally, the inclusion of the RMSE and MAE provided a more comprehensive assessment of model performance, quantifying the average magnitude of the prediction errors and ensuring robustness [71,72,73]. These metrics not only confirmed the model’s predictive accuracy but also facilitated meaningful comparisons with our previous modeling approaches [12], further strengthening the reliability of our findings.

2.3.4. Feature Importance

We assessed the contribution of each variable using the importance attribute of the characteristics of the best-performing model. This method provides a ranking of the predictor variables based on their influence on the model’s predictions. The results were normalized into percentages, highlighting the relative importance of each landscape index, environmental factor, and meteorological factor in explaining the variations of PM₁₀.

2.3.5. Spearman Coefficient Correlation

The Spearman correlation coefficient was applied to explore the relationships between environmental variables and the monthly average PM₁₀ concentrations in different periods and buffer zones. The dataset, initially stored in a Spark DataFrame, was processed in Databricks and converted to a Pandas DataFrame for ease of manipulation. The categorical variables were transformed using one-hot encoding to allow for numerical comparisons. Spearman correlation was chosen for its ability to measure both linear and non-linear monotonic relationships. Correlation coefficients and p-values were calculated to evaluate the strength and significance of each variable’s impact on PM₁₀ concentrations, helping to identify key ecological and environmental drivers of the landscape at the different spatial scales. To address potential multicollinearity among the independent variables identified through Spearman correlation, Variance Inflation Factor (VIF) values were calculated. This step ensures that the predictor variables used in the modeling process are independent and do not exhibit problematic levels of multicollinearity. The VIF analysis was conducted in Python (Version 3.9.19) using the processed Pandas DataFrame. The results, presented in

Table A2. Variance Inflation Factor (VIF) of Independent Variables Investigated for Correlation with PM₁₀ Concentration Using Spearman Correlation Across Cooling and Heating Periods at 1000 m and 3000 m Scales in Urban Landscapes.

Independent Variable	VIF
	Cooling Period		Heating Period
	1000 m	3000 m	1000 m	3000 m
MPS of PM10 Source LULC Categories	1.911885	3.028666	1.962884	3.302535
MPS of PM10 Barriers	4.480291	2.597356	4.250656	3.024115
MPS of LULC Categories with Changeable Effect on PM10	2.393220	1.234664	2.180816	1.229374
SHI of PM10 Source LULC Categories	1.250877	2.586977	1.197091	2.005047
SHI of PM10 Barriers	1.662593	1.385610	1.600599	1.365938
SHI of LULC with Changeable Effect on PM10	1.867300	1.215057	1.828440	1.243012
SHDI	1.991132	1.754137	1.990479	1.598674
CI1	4.958673	3.818949	5.124483	4.751261
CI2	7.327620	8.215971	8.187491	8.902750
CCE	8.344573	8.360450	7.445902	8.745668
Air Pressure	1.264055	1.281203	1.070224	1.095389
Total Precipitation	1.370848	1.500650	1.174197	1.181026
Temperature	1.417126	1.416542	1.123252	1.149163
Wind Speed	1.210010	1.430629	1.101249	1.259422
PLAND of Built-Up Area	8.380624	5.520451	8.361314	6.866883
PLAND of Industrial Unit	7.301525	2.820040	8.230368	3.356350
PLAND of Roads	4.815526	3.281017	3.957730	3.625052
PLAND of Railways	2.701912	1.511417	2.355532	1.600790
PLAND of Mine, Dump, and Construction Sites	1.139129	1.331924	1.174259	1.340900
PLAND of Vacant Lands	1.473016	1.254967	1.335742	1.277378
PLAND of Urban Parks	8.808699	2.836596	6.219159	3.207821
PLAND of Arable Lands	8.375765	6.910781	4.691872	7.744227
PLAND of Grasslands	6.632918	3.712766	4.380273	4.683871
PLAND of Forests	6.945185	3.460085	6.028392	4.489963
PLAND of Water	2.848026	2.002094	2.481236	2.822331

, show that all the VIF values were below 9, confirming that multicollinearity was not a concern in our analysis. This additional step strengthens the reliability of our modeling approach.

3. Results

3.1. Improvements to the Model’s Performance

The performance of the RF model was evaluated using a combination of landscape metrics, climatological variables, and soil texture categories. The performance of the RF model was improved by incorporating new landscape metrics along with the existing variables.

Table 3. Random forest modeling performance after adding the landscape metrics and considering all variables on the Databricks platform using Python 3.9.19.

