Spatial Autocorrelation Analysis of CO and NO₂ Related to Forest Fire Dynamics

Abstract

The increasing frequency and severity of forest fires globally highlight the critical need to understand their environmental impacts. This study applies spatial autocorrelation techniques to analyze the dispersion patterns of carbon monoxide (CO) and nitrogen dioxide (NO₂) emissions during the 2021 Manavgat forest fires in Türkiye, using Sentinel-5P satellite data. Univariate (UV) Global Moran’s I values indicated strong spatial autocorrelation for CO (0.84–0.93) and NO₂ (0.90–0.94), while Bivariate (BV) Global Moran’s I (0.69–0.84) demonstrated significant spatial correlations between the two gases. UV Local Moran’s I analysis identified distinct UV High-High (UV-HH) and UV Low-Low (UV-LL) clusters, with CO concentrations exceeding 0.10000 mol/m² and exhibiting wide dispersion, while NO₂ concentrations, above 0.00020 mol/m², remained localized near intense fire zones due to its shorter atmospheric lifetime. BV Local Moran’s I analysis revealed overlapping BV-HH (high CO, high NO₂) and BV-LL (low CO, low NO₂) clusters, influenced by topography and meteorological factors. These findings enhance the understanding of gas emission dynamics during forest fires and provide critical insights into the influence of environmental and combustion processes on pollutant dispersion.

1. Introduction

The frequency of forest fires has significantly increased in recent years, primarily attributed to the impact of climate change [1,2,3]. This rise in both frequency and intensity leads to detrimental effects such as habitat destruction and loss of biodiversity [4,5]. Despite these negative impacts, forest fires are a natural component of forest ecosystems that play a crucial role in supporting biodiversity and facilitating the regeneration of certain species [6,7,8]. Maintaining this balance is crucial in the Mediterranean region, where wildfires constitute one of the most pressing environmental challenges, significantly impacting countries like Greece, Spain, Italy, Portugal, Türkiye, and France [9]. Additionally, neighboring regions, such as northeast Iraq, which shares ecological and climatic similarities with Türkiye, have also experienced severe forest fires [10,11,12]. The 2021 forest fire season, particularly severe in the Mediterranean countries and nearby regions, resulted in numerous human fatalities and fire-related injuries such as burns and respiratory issues due to smoke inhalation, extensive natural and rural damage, and significant environmental losses. Furthermore, these fires significantly impact ecosystem stability, leading to the loss of biodiversity, disruption of nutrient cycles, and alteration of forest structures, which can lead to extensive consequences for the broader environment [13].

While forest fires destroy life across hundreds of hectares of land, they also significantly affect the Earth’s atmospheric composition by releasing carbon emissions, impacting greenhouse gases (GHGs) such as carbon dioxide (CO₂), methane (CH₄), nitrous oxide (N₂O), and halogenated gases, all of which contribute to heat-trapping [14,15]. CO, a by-product of incomplete combustion from fossil fuels and biomass combustion, although not a greenhouse gas (non-GHG), alters atmospheric chemistry and indirectly affects surface temperatures [15]. Additionally, NO₂, also a non-GHG, plays an essential role in understanding the global and local impacts of biomass burning [16,17,18,19]. Therefore, the increasing frequency of forest fires can potentially increase total carbon emissions and reduce ecosystem carbon storage by disturbing the balance between fire-induced CO₂ emissions and forest carbon sequestration [20].

Remote sensing technology provides powerful tools for monitoring non-GHGs such as CO and NO₂ in the Earth’s atmosphere [21,22]. This technology enables satellites and global fire emissions inventories to detect and measure concentrations of these gases across extensive geographical areas, utilizing distinct spectral signatures found in the ultraviolet, visible, infrared, and shortwave infrared spectra of various gases [23]. For example, the Tropospheric Monitoring Instrument (TROPOMI) aboard the Copernicus Sentinel-5P satellite provides essential daily global data on non-GHGs and aerosols for air quality monitoring, by utilizing a high-resolution spectrometer system with seven spectral bands from ultraviolet to short-wave infrared wavelengths [24]. On the other hand, commonly used global fire emission inventories such as the Global Fire Assimilation System (GFAS), Global Fire Emissions Database (GFED), and the Fire Inventory from the United States National Center for Atmospheric Research (NCAR-FINN) use MODIS/VIIRS burned area products which serve to identify and disseminate information regarding the boundaries of areas affected by fire [9,25]. While GFED and FINN estimate fire emissions based on the burned area, GFAS estimates fire emissions based on fire energy from MODIS-derived fire radiative power [26]. Especially in remote areas where ground-based measurements and other ancillary information are lacking, satellite measurements of regional non-GHG measurements provide valuable information on combustion efficiency, combustion practices, fuel types, and their variability [27].

Regarding forest fire emissions studies, several studies have estimated GHGs and non-GHGs and particulate emissions from forest fires using global fire emission inventories or emission factors [9,28,29]. For example, Bar et al. (2022) calculated emission amounts by considering factors such as forest type, burned area, dry biomass load, combustion efficiency, and emission factors, which represent the emission of a component per unit of dry matter combustion [28]. At the local scale, Bilgiç et al. (2023) applied an emission estimation method for significant forest fires in the Eastern Mediterranean between July and August 2021, estimating emissions of NO_x, CO, SO₂, TSP, PM_2.5, and BC, and found their estimates were similar to GFAS but higher than GFED and FINN inventories [9]. Similarly, Yılmaz et al. (2023) analyzed burn severity in the Mediterranean region of Türkiye using Sentinel-2 data and assessed post-fire CO levels with Sentinel-5P [30]. Eke et al. (2024) further highlighted substantial increases in CO, NO₂, and HCHO concentrations during intense forest fires in Antalya and Muğla in the summer of 2021, emphasizing the utility of satellite retrievals in capturing forest fire impacts [31]. Likewise, a study on the 2021 Marmaris and Manavgat fires in Türkiye reported significant increases in NO₂ (up to 260.43%) and CO (up to 107.07%) levels, with increased aerosol levels correlating with fire severity observed via Sentinel-2 dNBR maps [32].

