The Republic of Guatemala, a Central American country situated between latitudes 13.5 to 17.8° N and longitudes 88.2 to 92.2° W. According to the National Institute of Statistics of Guatemala (INE, 2020), the country is administratively divided into 22 departments, which differ in terms of land use, population density, and economic activity. Guatemala encompasses a diverse range of ecological regions, topographical variations, and land use patterns. Its landscape is characterized by a complex zone of agricultural lands, forested highlands, volcanic zones, and rapidly urbanizing areas. Agriculture remains a cornerstone of the Guatemalan economy and landscape with extensive cultivation of crops such as maize, beans, coffee, sugarcane, oil palm, and bananas (FAO, 2019; Hervas, 2019; Lopez-Ridaura et al., 2019). Fertilizer application and livestock farming are particularly prevalent in rural regions. Meanwhile, increasing urbanization and industrialization, especially around Guatemala City and its metropolitan zone, have introduced additional anthropogenic activities, such as vehicular exhaust, industrial activity, and waste treatment processes.

Guatemala experiences a pronounced dry season (typically November to April) and rainy season (May to October) (Alfaro Marroquín and Gómez, 2019; Bardales Espinoza et al., 2019; INSIVUMEH, 2018; Orrego León et al., 2021), which strongly modulate fire activity, atmospheric transport conditions, and deposition processes relevant to NH₃ dispersion. Frequent wildfires, especially in the Petén department (northern area of Guatemala) and central highlands during the dry season, contribute further to the atmospheric burden of reactive nitrogen species. Topographically, Guatemala varies from sea level to 3000 m above sea level (a.s.l.) in the central plateau, and with volcanic peaks that reach 4200 m a.s.l. (IGM, 2002). Figure 1 shows the digital elevation model (DEM) of the topographic surface in meters over Guatemala using the European Space Agency (2024).

This study applies a multi-step analytical framework to investigate the spatiotemporal variability of satellite-based atmospheric NH₃ over Guatemala from 2015 to 2023. The analysis is based on a temporally concatenated monthly NH₃ total column Level 3 (L3) observations, from the IASI instrument onboard the MetOp A/B/C satellites launched in 2006, 2012, and 2018, respectively (Clarisse et al., 2009, 2019; Coheur et al., 2009). The concatenation produces the dataset, which serves as the basis for all spatial distributions, time series, climatological analyses, clustering, and correlation, incorporating both morning and night observations for enhanced representativeness and discarded negative values. While IASI-B maintains the spatial resolution of sensors A and C, integrating all three sensors could improve both temporal and spatial coverage, thereby yielding more accurate and robust results. While the study period spans 108 months, the concatenated analysis incorporates a total of 228 sensor months, corresponding of IASI A (January 2015–December 2020), B (January 2015–December 2023), and C (January 2020–December 2023).

IASI works with a cross-track scanning swath approximately 2200 km wide and provides measurements with a nadir pixel diameter of about 12 km. Positioned aboard sun-synchronous polar-orbiting satellites, the instrument enables near-global coverage twice daily during both daytime and nighttime corresponding to local solar overpass times of approximately 09:30 and 21:30, respectively (Clarisse et al., 2010; Van Damme et al., 2017; Whitburn et al., 2017). The NH₃ dataset was obtained from version 4 (v4.0.0R) (L3) of the ANNI-NH₃ retrieval product (Clarisse et al., 2023), and regridded to a 0.25° × 0.25° spatial resolution. A major enhancement in this version is the incorporation of the vertical column averaging kernel (AK), which is essential for aligning satellite retrievals with chemistry-transport model simulations by mitigating the influence of the priori NH₃ vertical profiles applied during retrieval. While spatial patterns remain consistent with previous versions, total column values are approximately 15 %–20 % higher due to improvements in the high-resolution infrared retrieval configuration (HRI). Version 4 offers improved uncertainty characterization and enhanced temporal uniformity of the dataset from 2007 to 2023 (Clarisse et al., 2023). The retrieval algorithm estimates a hyperspectral range index from IASI spectra and converts it into NH₃ total column densities using an artificial neural network (Van Damme et al., 2021; Franco et al., 2018; Whitburn et al., 2017). The uncertainty was quantified using the standard error and the relative uncertainty was derived from the distribution of the observation within of each grid cell (Sect. S1).