Buffer Zone	Period	R2 Training Set (70%) Previous Study CRF Modeling	R2 Training Set (70%)	R2 Test Set (30%)	Best Hyperparameters
					Max_Depth	N_Estimators
1000 m	Cooling	0.36	0.524	0.585	30	140
3000 m	Cooling	0.41	0.481	0.508	50	70
1000 m	Heating	0.57	0.593	0.619	20	140
3000 m	Heating	0.61	0.652	0.666	45	140

summarizes the R² values of the model for the training and test sets, as well as the hyperparameters that were used to achieve the best performance.

The addition of landscape metrics led to a noticeable improvement in the model’s predictive capability compared to that seen in the previous study [6]. For example, in the 1000 m buffer zone during the cooling period, the R² for the training set increased from 0.36 (in the previous study) to 0.52, while the R² for the test set reached 0.58. Similarly, in the 3000 m buffer zone during the cooling period, the R² value improved from 0.41 to 0.48 for the training set and 0.50 for the test set, indicating better model performance. During the heating period, the model also showed a significant improvement. In the 1000 m buffer zone, the R² of the training set increased from 0.57 to 0.593, with the test set showing an R² of 0.619. The best performance was observed in the 3000 m buffer zone during the heating period, where the R² for the training set improved from 0.61 to 0.652, and the test set achieved an R² of 0.666. To enhance model validation, the RMSE and MAE metrics were calculated, showing clear improvements compared to the previous study. Incorporating landscape metrics significantly reduced prediction errors. For example, in the 1000 m buffer zone during the cooling period, the RMSE decreased from 4.84 to 3.72 and the MAE decreased from 3.58 to 2.74. Similar improvements were observed across all buffer zones and periods. These findings confirm the enhanced predictive accuracy of the RF model. The detailed metrics are presented in

Table A1. Random Forest PM₁₀ Prediction Modeling Performance Validation Metrics, Incorporating Landscape Metrics Across Cooling and Heating Periods in Urban Landscapes.

Buffer Zone	Period	RMSE Test Set (30%) Previous Study CRF Modeling	MAE Test Set (30%) Previous Study CRF Modeling	RMSE Test Set (30%) RF Modeling	MAE Test Set (30%) RF Modeling
1000 m	Cooling	4.84	3.58	3.72	2.74
3000 m	Cooling	4.63	3.35	4.02	2.95
1000 m	Heating	6.83	4.92	5.97	4.29
3000 m	Heating	6.64	4.70	5.96	4.18

.

3.2. Analysis of Independent Variable Importance and Spearman Correlation Coefficient

3.2.1. Heating Period

During the heating period within the 1000 m buffer zone, the temperature emerged as the most critical variable that influenced the concentration of PM₁₀, with an importance of 20.08%. This was closely followed by total precipitation (11.99%) and wind speed (11.10%). These variables exhibited significant negative correlations with the PM₁₀ concentration, suggesting a reduction in PM₁₀ levels with higher values of these climatic factors. Specifically, temperature showed a Spearman correlation coefficient of −0.197 (p < 0.001), while total precipitation and wind speed had coefficients of −0.273 (p < 0.001) and −0.231 (p < 0.001), respectively. These findings, illustrated in Figure 2, indicate that higher temperatures, increased precipitation, and stronger winds are associated with decreased PM₁₀ concentrations in this buffer zone during the heating period. The SHI for LULC Group 3 which comprises LULC classes that demonstrated varying impacts on the monthly average PM₁₀ concentration in our earlier research [37] shows that it contributed to PM₁₀ predictions by up to 5.23% and exhibited a positive correlation with PM₁₀ levels.

In the 3000 m buffer zone during the heating period, temperature continued to exert the greatest influence on the PM₁₀ concentration, with an increased importance of 22.46%. The second and third most influential variables were wind speed (14.58%) and total precipitation (11.77%). All three variables maintained negative correlations with the PM₁₀ concentration, underscoring their importance in reducing particulate levels during this period. The temperature correlation with the concentration of PM₁₀ was −0.24 (p < 0.001), while the correlations of wind speed and total precipitation were −0.272 (p < 0.001) and −0.299 (p < 0.001), respectively. These negative associations, represented in Figure 3, emphasize the importance of these variables, especially precipitation, in the potential dispersal or mitigation of PM₁₀ levels during the heating period.