To further understand the dynamics of these emissions, it is crucial to monitor and analyze their spatial patterns during fire events. This is where spatial autocorrelation becomes essential, as it allows for the examination of how parameter values—such as pollutant concentrations—are correlated with their geographic neighbors, revealing underlying patterns that may not be immediately apparent. Spatial autocorrelation summarizes the clustering of spatial data based on proximity and helps identify spatial heterogeneity and uniformity between parameters [33,34,35,36]. In spatial statistics, various indexes are commonly used, including Getis’ G-index (refer also Getis-Ord G; Global scale) [37], Geary’s C-index [38], Tango’s C-index [39], and Moran’s I [33]. Among these, Moran’s I—used either univariately or bivariately depending on the number of variables analyzed—is a widely used method in spatial clustering analysis [40]. This measure enables to determine whether parameters across a geographic area exhibit patterns of spatial clustering, randomness, or dispersion [41]. The univariate form examines the distribution of a single variable across areas, while the bivariate form analyzes spatial distributions between two variables [36,42]. Empirical studies have demonstrated the efficacy of Bivariate Moran’s I in analyzing the spatial interdependencies between two datasets, facilitating a deeper understanding of how changes in one variable may be related to changes in another within a geographic space [43,44,45,46]. This index has been used to investigate the effects of land use on monthly air pollution levels for pollutants such as PM_2.5, PM₁₀, O₃, NO₂, SO₂, and CO in northern Italy using the CAMS (Copernicus Atmosphere Monitoring Service) European air quality re-analysis dataset, with ~10 km spatial resolution [47]. Habibi et al. (2017) used Moran’s I and Getis-Ord G statistics to characterize the spatial patterns of CO and PM_2.5 pollution in Tehran, identifying high-clustering areas and suggesting methods for their remediation [48].

Regarding forest applications, a Moran’s I of 0.80 (p = 0.001, indicating strong statistical significance) reflects strong spatial autocorrelation between fire observations, with 60% of cells aggregated at a 0.5° resolution (≈55 km at the equator) showing significant positive or negative spatial correlation [49]. In a recent study, the forest fire spread hazard index, which calculates the likelihood and intensity of forest fire spread based on forest fuel, weather, topography, and fire behavior, showed positive spatial autocorrelation with Moran’s I, indicating a clustered distribution pattern [50]. In addition to its use in fire spread hazard studies, Moran’s I is frequently used in fire risk assessment studies, facilitating the identification of high-risk areas [51,52]. Moreover, Kganyago and Shikwambana (2020) used Moran’s I to assess the characteristics of 2018–2019 fire events in the USA, Brazil, and Australia using MODIS data, including active fires, Fire Radiative Power (FRP), and burned area [41]. Studies have also explored the carbon emissions from industrial land and carbon absorption of forest land with Moran’s I and Getis-Ord Gi* (Local scale) [53].

While Sentinel-5P data have been used in several studies to monitor non-GHG emissions in urban and industrial environments, its application to forest fire-induced emissions remains rather limited. Unlike the more stable emission patterns in urban areas, forest fires generate transient and highly variable emission bursts that change rapidly in both space and time. This study specifically addresses the critical gap in the literature by investigating the spatial distribution dynamics of fire-induced non-GHG emissions, focusing on the devastating Antalya Manavgat fires of July 2021. To capture these rapidly changing emission patterns, advanced spatial analysis techniques, such as Moran’s I and LISA, are applied to fire-affected landscapes. Furthermore, fire-specific variables, including wind speed, digital elevation models (DEMs), and FRP, are integrated to contextualize and enhance the spatial analysis. By utilizing Sentinel-5P’s high temporal resolution and improved spatial resolution (≈5.5 km × 3.5 km), in contrast to commonly used CAMS data (≈10 km), this study offers a novel perspective on the dynamic and transient nature of fire-induced emissions. This approach provides new insights into the spatial distribution, environmental impacts, and specific dynamics of forest fire emissions, addressing a critical gap in our understanding of fire-related non-GHG emissions, which contrasts with the more common focus on urban environments in existing studies. The investigation is guided by three key questions:

• What are the spatial distributions of non-GHGs, specifically CO and NO₂, during forest fire events in the Manavgat?
• How do variations in factors such as topography, wind speed, and fire ignition points affect the spatial distribution of CO and NO₂ emissions detected by Sentinel-5P satellite data?
• What do the Univariate and Bivariate Moran’s I indices reveal about the spatial distribution of CO and NO₂ during forest fire events?

2. Study Area and Data Used

Antalya is located in the Mediterranean region of southern Türkiye (Figure 1a) and exhibits a Mediterranean climate characterized by hot, dry summers and mild, wet winters. The study area, the Manavgat fire zone in Antalya (Figure 1b), was the most significantly affected region by the 2021 forest fires in Türkiye [54], with the majority of the burned area located there, while smaller portions extended into the neighboring Ibradı and Akseki districts to the northeast. The study area includes three reservoirs—Naras, Manavgat, and Oymapınar—located within Black Pine and Calabrian Pine forests and primarily fed by two major water sources: the Manavgat River and Naras Stream. The burned area surrounding these reservoirs encompasses approximately 54,769 ha, underlining their critical role in firefighting operations by supplying water for both aerial and ground vehicles [1,55] (Figure A1 and Figure 2). The elevation of the study area ranges from 2 to 1657 m with an average of 396 m, and the slope varies from 0° to 80°, averaging 15°. The Manavgat district primarily comprises Southern Anatolian montane conifer and deciduous forests, as well as Eastern Mediterranean conifer–broadleaf forests, and belongs to the Mediterranean forests, woodlands, and scrub biome. These forests, predominantly located in lowland areas with fragmented or non-intact cover [56], consist of highly combustible species such as Black Pine (Pinus nigra Arnold) and Calabrian Pine (Pinus brutia Ten.), which are also found in the study area [57,58]. Figure 1c,d depicts changes in vegetation biomass in Sentinel-2 false-color composite images captured before and during the fire event.

Figure A1 presents detailed land use information for the fire zone, based on the 2018 CORINE Land Cover (LC) classification system, which features a spatial resolution of 100 m and an 85% thematic accuracy [59,60]. Approximately 90% of the burned area falls within the following CORINE LC categories: 312 (46%, Coniferous Forest), 324 (19%, Transitional Woodland-Shrub), 242 (13%, Complex Cultivation Patterns), and 243 (9%, Land Principally Occupied by Agriculture with Significant Areas of Natural Vegetation).

The fire in this region originated from seven distinct ignition points. Figure 2 illustrates the daily burned areas and fire ignition points in the Manavgat fire zone, as obtained from the General Directorate of Forestry.

Table 1. Characteristics of fire ignition points associated meteorological conditions recorded at the time of ignition, and dominant tree species.