Additionally, NH₃ vertical profiles data were obtained from the ECMWF, Atmospheric Composition Reanalysis 4 (EAC4), produced by the CAMS. This global reanalysis provides NH₃ with mass fraction in the air, provided as a mass mixing ratio (kg kg⁻¹), the model is defined by 60 hybrid sigma-pressure levels. It provides a detailed representation of the atmospheric structure, particularly near the surface (∼ 0–100 m) where most emissions occur (Inness et al., 2019). These reanalysis datasets incorporate satellite observations and model simulations to provide globally consistent, spatially and temporally resolved atmospheric composition data (Flemming et al., 2017; Inness et al., 2019). In our study, the CAMS analysis of vertical structure and latitudinal distribution of NH₃ during a specific period of interest, rather than for a quantitative comparison with satellite of total column densities.

An unsupervised clustering approach was applied to classify NH₃ total columns field based on their statistical similarity. Clustering is a type of unsupervised machine learning that groups large datasets based on their similarity, allowing for more manageable and interpretable analyses (Ahmad and Dey, 2007; Pham et al., 2005). This analytical approach is particularly beneficial when there is limited or no prior knowledge about the underlying structure of the data (Gašparović et al., 2019). Among clustering methods, the k-means algorithm is one of the most commonly applied and reliable techniques. It partitions data into a predefined number of k clusters by iteratively determining central points, or centroids, for each group (Pham et al., 2005). The algorithm requires the number of clusters to be specified in advance (Ahmad and Dey, 2007; Mahata et al., 2020). The k-means operates using the squared Euclidean distance as its primary metric (Spencer, 2013) which quantifies the separation between two points within a Euclidean space of any dimension (Li et al., 2022). The Euclidean distance between two points, p and q, in a j-dimensional space is mathematically expressed as dp,q. 1dp,q=(p1-q1)2+(p2-q2)2+…+(pj-qj)2 The k-means algorithm divides a dataset into k clusters by aiming to minimize the variance within each group (Li et al., 2022). It operates on a set of input samples X=x1,…,xn, where each Xm=x1m,x2m,…,xjm represents an observation composed of j variables, and m ϵ X=1,…,n, with n being the total number of observations. These samples are categorized into k clusters S1,S2,…Sk with the goal of minimizing the total intra-cluster variance Si=1,…,k, as described in Eq. (2): 2Sargmin=∑i=1k∑X∈SiX-μi2 The notation X-μi represents the squared Euclidean distance between each data point (X) and its assigned cluster centroid μi within a multidimensional Euclidean space (Li et al., 2022). In this study, k-means clustering was utilized to examine the spatiotemporal patterns of NH₃ throughout the specified timeframe (Fouilloux, 2018; Pedregosa et al., 2011). To determine the most suitable number of clusters k, the elbow method was employed. This approach assesses the Cluster Sum of Squares (CSS) to identify the point at which the inclusion of additional clusters no longer significantly improves the model performance (Umargono et al., 2019).

The ESRI Land Use Land Cover (LULC) 2020 dataset, released in June 2021, was developed using deep learning models trained on more than 5 billion manually labelled pixel samples derived from the European Space Agency's Sentinel-2 satellite imagery with 10 m spatial resolution (Karra et al., 2021), enables land cover mapping at both national and local scales. According to (ESRI, 2021) these training samples were collected from 24 000 image tiles globally, each measuring 510 × 510 pixels, and reflect a wide range of land surface types. The LULC 2020 product is classified into nine classes: Water, Trees, Flooded vegetation, Crops, Built area, Bare ground, Snow/Ice, Clouds, and Rangeland (Karra et al., 2021). In this study, the analysis was restricted to classes related to Water, Trees, Flooded vegetation, Crops, Built area, and Rangeland.