3.2.2. Cooling Period

During the cooling period in the 1000 m buffer zone, total precipitation became the main factor influencing the PM₁₀ levels, with an importance of 9.40%. Three landscape metrics emerged as notable factors, with 7.55%, 7.51%, and 6.68% importance, respectively. The MPS of LULC categories with a changeable effect on PM₁₀ levels and the SHI of LULC categories with a changeable effect on PM₁₀ levels exhibited significant positive correlations with the PM₁₀ concentration, with coefficients of 0.3 (p < 0.001) and 0.337 (p < 0.001), respectively. In contrast, total precipitation showed a negative correlation with PM₁₀ levels (−0.098, p < 0.001), suggesting a mitigating effect. Furthermore, a greater diversity in LULC categories was associated with a reduction in monthly average PM₁₀ pollution within the 1000 m buffer zones. These findings are illustrated in Figure 4.

In the 3000 m buffer zone during the cooling period, air pressure emerged as the most significant variable associated with the PM₁₀ concentration, accounting for 12.76% of the importance. This was followed by total precipitation (9.55%) and the SHI of the PM₁₀ source LULC categories (6.04%). The air pressure exhibited a negative correlation with PM₁₀ (−0.104, p < 0.001), suggesting that a higher air pressure could help disperse PM₁₀. However, the SHI of the PM₁₀ source LULC categories showed a negative correlation coefficient of −0.212 (p < 0.001), indicating that a higher SHI for these types of LULC categories can contribute to higher levels of PM₁₀ under certain conditions. Figure 5 provides a visual representation of these associations, highlighting the important effects of air pressure and landscape metric variables on PM₁₀ pollution levels during cooling periods.

3.2.3. Effect of Different Soil Texture Categories

The findings revealed that the texture of the soil exerted a stronger influence during the heating period and in areas closer to the AQ stations. The Spearman correlation coefficients indicated similar significance within the 1000 m and 3000 m buffer zones around the stations. During the cooling period, soil textures, such as silty clay, were positively correlated with PM₁₀ concentrations, suggesting that these soil types contribute to higher PM₁₀ levels. On the contrary, loam, sandy loam, and silt loam showed negative correlations, which implies that these soil textures were associated with lower PM₁₀ concentrations. During the heating period, the correlation patterns changed. The sandy clay and loam showed significant negative correlations with PM₁₀ concentrations, indicating a reduction in PM₁₀ levels. However, loamy sand, silt loam, and silty clay loam were positively correlated with PM₁₀ levels, suggesting that they exacerbated pollution levels.

4. Discussion

4.1. Model Improvement

The improvement in model performance can be attributed to the inclusion of new landscape metrics, which served as important ecological landscape indicators. Metrics such as shape indices, patch sizes, and diversity indices offered a more detailed view of LULC patterns and their relationship with PM₁₀ concentrations. By incorporating these ecological indicators of the landscape, the model was able to capture more complex spatial relationships, leading to better predictions of levels of air pollution. Compared to the previous approach, the current model provided a more accurate representation of the ecological landscape and its influence on air quality. The analysis of feature importance for both the cooling and heating periods highlighted significant changes in the key variables influencing PM₁₀ concentrations in urban landscapes, and compared the results of the previous study [12] with those of the current study that incorporates landscape metrics as landscape ecological indicators. Furthermore, the spatial scale of the analysis, represented by the buffer zones, significantly affected the relationship between the landscape metrics and PM₁₀ concentration, which is comparable to that of [25].

Table 4. Comparing the current study (improved Random Forest regression PM₁₀ prediction modeling by adding landscape metrics) to the previous study (conditional inference regression Random Forest (CRF) PM₁₀ prediction model), showing variables with more than 5% importance in urban landscapes.

Cooling Period		Heating Period
1000 m	3000 m	1000 m	3000 m
Previous study
Soil texture (20.75%) Roads (11.77%) Temperature (10.26%) Forest (7.94%) Total precipitation (6.62%)	Forests (15.44%) Soil texture (12.83%) Vacant land (8.25%) Wind speed (7.49%) Temperature (7.04%) Total precipitation (6.43%) Roads (5.47%)	Temperature (26.33%) Wind speed (20.43%) Total precipitation (12.76%) Soil texture (8.64%)	Total precipitation (24.02%) Wind speed (21.28%) Temperature (12.79%) Soil texture (6.45%)
Current study
Total precipitation (9.39%) MPS of LULC categories with changeable effect on PM10 (7.55%) SHI of LULC categories with changeable effect on PM10 (7.51%) SHDI (6.67%) MSL air pressure (6.48%) Temperature (6.36%) Wind speed (5.76%) Urban parks (5.02%)	MSL air pressure (12.76%) Total precipitation (9.55%) SHI of LULC with changeable effect on PM10 (6.04%) Temperature (5.80%) Wind speed (5.75%)	Temperature (20.08%) Total precipitation (11.99%) Wind speed (11.0%) MSL air pressure (8.38%) SHI of LULC with changeable effect on PM10 (5.23%)	Temperature (22.46%) Wind speed (14.58%) Total precipitation (11.77%) MSL air pressure (7.82%)