	Fire Start		First Response		Controlled		Extinguished
#	Date	Local Time	Date	Local Time	Date	Local Time	Date	Local Time	Humidity (%)	Max Temperature (°C)	Wind speed (km/h)	Wind Direction	Tree Species
1	28 July 2021	11:58	28 July 2021	12:08	30 July 2021	15:00	7 September 2021	09:30	13	42	58	NE	Redpine
2	28 July 2021	16:00	28 July 2021	16:10	7 August 2021	09:30	1 September 2021	16:00	10	38	30	N	Redpine
3	28 July 2021	23:45	29 July 2021	00:10	3 August 2021	12:00	1 September 2021	14:30	25	32	74	E	Redpine
4	29 July 2021	01:30	29 July 2021	02:00	6 August 2021	08:30	3 September 2021	10:30	11	42	32	NW	Redpine
5	31 July 2021	19:00	31 July 2021	19:03	3 August 2021	12:00	1 September 2021	14:30	15	35	20	NW	Redpine
6	2 August 2021	13:00	2 August 2021	13:15	7 August 2021	09:30	5 September 2021	13:00	10	41	36	N	Redpine
7	3 August 2021	07:30	3 August 2021	08:30	7 August 2021	09:30	13 September 2021	07:30	15	32	30	N	Blackpine

presents meteorological data and operational details related to these ignition points and the control efforts, based on fire record slips provided by the General Directorate of Forestry of Türkiye, with timestamps matching the fire ignition time as recorded during the event. The table includes information on the fire’s start date, first response, and the dates when the fire was brought under control and fully extinguished. The fire, which ignited on 28 July 2021, was brought under control on 7 August 2021 and fully extinguished by 13 September 2021. On 28 July 2021, four ignition points (numbered 1–4 in Table 1 and Figure 2) were identified, with ignition occurring during both day and night hours. The average response time was 7 min, though it increased for nighttime ignitions. Adverse meteorological conditions contributed to the prolonged control and extinguishment efforts, as humidity dropped to 10% as the minimum, maximum temperatures reached 42 °C, and wind speeds peaked at 74 km/h, with the wind direction generally toward the north.

The dataset used in this study is outlined in the following:

• Sentinel-5P: As a satellite in the Copernicus program, it is designed to monitor atmospheric composition, including gases like nitrogen dioxide, ozone, and methane, providing crucial data on air quality and climate change. Over the fire zone, the satellite passes approximately between 10:08 and 11:28 UTC on the specified dates, subject to orbital variations.
• CO: The vertical column of CO has a spatial resolution of 1113.2 m downloaded from the Google Earth Engine (GEE) data catalog [61]. The Sentinel-5 mission consists of a high-resolution seven different spectral bands: UV-1 (270–300 nm), UV-2 (300–370 nm), VIS (370–500 nm), NIR-1 (685–710 nm), NIR-2 (755–773 nm), SWIR-1 (1590–1675 nm), and SWIR-3 (2305–2385 nm) [62]. TROPOMI on board the Sentinel 5 Precursor (Sentinel-5P) satellite observes the global abundance of CO using measurements of the Earth’s radiance in the 2.3 μm spectral region of the SWIR part of the solar spectrum [63].
• NO₂: Similarly to Sentinel-5P CO, the vertical column of NO₂ has the same spatial resolution (1113.2 m) and was downloaded from the GEE data catalog [61]. The TROPOMI NO₂ processing system utilizes adapted algorithms from two key sources: the DOMINO-2, which is part of the Dutch OMI NO₂ data products for the Ozone Monitoring Instrument (OMI), and the European Union’s QA4ECV (The European Quality Assurance for Essential Climate Variables) NO₂ reprocessed dataset for OMI. These algorithms have been adapted and customized to meet the specific requirements of the TROPOMI [64].

To ensure the quality of the measurements for these two gases and reduce the risk of misinterpretation, only pixels with a Quality Assurance (QA) value above 50% were included in the analysis. This threshold (QA > 50%) filters out data affected by cloud cover, snow/ice, atmospheric interferences, processing errors, sun glint, and specific regional anomalies, as recommended by Sentinel-5P technical guidelines, thereby enhancing the reliability and accuracy of the observed gas concentration patterns [63,64,65].

• Digital Elevation Model (DEM): The digital elevation data used in this study was obtained from NASA’s SRTM V3 (SRTM Plus) via GEE, with a resolution of 1 arc-second (≈30 m). The dataset underwent a void-filling process using open-source data, including ASTER GDEM2, GMTED2010, and NED, ensuring complete coverage [66,67].
• VIIRS FRP values, obtained from the NASA website, were utilized in this study due to their superior spatial resolution (375 m for the Suomi NPP) compared to MODIS, enabling more precise detection of small fire hotspots [25].

3. Methodology

The methodology employed in this study consists of three primary processing stages, as illustrated in Figure 3. The first stage focuses on preparing the data for subsequent analysis, including reprojection to Universal Transverse Mercator (UTM zone 36) and resampling the Sentinel-5P data from its original resolution of approximately 5.5 km × 3.5 km to a finer resolution of 1.1 km by GEE. Moreover, as recommended by Sentinel-5P technical guidelines, only data with a QA threshold above 50% were included from GEE to enhance the reliability of the gas concentration models by filtering out data affected by cloud cover, snow, atmospheric disturbances, and processing errors. The second stage involves conducting spatial autocorrelation analyses, utilizing both Univariate and Bivariate Moran’s I to explore the spatial relationships of the variables. The results are further evaluated in the third stage, through scatterplots and quadrants, providing a comparative analysis of the two gases. Each of these stages is described in more detail in the following subsections.

3.1. Preprocessing

At this stage, several key steps are undertaken to prepare the data for spatial analysis. First, datasets are collected and organized based on the study’s spatial and temporal framework. Data cleaning is performed to remove errors and inconsistencies, ensuring reliability for subsequent analysis. Temporal alignment is then performed to align date/time ranges across datasets. Reprojection is applied to align all spatial datasets to a common coordinate reference system, enabling precise spatial comparisons. The Nearest Neighbor method is used for resampling to match the spatial resolution across datasets, ensuring consistency throughout the study area.

3.2. Spatial Autocorrelation Analysis

The analysis evaluates spatial clustering patterns using Univariate and Bivariate Moran’s I indexes, supported by scatterplots and quadrant maps to further explore spatial relationships. Since the study focuses on identifying clustering patterns of pollutant concentrations, dispersion-modeling techniques were not employed, as these are more appropriate for simulating pollutant transport rather than analyzing clustering patterns. The detailed methodology for the clustering pattern analyses is presented in the following subsections.