Meteorological variables used in this study were obtained from the ERA5 reanalysis (Soci et al., 2024) dataset provided by ECMWF (Hersbach et al., 2020). Originally available at a horizontal resolution of approximately 31 km, the data were resampled to a 0.25° × 0.25° grid and both spatially and temporally to match the IASI NH₃ observation times and locations. The selected meteorological parameters were subsequently used in statistical analysis with atmospheric NH₃, these variables include 2 m air temperature (t2m), 10 m wind speed (ws), boundary layer height (BLH), relative humidity (RH, derived using 2 m air temperature and dew point temperature), and total precipitation (tp) from the monthly average reanalysis product, representing the twelve monthly precipitation values within each year, and monthly climatology for each calendar month. Pearson correlation matrix coefficient was calculated between monthly meteorological variables means and total column means of NH₃, including total fire counts per month, to quantify their linear relationship.

Active fire detections were obtained from MODIS Aqua/Terra Thermal Anomalies/Fire Locations product (MOD14/MYD14). Data was accessed via the Fire Information for Resource Management System (FIRMS) Near Real-Time (NRT), collection 6.1 (Giglio et al., 2021a). These data are produced by NASA's Land, Atmosphere Near-real-time Capability for EOS (LANCE) and provide global active fire information at a nominal spatial resolution of 1 km.

MODIS fire detection recognizes thermal anomalies in individual pixels using mid-infrared radiance signals. It assigns each detected fire's geographic location to the center of the corresponding 1 × 1 km² spatial resolution (Giglio et al., 2003, 2006; Justice et al., 2010). Each detection includes fire location, detection confidence, and FRP. Detection confidence, indicating the likelihood of an active fire in a pixel, is classified as low level: 0 % to <30 %, nominal level: 30 % to <80 %, high level: 80 % to 100 % (Giglio et al., 2016, 2021b, a; Hantson et al., 2013).

To reduce false detections, and ensure strong fire identification, only fire pixels with nominal to high confidence levels ≥70 % were retained for analysis, in accordance with recommendations from previous studies (Giglio et al., 2016, 2021b, a; Hantson et al., 2013). Fire activity was quantified using fire counts, defined as the number of detected fire pixels, which serve as a proxy for fire occurrence and spatiotemporal variability. Monthly fire counts were aggregated for each year to characterize seasonal and interannual variability in fire activity.

FRP served as an indicator of fire intensity with energetic significance providing as an additional filtering. Moreover, it quantifies the rate of radiative energy released from combustion and is directly related to biomass consumption (Wooster et al., 2003). In this study only fire detections with FRP values ≥40 Megawatt (MW) were included, thereby excluding low-intensity or small-scale fires to ensure consistency between temporal statistics and spatial fire distribution.

FRP is derived from MODIS 4 µm radiance measurement using the methodology proposed by Wooster et al. (2003) in Eq. (3): 3FRP=σ×Apixelα×τ4(L4-L4b) Apixel is the fire pixel area σ=5.6704 × 10⁻⁸ Wm⁻² K⁻⁴ is the Stefan–Boltzmann constant, α=3.0 × 10⁻⁹ Wm⁻² sr⁻¹ m⁻¹ k⁻⁴ is the MODIS empirical constant, τ4 is the atmospheric transmittance at 4 µm, and L4-L4b are radiances at 4 µm of the fire and background pixels, respectively (Wooster et al., 2003).