shows the main variables with more than 5% importance in both studies, categorized by buffer zones (1000 m and 3000 m) during the cooling and heating periods.

In the cooling period in the 1000 m buffer zone, the previous study identified soil texture (20.75%) and roads (11.77%) as the most important variables, while in the current study, total precipitation (9.39%) and the MPS of LULC categories with a changeable effect on PM₁₀ (7.55%) ranked the highest. Temperature appeared in both studies but was more important in the previous study (10.26%) compared to the current study (6.36%). Forests were important in the previous study (7.94%), but not in the current one. In the 3000 m buffer zone, the previous study highlighted forests (15.44%) and soil texture (12.83%) as the main variables, while the current study showed that MSL air pressure (12.76%) and total precipitation (9.55%) were the most significant. Temperature and wind speed were present in both studies but with slight differences in their degrees of importance.

In the heating period in the 1000 m buffer zone, the temperature had the highest importance in both studies, with 26.33% in the previous study and 20.08% in the current study. Wind speed and total precipitation were also key variables, and both studies showed similar importance levels. The current study added MSL air pressure (8.38%) and the SHI of LULC categories with a changeable effect on PM₁₀ (5.23%) to the main variables, which were not present in the previous study. In the 3000 m buffer zone, total precipitation and wind speed were important in both studies, although the current study assigned lower percentages (11.77% and 14.58%, respectively) compared to the previous study. The temperature also remained significant, with a slight reduction from 12.79% in the previous study to 22.46% in the current one. The MSL air pressure was identified as a significant variable only in the current study (7.82%).

4.2. Effects of Climatological Variables

Our study confirmed the significant influence of temperature, wind speed, total precipitation, and air pressure on PM₁₀ concentrations, highlighting their importance in predicting and managing PM₁₀ levels, especially during the heating season. In particular, temperature exerted a particularly pronounced effect on PM₁₀ levels during the heating period, with its predictive influence nearly quadrupling compared to the cooling period (ranging from 5.80% in the cooling period within a 3000 m radius to 22.46% during the heating period within the same distance). This substantial increase indicates that temperature fluctuations during the heating season may enhance the retention of pollutants, likely due to limited atmospheric mixing and reduced dispersal. In contrast, during the cooling period, total precipitation emerged as the dominant predictor, suggesting that precipitation plays a more active role in mitigating PM₁₀ concentrations through washout effects when heating sources are minimal.

Additionally, we identified distinct seasonal correlation patterns: a negative correlation between temperature and PM₁₀ concentration during the heating season, likely due to heightened emissions from heating sources in colder conditions, and a positive correlation during the cooling season due to enhanced evaporation and the formation of secondary particles [74,75]. For example, a study in Switzerland found a positive correlation between high afternoon temperatures and PM₁₀ concentrations, particularly in summer [74]. This suggests that as temperatures increase, the evaporation of volatile compounds increases, contributing to the formation of secondary aerosols that can increase PM₁₀ levels. On the contrary, low temperatures, particularly during winter, can also lead to increased PM₁₀ concentrations. This is attributed to the condensation of volatile compounds and the increase in residential heating emissions, particularly from wood burning [76,77,78]. The Swiss study also found a positive correlation between low temperatures in winter and high PM₁₀ concentrations [74]. This finding highlights the need to consider seasonal variations when examining the impact of temperature on PM₁₀ levels [79]

Our findings also revealed a consistent negative correlation between the monthly average total precipitation and monthly average wind speed with monthly average PM₁₀ concentrations in both seasonal periods in 2018. These results suggest that higher wind speeds and precipitation rates are effective in reducing PM₁₀ levels throughout the year, either by dispersing particulate matter or by facilitating its removal from the atmosphere. Higher wind speeds promote PM₁₀ dispersion and dilution, leading to lower concentrations. This is supported by previous studies such as [12,77,78,79,80,81,82,83,84]. Precipitation, especially rainfall, is widely recognized as a significant factor in reducing PM₁₀ concentrations [12,77,80,85,86,87]. Rainfall effectively removes PM₁₀ from the atmosphere through wet deposition, leading to lower PM₁₀ levels.