3.2.1. Moran’s I

Moran’s I, first introduced by Moran in 1948 [68], evaluates the pattern of a dataset and determines whether it is dispersed, clustered or random [69]. The values of Moran’s I range from −1, indicating perfect dispersion, to +1, signifying perfect spatial clustering. A zero value suggests a random spatial pattern [36]. Moran’s I can be divided into Univariate and Bivariate Moran’s I, which are used to examine the distribution of a single variable or two variables, respectively [70], and explained in more detailed in the following subsections. These indexes have two subindexes: Global Moran’s I and Anselin’s Local Moran’s I [71]. Anselin’s Local Moran’s I (also known as LISA, Local Indicators of Spatial Autocorrelation) was first introduced by Anselin in 1995 [33]. While Global Moran’s I defines a single value indicating overall spatial autocorrelation (i.e., amount of clustering across the area), Anselin’s Local Moran’s I, defines local clusters and outliers with similar or dissimilar values [72]. Global Moran’s I is a widely used metric for quantifying overall spatial autocorrelation across an entire study area; however, it does not identify the specific locations of clusters. To overcome this limitation, local measures of spatial autocorrelation are applied, allowing for the detection of spatial patterns by assessing the similarities or differences in each location relative to its neighbors [48].

Moran’s I is calculated by constructing a spatial weight matrix $\left(w_{ij}\right)$ that measures the strength of the relationship between spatial units [33,34,36]. The spatial weight matrix has been developed by different researchers using different techniques such as queen contiguity, rook contiguity, and nearest neighbors [34,36]. The queen contiguity approach was used in this study to calculate the spatial weights, as it has proven to be effective in earlier studies [36,43,73]. The queen contiguity defines weights based on adjacency between spatial units. If spatial unit ‘i’ shares any edge with unit ‘j’ including north, east, west, south, and corners, they are recognized as neighbors and assigned a weight of 1, otherwise the weight is assigned as 0. The diagonal of the weight matrix consist of zeros [36,69,74]. To ensure each spatial unit’s influence is equalized relative to its number of neighbors, a row-standardization of the spatial weight matrix is performed. This involves adjusting the weights in each row to sum to one [75].

Univariate (UV) Moran’s I

UV Moran’s I is a statistical measure used to explore clustering patterns by analyzing the distribution of a single variable across geographic regions and assessing the degree of spatial autocorrelation within the dataset [36,42]. It examines whether similar or dissimilar values are spatially clustered and evaluates the statistical significance of these clusters. Global Moran’s I, a specific form of UV Moran’s I, provides an overall measure of spatial autocorrelation across the entire dataset, indicating whether values are generally clustered or dispersed throughout the region [42]. The equation for calculating UV Global Moran’s I is as follows [76]:

$$\mathrm{UV}\mathrm{Global}\mathrm{Moran}’\mathrm{s}\mathrm{I}={\displaystyle \frac{n\sum _{i=1}^n\sum _{j=1}^n{w}_{ij}\left(x_i-\overline{x}\right)\left(x_j-\overline{x}\right)}{(\sum _{i=1}^n\sum _{j=1}^n{w}_{ij}){\sum _{i=1}^n\left(x_i-\overline{x}\right)}^2}}$$

(1)

On the other hand, UV Anselin Local Moran’s I examines spatial distributions within a region to differentiate between random and non-random patterns. By applying LISA to each spatial unit, it assesses the extent of significant clustering of similar values around each unit [33]. This method effectively identifies spatial clusters and outliers by comparing each unit with its neighbors [77]. The equation for UV Anselin Local Moran’s I is as follows [76]:

$$\mathrm{UV}\mathrm{Anselin}\mathrm{Local}\mathrm{Moran}’\mathrm{s}\mathrm{I}={\displaystyle \frac{n\left(x_i-\overline{x}\right)\sum _{j=1}^n{w}_{ij}(x_j-\overline{x})}{\sum _{j=1}^n{(x_j-\overline{x})}^2}}$$

(2)

In both equations, x represents the variable, and $\overline{x}$  denotes its mean. The term $x_i$ signifies the value of the variable at a specific spatial unit (e.g., pixel), while $x_j$ represents the value of at another unit. Here, n refers to the total number of spatial units, and $w_{ij}$ is the row-standardized weight matrix.

Bivariate (BV) Moran’s I

BV Moran’s I is used to examine spatial clustering between two variables by analyzing their spatial distributions [36,42]. The BV Global Moran’s I provides an overall measure of the spatial relationships between the variables across the entire region, while the BV Local Moran’s I identifies local spatial correlations at individual spatial units (e.g., pixels) [70]. The equations for BV Global and Local Moran’s I are as follows [78]:

$$\mathrm{BV}\mathrm{Global}\mathrm{Moran}’\mathrm{s}\mathrm{I}={\displaystyle \frac{n\sum _{i=1}^n\sum _{j=1}^n{w}_{ij}\left(x_i-\overline{x}\right)\left(y_j-\overline{y}\right)}{\left(\sum _{i=1}^n\sum _{j=1}^n{w}_{ij}\right)\sum _{i=1}^n{\left(x_i-\overline{x}\right)}^2}}$$

(3)

$$\mathrm{BV}\mathrm{Anselin}\mathrm{Local}\mathrm{Moran}’\mathrm{s}\mathrm{I}={\displaystyle \frac{n\left(x_i-\overline{x}\right)\sum _{j=1}^n{w}_{ij}\left(y_j-\overline{y}\right)}{\sum _{i=1}^n{\left(x_i-\overline{x}\right)}^2}}$$

(4)

In these equations, x and y represent the variables of interest, while $\overline{x}$ and $\overline{y}$  denote their respective means. The term n refers to the total number of spatial units (e.g., pixels), and $w_{ij}$ represents the row-standardized weight matrix.

3.2.2. Moran’s I Scatter Plots

The Moran’s I statistic is visualized using a scatter plot, known as a Moran’s I scatter plot, where the spatially lagged variable (Wx) is plotted against the original variable (x). As shown in Figure A2, this plot illustrates spatial autocorrelation, serving as a valuable tool for identifying spatial patterns and understanding the spatial structure of the data. The spatial lag represents the weighted average of values from neighboring units, where weights are determined by their spatial proximity [75]. The equation for calculating the spatial lag $({lag}_i$) is as follows:

$${lag}_i={\displaystyle \frac{\sum _{j=1}^n{W}_{ij}\times x_j}{\sum _{j=1}^n{W}_{ij}}}$$

(5)

where x denotes the variable of interest and $w_{ij}$ represents the row-standardized weight matrix.

In Figure A2, the polygons at the top represents varying degrees of clustering, corresponding to different values of Moran’s I, which range from −1 (perfect dispersion) to +1 (perfect spatial clustering), with 0 indicating a random distribution. Below, three scatter plots display the distribution of data points across four quadrants, each representing a distinct spatial relationship. In these scatterplots, the ‘High-High’ (HH) cluster in Q1 identifies regions with high values surrounded by high values, while the ‘Low-Low’ (LL) cluster in Q3 represents regions with low values surrounded by low values. Both quadrants reflect positive spatial autocorrelation, where similar values are clustered together. In contrast, the ‘High-Low’ (HL) cluster in Q4 highlights high values in proximity to low values, and the ‘Low-High’ (LH) cluster in Q2 indicates low values surrounded by high values, reflecting negative spatial autocorrelation and identifying spatial outliers or contrasting values [75]. When similar values cluster, as shown in the top right, the scatterplot forms a clear trendline, signifying stronger spatial autocorrelation.