The burned analysis for this work was obtained by MODIS MCD64A1 Version 6.1 product, which combines observations from both Terra and Aqua platforms to produce a monthly global dataset gridded at 500 m that provides per-pixel burned-area and quality information (Giglio et al., 2018). The MCD64A1 MODIS burned area product integrates surface reflectance data with active fire detections to identify burned pixels based on temporal changes in a burn-sensitive vegetation index derived from shortwave infrared bands (bands 5 and 7) and spatial texture metrics (Dong et al., 2013; Freeborn et al., 2011; Giglio et al., 2003, 2021b; Justice et al., 2010). For each pixel, the product provides the estimated burn date as the ordinal day of the year, as well as quality assurance layers and uncertainty metrics. Pixels identified as unburned, water, or missing data are assigned to distinct flag values (Giglio et al., 2016).

Figure 2 shows the annual variability of NH₃ total columns over Guatemala from the period 2015 to 2023, while the corresponding anomalies relative to long-term are presented in Fig. S1. The respective boxplots show the distribution of monthly NH₃ values within each year. The annual medians of NH₃ remained relative stable, ranging from 3.95 × 10¹⁵ molecules cm⁻² in 2021 to 4.66 × 10¹⁵ molecules cm⁻² in 2020 (Table S1). In contrast, the standard deviation varied considerably between years, reaching a peak in 2020 of 5.83 × 10¹⁵ molecules cm⁻² following by 2016 with 5.26 × 10¹⁵ molecules cm⁻²and 2023 with 2.88 × 10¹⁵ molecules cm⁻².

The spatial distribution of NH₃ exhibited a consistent area-wide pattern characterized by elevated values in specific areas, and the corresponding annual anomalies relative to the long-term (Fig. S2). Frequently hotspots are observed in the northern (Petén), north-central (Quiché, Alta Verapaz, and Izabal), and along the southern coastal department of Escuintla. Conversely, lower annual mean NH₃ dominate in the central midlands, including the departments of Totonicapán, Sololá, Chimaltenango, Sacatepéquez, Guatemala, Jalapa, and Jutiapa. Interannual variability in hotspot intensity is evident, with particularly high values over Petén, Alta Verapaz and Quiché in 2016 and more widespread increased NH₃ across Petén, Quiché, Alta Verapaz, Izabal and Escuintla in 2020. Notably, 2018 presented lower annual mean NH₃ across Guatemala, however, the main spatial distribution of the main hotspots were still evident (Petén, Alta Verapaz, Izabal and Escuintla) showing frequent spatial patterns despite lower annual columns values.

The monthly climatology variability of NH₃ total columns is illustrated by boxplots in Fig. 3, showing a distinct seasonal pattern. Monthly medians were lowest in winter, reaching a minimum in February (3.52 × 10¹⁵ molecules cm⁻²) while January and March also exhibited relative low median values (4.25 × 10¹⁵ molecules cm⁻² and 4.00 × 10¹⁵ molecules cm⁻²), before increasing in the month of April with 7.34 × 10¹⁵ molecules cm⁻², and remaining elevated in May with 7.27 × 10¹⁵ molecules cm⁻² (Table S2). These months exhibit the largest standard deviation, that shows greatest variability with high NH₃. The values started decreasing thereafter in the month of June with a median of 4.83 × 10¹⁵ molecules cm⁻². From July through December, the median stabilizes, ranging between 3.58 × 10¹⁵ molecules cm⁻² (August) to 4.18 × 10¹⁵ molecules cm⁻² (November).

The spatial distribution of monthly climatology mean NH₃ displays the distinct seasonal pattern across Guatemala. From January to March, NH₃ remain generally low over most of the country, with only limited area exhibiting elevated columns (Escuintla). A substantial increase in both the total column values and spatial coverage of NH₃ is observed in April, with widespread elevated areas particularly in the northern (Petén), north-central (Quiché, Alta Verapaz, and Izabal), and southern coastal area (Retalhuleu, Suchitepéquez, and Escuintla). Specific defined areas with highest values primarily in the departments of Quiché, Alta Verapaz and Escuintla is observed in the month of May. Subsequently, a sharp decline in NH₃ is evident across the country from June through December. During these months, the spatial distribution of NH₃ is characterized by few hotspots and generally uniform low values in the central midlands (Totonicapán, Sololá, Chimaltenango, Sacatepéquez, Guatemala, Jalapa, and Jutiapa).