According to our findings, there was a significant negative correlation between the monthly average air pressure at sea level and the monthly average PM₁₀ concentration, especially during the cooling period. In contrast, several studies have indicated a positive correlation between air pressure and PM₁₀ concentrations [75,79,81,88] because traditionally, high atmospheric pressure is often associated with stable atmospheric conditions, which can trap pollutants near the surface, leading to higher PM₁₀ concentrations. The observed negative correlation during the cooling period suggests that other factors may be at work, potentially overriding the traditional influence of high pressure. Cooling periods are generally characterized by transitional weather patterns, which can significantly influence atmospheric dynamics, such as wind pattern mixing and precipitation effects. Cooling periods often experience changes in wind patterns, which can enhance the vertical mixing and dispersion of pollutants, even under high pressure conditions. The increase in wind speed during these periods could counteract the trapping effect of high pressure, leading to lower concentrations of PM₁₀. For example, Ref. The importance of wind patterns and their influence on the transport and dispersion of pollutants, particularly during seasonal transitions [81]. Furthermore, cooling periods, depending on the specific geographic location, could also be associated with increased precipitation, which can further contribute to the removal of PM₁₀ from the atmosphere through wet deposition. Analyzing long-term meteorological data and PM₁₀ measurements for the study area could provide a comprehensive understanding of the seasonal and temporal variations in this relationship. Understanding these relationships is crucial to developing effective strategies to manage and mitigate PM₁₀ pollution, ultimately improving air quality and protecting public health.

4.3. Effects of Landscape Metrics

Our results proved that landscape metrics significantly affect PM₁₀ concentrations; these effects varied across the heating and non-heating periods and different buffer zones. For instance, forests and urban parks played an essential role in reducing PM₁₀ levels. Specifically, we observed a strong negative correlation between the PLAND metric for forests (in both buffer zones) and urban parks (within the 3000 m buffer zone only) during both the heating and cooling periods. This indicates that expanding vegetated areascould effectively reduce PM₁₀ concentrations, which is in agreement with the results reported in [89]. The effectiveness of green spaces as PM filters is influenced by their shape, size, and configuration. Green areas with large surface areas (high AREA_MN) and complex shapes (high Shape Index) are more effective at capturing particulates [29]. This aligns with our current findings, which demonstrated a significant negative correlation between the SHI of LULC types acting as barriers to PM₁₀ pollution and distances of 1000 m from air quality monitoring stations during both the cooling and heating periods.

Increasing the area of grassland patches (measured by AREA_MN) was beneficial for PM₁₀ reduction in both buffer zones during the cooling period, because grassland acts as effective natural filters for air pollutants. The vegetation in grassland areas captures particulate matter, including PM₁₀, through processes such as sedimentation and deposition [90].

The analysis of land use metrics, particularly the SHI and MPA, across different groups of LULC categories (G1, G2, and G3) revealed distinct patterns in their correlation with monthly average PM₁₀ concentrations. Our findings demonstrated that for Group 3 (LULC categories with changeable effects on PM₁₀ concentrations), both the SHI and MPA had substantial impacts on PM₁₀ concentrations, particularly during the cooling (non-heating) period within a 1000 m buffer zone from the air quality (AQ) stations.

The positive correlation between PM₁₀ levels and the SHI of LULC categories with a changeable effect on PM₁₀ suggests that a more complex and fragmented urban structure leads to higher PM₁₀ pollution and is supported by several studies that highlight the impact of landscape configuration on PM concentrations. For instance, a study on green space buffers (GBSs) in Shanghai discovered that a higher landscape SHI was correlated with a longer diffusion distance for atmospheric particles such as PM_2.5 and PM₁₀. This indicates that more complex shapes were less effective in reducing PM pollution, potentially allowing for a wider spread of pollutants [91]. Studies have emphasized that a fragmented landscape can hinder the effectiveness of green spaces in mitigating air pollution, allowing pollutants to accumulate in urban areas. Research in Shenzhen [26] indicated that increasing the percentage of land and edge density in built-up areas contributes to higher PM₁₀ concentrations. Furthermore, a study in Seoul [92] found that areas with densely built-up areas and high traffic volumes, often characteristic of fragmented landscapes, experienced prolonged periods of high levels of PM_2.5. These findings support the idea that a more complex and fragmented urban structure, represented by a higher SHI in our analysis, can lead to higher PM₁₀ pollution. Additionally, during heating months (when temperatures drop below 18.3 °C), the demand for heating increases emissions and complex urban patches tend to trap pollutants due to reduced air circulation. In non-heating months (above 18.3 °C), when emissions are lower, the complex shapes still hinder pollutant dispersion. This means that irregular LULC patterns in urban areas can restrict airflow, causing pollutants such as PM₁₀ to accumulate, especially when heating demands add to pollution. More information using data on air circulation patterns and pollutant dispersion would improve our understanding of these mechanisms. Factors such as building height and compactness, wind direction, and temperature inversions, which are discussed in the cited references, can significantly influence air circulation and pollutant dispersion [26,34,93,94].