4. Results

4.1. Temporal Analysis of CO and NO₂ Values

To understand the dynamics of fire activity and its environmental impacts, it is crucial to examine the temporal variations in emitted gases and fire intensity. Figure 4 presents the temporal changes in CO and NO₂ concentrations and FRP values in the Manavgat fire zone from 28 July to 4 August 2021. The graph shows a distinct pattern in CO and NO₂ emissions, which aligns with the progression of fire intensity as indicated by FRP. The FRP peaked at 29,218 MW on 28 July, marking the onset of the intense forest fire in the region. However, CO and NO₂ concentrations on 28 July were relatively low, likely due to a time lag between fire ignition and the accumulation of detectable emissions in the atmosphere. This time lag corresponds to the interval between fire initiation (11:58) and satellite data collection (14:03–14:39). The concentrations of CO and NO₂ reached their maximum values on 29 July, with CO peaking at 0.13500 mol/m² and NO₂ at 0.00023 mol/m². This peak reflects the time required for these gases to accumulate in the atmosphere following the initial fire activity. After 29 July, both CO and NO₂ concentrations exhibited a general decline, mirroring the reduction in FRP values over the observed period. Interestingly, the sustained variations in CO and NO₂ concentrations, despite the decline in FRP, suggest ongoing emissions from residual fires and potential contributions from atmospheric dispersion processes. These findings emphasize the complex interactions between fire dynamics, atmospheric conditions, and gas emissions during fire events.

4.2. UV Moran’s I Analysis

UV Global Moran’s I values for CO and NO₂ gases between 28 July and 4 August 2021 are shown in Figure 5. The consistently high Moran’s I values for both gases (ranging from 0.84 to 0.93 for CO and 0.90 to 0.94 for NO₂) indicate strong spatial autocorrelation, reflecting persistent clustering patterns throughout the observed period. While minor fluctuations are observed, the values for CO remain between 0.84 and 0.93, while those for NO₂ consistently exceed 0.90. Notably, on 1 August, Moran’s I value for CO shows a slight decrease, indicating a temporary reduction in spatial autocorrelation. This decline may be attributed to factors such as the absence of new fire ignitions on that day, firefighting efforts, or potential data capture limitations, which could have contributed to a less clustered spatial distribution.

Following the UV Global Moran’s I analysis, which provided a general indication of spatial autocorrelation across the study area, UV Local Moran’s I was applied to identify specific regions with significant clusters of high or low values. This approach offered a more detailed, location-specific perspective on the spatial distribution of CO and NO₂ concentrations. To further explain the dispersion of these gases, an in-depth analysis was conducted, incorporating factors such as elevation, wind direction and speed, and fire ignition points to better understand the main drivers behind the observed spatial clustering patterns (Figure 6). As shown in Figure 6c, pixels with QA values below 50% were filtered out, except in the data from 30 July, 31 July, and 4 August. This filtering led to missing data in the CO dataset, which was subsequently excluded from the analysis.

Topography (Figure 6a) plays a pivotal role in shaping fire behavior, which, in turn, affects the spatial distribution and levels of CO and NO₂ detected by the Sentinel-5P satellite (Figure 6c and Figure 9a). Lower elevations with flatter terrain and greater fuel continuity exhibited more aggressive and widespread fire activity, as indicated by the larger extent of burned areas in these regions. Conversely, at higher elevations, fire spread was slower due to natural barriers such as steeper slopes, cooler temperatures, and reduced fuel availability (e.g., thinner vegetation or moisture retention), resulting in lower fire intensity. This variation in fire behavior is evident in the CO concentrations mapped in Figure 6c, where lower altitudes consistently showed higher levels of CO. Similarly, NO₂ concentrations (Figure 9a) were significantly higher at lower elevations, suggesting that favorable fire conditions in these regions supported more intense combustion and greater emissions. The UV Local Moran’s I analysis (Figure 6d and Figure 9b) further corroborates these findings, highlighting UV-HH clusters of CO and NO₂ concentrations in lower-altitude regions. Notably, on 3 and 4 August, the UV-HH clustering pattern shifted, as fire ignitions began at higher elevations on 3 August, temporarily altering the spatial distribution of emissions. These observations emphasize the influence of topography on fire behavior and subsequent gas dispersal patterns.

Figure 7, derived from DEM data, provides a detailed analysis of elevation-dependent spatial clustering for CO and NO₂ concentrations, highlighting specific elevation ranges where significant clustering occurs. Both gases exhibit an inverse correlation with elevation, with higher concentrations predominantly occurring at lower elevations (0–500 m). Notably, substantial UV-HH CO clusters were observed on 31 and 30 July (Figure 7a), while UV-HH NO₂ clusters peaked on 29 and 31 July (Figure 7b). At higher elevations (500–1000 m and above), UV-HH clusters were negligible for both gases, underlining the role of elevation in limiting fire intensity and gas emissions. Conversely, UV-LL clusters for both gases were more prevalent at lower altitudes, particularly on 2 and 3 August, although some UV-LL clusters persisted at higher altitudes. Despite minor differences in clustering patterns, CO and NO₂ displayed broadly similar spatial clustering trends during fire events, reflecting the interaction of fire dynamics, elevation, and atmospheric dispersion processes.

After analyzing the effects of elevation, the relationship between wind speed and CO concentrations was examined. Variations in wind speed (Figure 6b) significantly influenced the spatial distribution of CO and NO₂ concentrations, as well as the formation of clusters, particularly during shifts in wind patterns. On 28 July, a notable wind speed of 74 km/h likely accelerated fire spread, enabling rapid combustion and contributing to widespread emissions. While lower wind speeds are generally associated with the accumulation of gas concentrations, the UV Local Moran’s I analysis (Figure 6d and Figure 9b) indicates that significant CO and NO₂ clusters also formed at high or moderate wind speeds. This suggests that high wind speeds might not entirely prevent the localized accumulation of gases due to the interaction of other factors, such as terrain and fire dynamics. As wind speeds decreased in the subsequent days, both CO and NO₂ concentrations tended to accumulate (Figure 6d and Figure 9b), reflecting reduced dispersal capacity under calmer conditions. This is evident in the UV Local Moran’s I analysis, which shows a growing prevalence of high-concentration (UV-HH) clusters, particularly in the southern and western regions.