The spatial distribution of NH₃ in Guatemala, based on the overall mean for the period 2015–2023, reveals several distinct hotspots areas, most notably in the southern coast (Escuintla), central-northern (Alta Verapaz and Quiché and Izabal), and northern (Petén) (Fig. 4a). Among these, the most prominent hotspot is observed in the central-northern area, which exhibits the highest mean NH₃ total column during the study period.

The k-means clustering partitions the country into three emission zones (Fig. 4b), each with distinct land cover profiles when integrated with ESRI data (Fig. 4c). Cluster 1 with smallest area (∼ 28 254 km²) (red) associated to the highest NH₃, is located in the northern (Petén), central-north (Quiché, Alta Verapaz, Izabal), and southern regions (Escuintla, Suchitepéquez, Retalhuleu and San Marcos) is dominated by extensive tree cover (53.2 %; 15 024.6 km²), together with rangeland (24.5 %; 6934.2 km²) and cropland (16.7 %; 4710.8 km²), while built-up areas are limited to 3.4 % (948.6 km²) as shown in Fig. 4d. The complete land cover distribution across clusters is provided in Table 1.

Cluster 2 with a largest extend area around 45 662 km² (yellow) represents moderate zones forming a transition around Cluster 1 and extending into northern (Petén), north-central (Huehuetenango, Quiché, Alta Verapaz, Santa Rosa, Izabal), and south (Quetzaltenango, Escuintla, Suchitepéquez, Retalhuleu, Santa Rosa, Jutiapa and San Marcos) (Fig. 4e). It is dominated by tree cover (58.9 %; 26 891.6 km²) and rangeland (30.3 %; 13 844.6 km²), with a comparatively small share of cropland (6.0 %; 2724.7 km²). Built-up areas account for 2.8 % (1261.8 km²), slightly larger in absolute terms than in Cluster 1.

The intermediate extent area Cluster 3 with ∼ 34 640 km² (green), associated with the lowest NH₃, covers the central region with departments of Huehuetenango, San Marcos, Totonicapán, Sololá, Chimaltenango, Sacatepéquez, Guatemala, El Progreso, Jalapa, Chiquimula and Jutiapa. This cluster is characterized by high rangeland cover (34.3 %; 11 893.3 km²) and substantial tree cover (47.1 %; 16 298.9 km²), while cropland is minimal (3.1 %; 1088.6 km²) (Fig. 4f). Notably, built-up areas occupy the largest proportion across clusters (14.8 %; 5137.1 km²).

Figure 5 illustrates the annual variability and monthly climatology of meteorological variables over Guatemala derived from ERA5 reanalysis. Figure 5a, c, exhibits the BLH and precipitation, respectively. The BLH shows fluctuations with the highest median recorded in 2019 (524 m). contrasted by the lowest median observed in 2017 (415 m) which also recorded the minimum value (305 m) (Table S6). Although the annual medians varied moderately, the peak monthly BLH values detected during 2020 (791 m) indicate high intra annual variability during that year. At the monthly scale, April shows the highest BLH (median 671 m), while the lowest occurred in October (median 348 m). Furthermore, the distribution of the twelve-monthly total precipitation values within each year shows that the median monthly precipitation was highest in 2020 (164 mm) and 2022 (142 mm), while the lowest median monthly occurred in 2019 (89 mm). The maximum monthly precipitation values exceeded 840 mm in 2017, 2021, and 2022, reaching an overall maximum of 1283 mm in 2022. The monthly climatology indicates that precipitation increased rapidly from May onward, with median monthly precipitation increasing from 49 mm in April to 135 mm in May, peaking in June (269 mm), before remained elevated through October. Subsequently, precipitation declined during the dry season, with the lowest median climatological precipitation occurred in February (28 mm).