The findings indicate that there is a positive correlation between PM₁₀ levels and the MPA for Groups 1 and 3. Group 1, identified as sources of PM₁₀, likely includes areas with high emissions. Larger patches in these high-emission areas appeared to concentrate PM₁₀ pollution, particularly during the heating season, when emissions were high due to heating demands. This is consistent with the understanding that emission sources and urban morphology can interact to influence air quality. A study in Poland [29] found that the highest concentrations of PM₁₀ occurred in autumn and winter, suggesting the importance of residential and traffic sources. PM concentration was observed to be higher inside urban landscapes than outside, particularly for PM₁₀, highlighting the role of urban emissions in pollution levels. The increased concentration of PM₁₀ in larger patches within the LULC categories that are sources of PM₁₀ pollution during the heating season is consistent with the understanding that heating demands contribute significantly to PM emissions. A study in Weifang, China [95], found that PM_2.5 pollution was exacerbated in winter due to increased emissions from coal burning for heating. Research in Poland [29] also highlighted the influence of household heating using solid fuels on PM₁₀ levels, particularly during the colder months.

For LULC categories with a changeable effect on PM₁₀, larger patches were associated with an increase in PM₁₀ during the heating period within a radius of 1000 m to 3000 m, and within 1000 m during the cooling period. This pattern suggests that the relationship between the MPA and PM₁₀ concentration is complex and is influenced by factors beyond emissions alone. One of the possible reasons could be that compact urban patterns in Group 3 contributed to pollutant accumulation, even in periods of lower emissions, due to limited air dispersion in dense environments. Research in Nanchang, China [96], found that PM_2.5 concentrations were higher in compact urban areas compared to open areas, indicating that the density of buildings can influence air quality. Our results highlight the importance of considering scale when analyzing the relationship between urban form and air quality.

On the contrary, Group 2 (railways and forests within 1000 m and industrial units, urban parks, grasslands, and water bodies in the 3000 m buffer zone), as barriers to PM₁₀ pollution, demonstrated a negative correlation between PM₁₀ concentrations and the MPS, with larger patches acting as barriers to PM₁₀ pollution within 1000 m of the AQ stations in both the heating and cooling seasons. These areas probably consist of green spaces or low-emission zones that buffer pollutants and promote cleaner air. Larger patches in the LULC PM₁₀ barrier may have provided a dispersive effect, allowing PM₁₀ to dissipate more easily and decreasing pollutant levels near the AQ stations. This suggests that integrating larger green or low-emission areas around urban centers can be effective in managing air quality, particularly near emission sources. Forests play a vital role in PM₁₀ mitigation and their effectiveness increases with greater coverage and connectivity [25,97]. The presence of forests within both 1000 m and 3000 m radii could contribute significantly to reducing PM₁₀ concentrations. According to [94], the negative regression coefficient of the building agglomeration index within 1000 m in winter suggested that increased building density, which might indirectly indicate reduced forest coverage, was associated with higher PM₁₀ concentrations. Moreover, the combined presence of forests, urban parks, and grasslands can create a synergistic effect, maximizing PM₁₀ mitigation potential [94,95]. Larger and more connected patches of these green spaces within both radii could effectively act as barriers to PM₁₀ pollution.