Moreover, the locations of fire ignition points significantly influenced the spatial distribution of CO and NO₂ concentrations (Figure 6c and Figure 9a), with the highest concentrations observed near these points and gradually decreasing with distance. The presence of coniferous tree species that shed highly flammable cones, further accelerated fire spread and intensity, resulting in seven ignition points over the fire duration. As the fire progressed and new ignition points emerged, UV-HH clusters of CO and NO₂ concentrations expanded spatially (Figure 6d and Figure 9b), particularly near active fire zones. Conversely, UV-LL clusters were more prevalent in regions farther from the ignition points or in areas where firefighting efforts had effectively reduced fire severity.

As a further step, the areas affected by CO concentration levels within the identified UV-HH and UV-LL clusters were assessed (Figure 8). The analysis of daily trends revealed that the largest affected areas in the UV-HH clusters occurred on 31 July, followed by 30 July and 1 August. Although the CO concentration ranges associated with the affected areas varied slightly across days, the dominant pattern showed that most UV-HH-affected regions were linked to high-concentration CO clusters (0.10100–0.36000 mol/m²). This pattern highlights the increased intensity of fire activity in areas with higher CO concentrations, particularly on 31 July, when the largest UV-HH clusters were observed.

In contrast, the UV-LL clusters showed the largest affected areas on 3 August, with smaller but notable peaks on 2 August and 4 August. Additional UV-LL peaks were observed between 28 July and 1 August; however, these were less prominent due to the dominance of UV-HH clusters during this period. UV-LL clusters were predominantly associated with the low-concentration range (0–0.05000 mol/m²), indicating that fires with lower CO concentrations had a significant impact over larger areas. Unlike UV-HH clusters, higher CO concentration ranges had minimal impact on the UV-LL clusters, with most affected areas concentrated at the lowest concentration levels.

Another gas parameter, NO₂ (Figure 9a), exhibited spatial clustering patterns similar to CO (Figure 6a–d), particularly near the core fire activity zones, indicating the significant influence of combustion processes on emissions of both gases. Dense UV-HH clusters (Figure 9b) were primarily observed in low-lying areas with low wind speeds, while UV-LL clusters were more prevalent in regions farther from the fire zones or where firefighting efforts reduced fire intensity. Despite these similarities, NO₂ concentrations were generally lower, with values exceeding 0.00020 mol/m², compared to CO concentrations, which surpassed 0.10000 mol/m². Additionally, while CO spread over a larger area, NO₂ remained more localized around the most intense fire zones. These patterns suggest that environmental factors, including terrain, wind speed and direction, and proximity to ignition points, influenced the day-to-day fluctuations and spatial distribution of both gases.

Figure 10 illustrates the areas impacted by NO₂ concentration levels within the identified UV-HH and UV-LL clusters. Day-to-day variations in the extent of these NO₂ concentration clusters are evident, with significant changes in affected areas observed across dates. During the initial days of the fire, emissions were detected in the high-concentration range (0.00041–0.00060 mol/m²), though these contributed minimally to the total area. Most emissions within UV-HH clusters were concentrated in the low-concentration range (0.00016–0.00030 mol/m²), which accounted for the largest affected areas, particularly on 31 July, the peak day for UV-HH clusters. Notable impacts were also observed on 29 July and 3 August. Overall, low-range concentrations dominated the extent of affected areas compared to mid- and high-range concentrations, suggesting that the majority of NO₂ emissions during the fire were concentrated at lower levels.

NO₂ concentrations in UV-LL clusters (Figure 10) were substantially lower than those in UV-HH clusters, ranging from 0 and 0.00015 mol/m². The largest areas of UV-LL clusters were observed on 3 August, followed by 2 August and 29 July, predominantly within this low-concentration range. This pattern parallels that of CO, where UV-LL clusters covered broader regions of lower concentrations, while the intensity of fire-related emissions remained notably diminished compared to regions dominated by UV-HH clusters.

4.3. BV Moran’s I Analysis for CO and NO₂

The BV Moran’s I analysis was conducted to explore the spatial correlation between CO and NO₂ concentrations, identifying regions where high or low values of one variable correspond to similar patterns in the other. This analysis provides insights into the spatial interactions between these two pollutants during fire events, highlighting areas where increased concentrations of both gases coincide or diverge. The BV Global Moran’s I values ranged from 0.69 to 0.84 during the observed period, indicating a strong positive spatial relationship between the two gases (Figure 11).

Peaks were observed on 30 July, 29 July, and 3 August, with values near 0.85, while 1 August and 4 August showed lower Moran’s I values of approximately 0.69 and 0.70, indicating a slight decrease in the spatial association between CO and NO₂, likely due to changes in wind patterns, fire dynamics, or atmospheric conditions. The bivariate relationship illustrated in Figure 11 closely parallels the trends in daily mean CO and NO₂ concentrations shown in Figure 4, with similar fluctuations in their spatial association over time.

Following the BV Global Moran’s I values, the BV Local Moran’s I analysis (Figure 12a) was performed to identify spatial clustering patterns where CO and NO₂ exhibit similar or contrasting values. Figure 12a demonstrates the spatial relationships between the two pollutants over the duration of the fire event, revealing four distinct cluster types: BV-HH (High CO, High NO₂), BV-HL (High CO, Low NO₂), BV-LH (Low CO, High NO₂), and BV-LL (Low CO, Low NO₂), along with areas marked as “ns” where no statistically significant relationship was observed. The spatial clusters vary throughout the fire event. BV-HH clusters predominantly occur near active fire zones, with their areal extents peaking on 29 July, 31 July, and 1 August (Figure 12b). Conversely, BV-LL clusters are more common in the northern and southeastern regions, especially from 1 to 3 August (Figure 12b). BV-HL and BV-LH clusters are infrequent and irregular, although a notable BV-LH cluster is present on 31 July.

The BV Local scatterplot, a tool commonly used to assess spatial relationships such as clustering and dispersion patterns, was included in the analysis. Figure A3 illustrates an example from 31 July, when all cluster types were present. In the scatterplot (Figure A3b), Q1 corresponds to BV-HH concentrations of both gases, while Q3 represents BV-LL concentrations. Q2 (BV-LH) and Q4 (BV-HL) indicate scenarios where one gas is emitted at increased levels while the other is reduced. Pixels that are statistically significant, based on the spatially weighted average (lag) of NO₂ and neighboring units of CO, are assigned to one of the four quadrants. Notably, BV-LH clusters (low CO and high NO₂) are localized in the central fire zone (Figure A3a). This finding is supported by the scatterplot (Figure A3b), where BV-LH clusters align closely with BV-HH regions, showing similar spatial clustering patterns to the BV LISA cluster map. BV-HL clusters (high CO, low NO₂) are sparse and statistically insignificant; similarly, gray pixels represent areas with no statistically significant clustering.