Figure 5b, e presents the air temperature and relative humidity. The highest annual median temperature occurred during the 2023 (22.8 °C), whereas the maximum was detected in 2020 (27.9 °C). The lowest temperature was recorded in 2018 at 20.2 °C. Additionally, the warmest month is May (median 26.4 °C) and the coldest is January (21.9 °C). Relative humidity showed the highest annual median in 2016 (81 %), while the lowest annual median occurred in 2019 (76 %). The minimum value was detected in 2020 with 63 %, which also featured the largest relative humidity values. The years 2017 and 2022 exhibited the maximum values with 87 %. Monthly medians indicated that the lowest relative humidity occurs in April (71 %) and the highest in October (84 %).

In order to connect the previously described NH₃ patterns with biomass-burning, Fig. 6 combines the annual spatial distribution of MODIS active fires detection with the distribution of monthly fire counts in Guatemala during the study period. The fire activity reveals a marked interannual variability. The highest monthly fire count occurred in 2023 (median 59), followed by 2016 (median 50), where the lowest was observed in 2022 (median 19), (Table S4). Fire activity was also relatively low in 2018 (median 28) and 2021 (median 30.5) respectively. Although 2020 had a relatively low median monthly fire count (median 26) it exhibited the largest monthly variability (standard deviation 272) and included the highest monthly fire count during the study period (957). In contrast 2016 and 2017 showed the highest total annual fire counts (1779, 1783), while 2022 had the lowest total (807).

The spatial distribution of detected fires corresponds with the temporal variability in fire counts, showing pronounced interannual differences in both fire occurrence and spatial heterogeneity. Years with elevated fire occurrence (2016, 2017, 2019, 2020, and 2023), showed widespread spatial extents of fire detections, with recurrent hotspots concentrated primarily in the northern (Petén) and along the southern area (Escuintla, Suchitepéquez, Retalhuleu, and Santa Rosa). In contrast, years of reduced fire activity, including 2018, 2021 and 2022, showed decreased detections, and reduce spatial confined distribution of fires.

The monthly climatology of fire activity in Guatemala is illustrated in Fig. 7. The boxplots show a clear seasonal pattern in detected fires. Fire activity is relatively low at the beginning of the year, with a median of 78 in January, and 105 in February (Table S5). A substantial rise occurs in March (median 211), while April exhibited the highest fire activity (median 659) detected fires and a maximum of 957. Consequently, the month of May demonstrated a decrease of in the number of fires (median 238). Moreover, an observed sharp decline in June (median 15), and the lowest fire activity occurs in September (median 3). Finally, slight increase is observed at the end of the year in December (median 46).

The spatial distribution of FRP across Guatemala also confirms this seasonal pattern and shows the shifts in fire location and intensity. During the early months of the year, fire activity is concentrated mainly along the southern areas (Escuintla, Suchitepéquez, Retalhuleu, and Santa Rosa). As the fire season advances, the activity shifts northward with central-north (Huehuetenango, Quiché, Alta Verapaz, Santa Rosa, Izabal), and northern (Petén). In November and December, the fire remerged in the southern coast (Escuintla, Suchitepéquez, and Retalhuleu).

The variability of NH₃ total columns over Guatemala reflects the interaction between persistent agricultural sources, and episodic biomass burning. Although annual median of NH₃ are relatively stable throughout the study period (3.95–4.66 × 10¹⁵ molecules cm⁻², Table S1), interannual differences were associated with increases in maximum values and standard deviations in some specific years (2016, 2019, 2020, and 2023) indicating that episodic events were superimposed on a relatively stable background. Across the study period (9 years) annual median of NH₃ varied about 18 % where the maximum values and standard deviations increased markedly during nine years with increased fire activity. This indicates that episodic events primarily affected the magnitude and variability of NH₃ rather than its background values. These observations suggest that persistent agricultural activities likely maintain the regional NH₃ background, whereas biomass burning primarily contributes to short term enhancements specifically during the dry season.