Among landscape metrics, the Shannon diversity index (SHDI) plays an important role in PM₁₀ prediction within 1000 m buffer zones during the cooling period. SHDI reflects landscape complexity and balance in the distribution of various LULC patch types in the landscape [31,98]. Although some references discuss the relationship between landscape metrics such as the SHDI and PM_2.5 concentrations [31,34,89,93], Other authors have not investigated the relationship between the Shannon diversity index and PM₁₀ in the literature.. According to our results, the Shannon diversity index, a measure of landscape complexity, was negatively correlated with the monthly average PM₁₀ concentration in all buffer zones in both the heating and non-heating periods. This suggests that a more diverse landscape with a mix of LULC classes can contribute to lower PM₁₀ levels. These findings are consistent with previous studies, such as [94]. Diverse landscapes often include more green spaces. Green spaces act as sinks for air pollutants such as PM₁₀ [29,37,93]. These findings support the idea that a mixed land use strategy can reduce the risk of exposure to a high concentration of PM by dispersing PM in urban areas [99]. They also recommend securing sufficient open space in urban areas to minimize PM concentration and variability. Some authors confirmed our finding that a higher landscape complexity (LULC heterogeneity) can contribute to the reduction of PM pollution levels [89].

In our study, the contrast index (CI₁) measured the edge length of LULC categories that are sources of PM₁₀ pollution. This index was particularly relevant during the cooling period, where it showed a significant positive correlation with monthly average PM₁₀ concentrations. In other words, a longer edge length for LULC categories that contribute to PM₁₀ pollution is associated with higher PM₁₀ levels in the surrounding area. Several studies have investigated the connection between the edge length of specific LULC types, often represented by the landscape metric edge density (ED), and PM concentration [89,94]. Ref. [94] found a positive correlation between the edge density (ED) of architectural landscapes and PM₁₀ concentrations. This indicates that as the length of the edge and fragmentation of the built-up areas increase, PM concentrations also tend to rise. Our findings contribute to a growing body of research highlighting the importance of landscape configuration, specifically the edge length of PM source areas, in understanding and mitigating PM pollution.

4.4. Effect of Different Soil Textures

The influence of soil texture on PM₁₀ emissions, especially during the heating seasons, was significant due to factors such as reduced vegetation cover, intensified wind erosion, and increased particulate matter from human activities such as residential heating. Fine-textured soils—particularly those rich in clay and silt—play a crucial role in PM₁₀ emissions because these fine particles are more readily retained by the soil and subsequently resuspended in the air, leading to elevated PM₁₀ levels. During the colder months, when the vegetative cover is sparse, silty clay soils tend to trap more particles that can easily be lifted into the atmosphere. Ref. [100] highlighted that fine-textured soils, such as silty clay, significantly contribute to dust resuspension, especially in dry and windy conditions, a finding consistent with our study, which noted higher PM₁₀ emissions in areas with fine soils during heating periods.

The size of soil particles has been shown to affect PM₁₀ emission levels, with finer particles, such as clay, being more susceptible to resuspension. Ref. [101] reported that finer particles, such as clay, resulted in higher concentrations of PM₁₀ as they are more readily transported by wind. In contrast, coarser soils, such as loam and sandy loam, exhibited reduced susceptibility to resuspension, resulting in lower PM₁₀ emissions. Ref. [102] observed that coarser soils resist wind erosion more effectively, leading to a negative correlation between coarse soil textures and PM₁₀ concentrations, findings that align with our observations that coarser soils inherently contributed less to PM₁₀ pollution.

Ref. [103] reinforced the role of soil texture in particulate emissions, noting that PM₁₀ emissions were higher in soils with a higher silt and clay content but lower in sandy soils. This relationship emphasizes that soil composition plays a key role in PM₁₀ emissions and resuspension, especially during seasons with increased particulate release due to heating and wind. Other studies further support that finer soils containing more clay and silt have a higher PM₁₀ emission potential than coarser soils with a predominance of sand [101,104,105]. For example, studies using a dust resuspension chamber in the San Joaquin Valley of California Valley demonstrated a strong correlation between PM₁₀ emissions and clay content, with clay-rich soils showing a higher propensity for PM₁₀ emissions [104]. Wind tunnel experiments confirmed these findings, indicating that PM₁₀ emissions increase with silt content and decrease with sand content [101].

The distinct characteristics of clay and silt particles make them particularly influential in PM₁₀ pollution. Clay particles, due to their small size, are easily airborne, directly contributing to PM₁₀ emissions. Silt, although larger than clay, also falls within the PM₁₀ range and can be easily transported by wind [102,106]. In contrast, sandy soils, though vulnerable to wind erosion, may emit less PM₁₀ if they have already lost fine particles through previous erosion events. Research in the free state region of South Africa supports this, demonstrating that erodible, loose soils with clay and silt significantly increase PM₁₀ emissions [107].