The UV and BV Global Moran’s I analyses provided insights into the overall spatial autocorrelation of CO and NO₂, demonstrating their efficacy in identifying general patterns of spatial clustering. In contrast, the UV and BV Local Moran’s I analyses enabled the identification of specific locations where clusters of low or high values were formed. The UV Local Moran’s I analysis revealed distinct clustering patterns for CO and NO₂, highlighting their independent spatial distributions. This distinction reflects differences in their distribution dynamics, driven by factors such as atmospheric lifetimes, combustion dynamics, missing CO data, varying combustion phases, and chemical interactions with other gases. Conversely, the BV Local Moran’s I analysis highlighted the spatial similarities between CO and NO₂, revealing strong correlations in their dispersion patterns, which are influenced by topographic and meteorological factors. Together, these analyses highlight the utility of UV Moran’s I in capturing the distinct spatial distributions of each gas and BV Moran’s I in revealing their spatial correlations, providing complementary insights into gas distribution dynamics during forest fire events.

5. Discussion

Results highlight the complex interaction between CO and NO₂ concentrations during fire events, driven by factors such as elevation, wind speed, and combustion processes. These findings provide valuable insights into how natural and atmospheric conditions influence fire spread and the spatial distribution of fire-induced emissions through clustering patterns. The identified clusters highlight priority zones for targeted management actions, supporting spatially focused decision-making. By identifying statistically significant clusters of high pollutant concentrations, these methods offer actionable guidance for resource allocation, such as determining safe emergency response routes and prioritizing evacuations in severely affected areas. These applications greatly enhance the efficiency and effectiveness of management efforts. To further interpret these findings, the following sections evaluate the broader implications and limitations of the study, focusing on satellite data quality and resolution, topographical and meteorological uncertainties, and the effects of fuel types and combustion phases.

5.1. Image Resolution and Data Quality Issues

The resolutions of satellite imagery are a critical factor that directly affects the level of detail in spatial analyses. TROPOMI, onboard Sentinel-5P, offers a superior spatial resolution (≈5.5 km × 3.5 km) compared to its predecessors, such as OMI, which has a resolution of approximately 13 km × 24 km at nadir [79]. This advancement enables more detailed and accurate mapping of gases like CO and NO₂, particularly in localized events such as forest fires. However, despite these improvements in spatial resolution, the effective analysis of gases with short atmospheric lifetimes, such as NO₂ (2–6 h [80], remains constrained by temporal resolution limitations—particularly in mid-latitude regions, where imaging is typically restricted to one observation per day [79,81]. This limitation poses a significant challenge for accurately capturing rapid fluctuations in gas concentrations over time [80]. The limited temporal frequency may result in a time lag, overlooking rapid changes in gas concentrations and potentially impacting the accuracy of temporal assessment in dynamic fire events.

Data quality is also another critical factor in satellite-based CO and NO₂ monitoring, as it directly influences the reliability of spatial analyses. In this study, Sentinel-5P data were filtered to include only pixels with QA values above 50%, minimizing uncertainties associated with cloud cover, sensor noise, and atmospheric interference [65]. This filtering approach enhances data consistency across the study area but introduces data gaps. On specific days, such as 30 July, 31 July, and 4 August, complete data coverage was achieved for CO and NO₂ concentrations, making these dates particularly suitable for reliable analysis (Figure 6c and Figure 9a). However, missing pixels in the CO dataset on other fire days complicate spatial distribution analysis and hinder direct comparisons with NO₂ levels. The limited temporal resolution of CO data, with only one daily observation, further restricts the ability to fill gaps due to the lack of near-simultaneous data [81]. While suitable for applications such as daily air quality monitoring [82], this temporal resolution is insufficient to capture the rapid fluctuations typical of dynamic events like forest fires or industrial activities [24,83,84].

5.2. Uncertainties Due to Topographical and Meteorological Factors

Our results reveal that elevation plays a crucial role in fire behavior and pollutant levels, with higher altitudes typically exhibiting cooler, more humid conditions that retain moisture and consequently reduce fire intensity and spread. This finding aligns with existing studies demonstrating that elevation influences fire dynamics by affecting temperature, humidity, and fuel moisture [85,86,87]. Additionally, our results indicate that UV-HH CO and NO₂ clusters are more common at lower elevations (0–500 m) compared to mid and higher elevations (500–1657 m), where such concentrations are less frequent. However, UV-LL clusters were observed across all elevations, indicating a broader spatial distribution for lower concentrations (Figure 9). These patterns are consistent with prior research showing that mountainous terrains often limit pollutant dispersion, resulting in pollutant accumulation in low-lying areas [88,89,90,91].

Higher altitudes also tend to receive more rainfall and experience delayed seasonal drying, resulting in milder fire behavior compared to lower elevations, where fuels dry earlier due to warmer temperatures and reduced rainfall [87,92,93,94]. This relationship directly influences the spread of fires, which subsequently affects smoke plume dynamics and the transport and dispersion of fire emissions [95].

Moreover, wind speed and direction significantly influence the dispersion and concentration of gases emitted during forest fires. Low wind speeds often result in poor ventilation, causing pollutants to accumulate in confined areas, whereas higher wind speeds promote both horizontal and vertical dispersion, reducing local pollutant concentrations [96,97,98]. For instance, Garcia-Menendez et al. (2013) reported a linear relationship between declining wind speeds and pollutant accumulation, while Lotrecchiano et al. (2020) demonstrated that low wind speeds (e.g., ≈11 km/h) limit dispersion near ignition points, whereas higher wind speeds (e.g., 54 km/h) enhance dispersion and reduce ground-level concentrations [96,97]. In our study, wind speeds remained predominantly below 54 km/h, except on the first day, contributing to localized CO and NO₂ accumulation (Figure 6c and Figure 9a). Wind direction, on the other hand, plays a critical role in determining the spatial trajectory of emitted gases [99]. More specifically, prevailing wind directions can influence the dispersion of emissions, leading to uneven spatial distribution and localized deposition. In this study, the predominant wind direction was from the north, which directed emissions toward lower altitude regions. The topography in these areas contributed to greater trapping of these gases. Furthermore, the combination of relatively low wind speeds (20–30 km/h) and prevailing wind direction resulted in the formation of dense UV-HH clusters near fire ignition points (Figure 6d and Figure 9b), as well as the accumulation of gases in low-lying areas (Figure 7).

In addition to wind speed, variations in meteorological factors such as temperature and humidity further complicate fire behavior and smoke dispersion. For instance, Figure 13a illustrates hourly fluctuations in these parameters, highlighting an inverse relationship between temperature and relative humidity. The satellite image captured between 08:35 and 08:39 UTC shows smoke dispersing in the NNE direction (Figure 13c), consistent with the measured wind direction (N-NNE, Figure 13b).

Although models such as Eulerian, Lagrangian, and Gaussian Plume are widely used to simulate the effects of topography and meteorological factors on fire behavior [102,103,104], their accuracy remains constrained by the unpredictable nature of forest fires and the complexities of rugged terrains like mountainous forests. The incorporation of additional datasets, including high-resolution meteorological and terrain data, is critical for improving model reliability and enhancing the understanding of fire behavior and pollutant dispersion in such challenging environments [104,105].

5.3. Fuel Types and Combustion Phases

The burned area within the fire zone consists of various CORINE LC classes, but the most dominant fuel type, by areal extent, is highly flammable coniferous forest species, which constitutes approximately 50% of the area according to the 2018 CORINE LC map (Figure A1). For the entire Manavgat district, the 2022 report by the World Wide Fund for Nature (WWF) [106], which provides a comprehensive analysis of forest composition and species distribution in the region, indicates that Calabrian pine, a coniferous species, accounts for 77% of the tree species, with smaller contributions from black pine (2%) and juniper (1%). The report highlights the predominance of coniferous vegetation, which, along with broad-leaved species (6%) and agricultural land (14%), collectively shapes the landscape of the region [106]. The prevalence of coniferous vegetation significantly affects fire behavior and emissions, as its dry needles and dense foliage make it highly flammable, enhancing combustion and promoting fire spread [57,58,107]. Studies such as Guo et al. (2020) support this, showing that coniferous species like pine release higher levels of CO and PM_2.5, while broadleaved species emit more CO₂ and nitrogen oxides (NO_x = NO + NO₂) [108,109]. Additionally, factors such as vegetation composition, moisture content, and physical arrangement further affect combustion intensity, duration, and emission rates [98,110]. Although coniferous forests dominate our study area (approximately 50%), the lack of species-level data limits a more detailed fuel type analysis. Despite this limitation, the predominance of coniferous forests partially reduced uncertainties related to vegetation variability by providing a rather generalized assumption to assess gas distributions and fire dynamics.

Variations in both Moran’s Indexes for CO and NO₂ (Figure 6d, Figure 9b, Figure 11 and Figure 12) can partly be attributed to differences in combustion phases [111]. Smoldering combustion, characterized by lower temperatures near the surface, primarily emits CO (Figure 14). In contrast, NO₂ is predominantly released during high-temperature flaming combustion (Figure 14) [27,80,112]. The balance between these phases is influenced by factors such as biomass composition, growth stage, moisture content, temperature, and wind speed [27,111]. Additionally, atmospheric lifetimes differ significantly: CO persists for 7–14 days, whereas NO₂ degrades within 2–6 h [80]. These differences were evident in our study, where CO concentrations were notably higher than NO₂ (Figure 8 and Figure 10). Furthermore, rapid interconversion between NO and NO₂ likely contributed to the lower NO₂ levels observed [80,111]. Local Moran’s I analysis further highlighted these distinctions. UV-HH CO clusters were observed in the high-concentration range (0.10100–0.36000 mol/m²) (Figure 8), while UV-HH NO₂ clusters were found in the mid-concentration range (0.00016–0.00040 mol/m²) (Figure 10). These patterns reflect the limited transport and shorter atmospheric lifetime of NO₂ compared to CO.

The co-occurrence of flaming and smoldering combustion phases within the same area (as observed in our study, Figure 14) introduces significant complexity to fire behavior and associated processes. Such interactions amplify spatial and temporal variability in fire intensity, emission composition, and plume dispersion, posing challenges for accurate monitoring and modeling [114,115]. This underlines the need for integrated approaches that consider the distinct yet interconnected dynamics of smoldering and flaming combustion phases.

6. Conclusions

The increasing frequency and intensity of forest fires globally have raised significant concerns about their environmental impacts, including the release of substantial quantities of GHGs and non-GHGs. These emissions exhibit complex spatial distribution patterns, necessitating detailed spatial analyses to understand the relation between fire dynamics, environmental conditions, and pollutant dispersion mechanisms.

This study examined the spatial dynamics of CO and NO₂ emissions during the 2021 Manavgat forest fires in Türkiye using Sentinel-5P satellite data and spatial statistical methods. The analysis revealed distinct yet interconnected distribution patterns of these gases, shaped by their chemical properties, combustion phases, meteorological conditions, and topography. The key findings are summarized as follows:

• Global Spatial Autocorrelation:
  • UV Global Moran’s I values for CO (0.84–0.93) and NO₂ (0.90–0.94) indicated strong spatial autocorrelation;
  • BV Global Moran’s I values (0.69–0.84) demonstrated significant spatial relationships between CO and NO₂, reflecting interconnected dispersion patterns.
• Local Spatial Autocorrelation:
  • UV Local Moran’s I analysis identified UV-HH and UV-LL clusters for CO and NO₂, reflecting differences in combustion phases, chemical interactions, and atmospheric lifetimes. CO concentrations exceeded 0.10000 mol/m², exhibited wide dispersion, while NO₂ concentrations, exceeding 0.00020 mol/m², remained localized near intense fire zones due to their shorter atmospheric lifetime;
  • BV Local Moran’s I analysis revealed overlapping BV-HH (High CO, High NO₂) and BV-LL (Low CO, Low NO₂) clusters, shaped by topography and meteorological factors.

The analysis further showed an inverse relationship between gas concentrations and elevation, with UV-HH clusters occurring at lower altitudes (0–500 m) and UV-LL clusters more common at higher elevations (500–1000 m or above). Low wind speeds contributed to localized gas accumulation, leading to pronounced UV-HH clusters near fire ignition points, a pattern further amplified by the flammability of coniferous vegetation.

To advance the understanding of pollutant dispersion during forest fire events, future research should integrate additional factors, such as FRP, fuel type and moisture content, meteorological variables (e.g., humidity and soil moisture), topographical features (e.g., slope and aspect), and smoke plume height. Expanding the scope to include diverse combustion characteristics would provide a more comprehensive understanding of emission patterns and enable improved modeling of fire-related atmospheric processes. Additionally, incorporating machine learning techniques into future investigations is essential for accurately and efficiently mapping the spatial distribution of forest emissions, providing valuable insights to guide more effective management and mitigation strategies.