In summary, soil texture—specifically clay and silt content—emerged as a key factor influencing PM₁₀ emissions, with fine-textured soils showing a higher susceptibility to resuspension and particulate release than coarser soils. As our findings contribute to the growing evidence on the subject, it becomes increasingly clear that soil texture should be considered in PM₁₀ pollution mitigation strategies, particularly in regions with seasonal climatic changes that amplify particulate emissions.

According to our findings, the effect of climate change has opposite effects during cooling and heating periods: when the temperature increases, then PM₁₀ from heating is reduced, which could count as a benefit of climate change; however, during cooling periods (summer), the monthly average temperature can increase the monthly average PM₁₀ concentration because the top soil will be dryer, and therefore, it can be a source of PM₁₀ pollution. To mitigate the impact of fine-textured soils on PM₁₀ emissions, especially during heating seasons, shelterbelts can be planted to reduce wind erosion and the soil can be covered during the winter months with crops such as cabbage to limit soil exposure. Leaves of plant residues from previous harvests can also act as a natural barrier against wind erosion. These strategies not only reduce particulate pollution but also improve soil health and agricultural sustainability, offering practical solutions for regions prone to high PM₁₀ levels.

5. Conclusions

In conclusion, this study demonstrated that incorporating landscape metrics such as the SI, MPS, and SHDI significantly improves PM₁₀ prediction models, providing a more accurate understanding of how LULC influences air quality. Meteorological factors, such as monthly average temperature, wind speed, precipitation, and air pressure, had a strong impact on PM₁₀ concentrations. Notably, the temperature’s effect was amplified during the heating season, with its impact nearly quadrupling, which indicates increased pollutant retention during this time. Landscape composition, especially the presence of green spaces, is vital to reducing monthly average PM₁₀ levels. Larger patches in high-emission areas (LULC Groups 1 and 3) tended to increase PM₁₀, particularly during the heating periods due to concentrated emissions. In contrast, larger patches in barrier zones (LULC Group 2) helped mitigate PM₁₀ levels by acting as buffers or facilitating pollutant dispersion. This study highlights that the specific landscape configurations, such as the size, shape, and connectivity of green spaces, significantly impact the diffusion and distribution of PM₁₀. For instance, larger, contiguous patches of forests and urban parks are more effective in mitigating PM₁₀, while fragmented and complex built-up areas tend to exacerbate pollution levels. Soil texture, especially fine-textured soils rich in clay and silt, also contributes to PM₁₀ levels by resuspending particles, particularly in the colder months when vegetation cover is limited. These findings underscore the importance of adapting PM₁₀ management strategies to account for landscape structure, seasonal variations, and climate trends to enhance urban air quality and public health. Our research provides valuable information for landscape planners looking to create sustainable urban environments. Although we cannot fully control climate-related factors, modifying LULC and landscape configurations presents an effective tool for reducing PM₁₀ pollution and improving urban planning results. Climate change is likely to further influence PM₁₀ levels through altered temperature and precipitation patterns. Warmer temperatures could reduce heating-related emissions in winter but could increase dust generation due to the dryness of the soil in summer [108]. Changes in precipitation could also affect wet deposition processes, altering the removal of PM₁₀ from the atmosphere. Given the complex interplay of these factors, a comprehensive approach to mitigation is essential. This includes integrating urban planning strategies, emission control measures, and public health considerations. Promoting green infrastructure and optimizing urban forms to enhance ventilation are key interventions that can help address PM₁₀ pollution and improve urban air quality [96]. Cities can adopt design guidelines to enhance air circulation, such as lower aspect ratios for street canyons, strategic building orientations aligned with wind patterns, and ground-level porosity through features like arcades or elevated structures [109,110,111]. Zoning regulations can restrict building heights and densities in ventilation corridors or near sensitive areas like schools and hospitals [110]. Mixed-use development, reduced vehicle reliance, and green infrastructure, such as green roofs and strategically placed trees, could further improve air quality when carefully planned [109,112,113]. This study was limited by the thematic accuracy (≥80%) of the Urban Atlas LULC dataset, which may have introduced uncertainties in the spatial modeling. While landscape metrics such as the PLANDs, SHDI, SHI, and MPA were effectively used, future research could improve predictions by integrating urban form indexes (e.g., connectivity, compactness, and 3D structural metrics) to better capture spatial configurations. Additionally, using daily PM10 concentrations instead of monthly averages could enhance temporal precision and provide deeper insights into short-term variations in air quality.