The spatial patterns further support this interpretation. Increased NH₃ columns were consistently observed in the departments of Petén, Alta Verapaz, Quiche, Izabal and Escuintla during the study period, including years with relatively low annual NH₃, suggesting the presence of continuous NH₃ sources rather than isolated episodic events. Land cover analysis indicates that cluster 1, which contains the highest NH₃ columns, is characterized by a greater proportion of cropland (16.7 %) while tree cover remains extensive across all clusters. Although rangeland is also abundant, its relatively high proportion in cluster 3 (low NH₃) suggests that rangeland alone cannot explain the observed NH₃ distribution. Instead, the larger cropland area in cluster 1 suggests a stronger contribution of agricultural activities to elevated NH₃ total column. The seasonal timing of these persistent hotspots is also consistent with agricultural practices in Guatemala, where land preparation, shifting cultivation, and biomass burning for crops are concentrated during the dry season (Liu et al., 2024b; Monzón-Alvarado et al., 2012; Ríos and Raga, 2018). These findings are consistent with previous studies showing that agricultural practices dominate global source of atmospheric NH₃ (McDuffie et al., 2020; Sutton et al., 2013), and with satellite observations from South Asia (Abeed et al., 2023), and Western Europe (Van Damme et al., 2018). Furthermore, cluster 1 has the smallest spatial extent but the highest NH₃ values, demonstrating that NH₃ is more strongly associated with localized agricultural land use than the total area occupied. In contrast, the central highlands exhibit low NH₃ despite the presence of urban areas (Huehuetenango, Quiché, San Marcos, Sololá, Guatemala, Fig. 4), suggesting that urban emissions play a secondary role at the national scale.

While persistent agricultural activities appear to maintain the NH₃ background, the pronounced seasonal cycle demonstrates that additional processes regulate the short-term variability. Monthly NH₃ levels reached their highest values during the months of March–April–May (MAM, mean 7.49 × 10¹⁵ molecules cm⁻²), which coincides with the regional fire season (mean total fire count 3321), and also increases the variability where events of high columns of NH₃ appeared. This seasonal increase points to episodic biomass burning and meteorological conditions that modulate NH₃ above persistent agricultural background.

Although persistent agricultural activities likely maintain the background atmospheric NH₃ (Sect. 4.1), the seasonal and interannual variability is primarily explained by biomass burning. Monthly NH₃ exhibited a strong positive relationship with fire activity (r=0.65, Fig. 8), with the highest values occurring during the fire season (MAM). During this period, mean NH₃ columns reached 7.49 × 10¹⁵ molecules cm⁻², coinciding with a mean fire count of 3321. Following the onset of the wet season, fire activity declined by approximately 94 % between May and June, accompanied by a 42 % decrease in mean NH₃ (9.53 × 10¹⁵ molecules cm⁻² to 5.56 × 10¹⁵ molecules cm⁻²). This indicates that biomass burning is the dominant driver of short-term NH₃ enhancements superimposed on the persistent agricultural background.

The spatial distribution of fires further supports this interpretation, with recurrent fire hotspots in northern Guatemala (Petén) and the southern coast (Escuintla). Petén shows the strongest area of fire activity, including the officially designated protected areas that correspond to the large Maya Biosphere reserve (CONAP, 2016). The observed pattern of fire activity is predominantly anthropogenic. Previous studies indicate that most fires globally are human induced (FAO, 2007), and in Central America lighting occurrence is minimal during the dry season (November–April) (Kucieńska et al., 2010) when fire activity is highest. The coincidence between fire occurrence hotspot, cropland regions, and elevated NH₃ suggests that agricultural burning and land preparation practices substantially contribute to the seasonal increase of NH₃.

Meteorological conditions further modulate the magnitude of biomass burning on atmospheric NH₃. Periods with increased fires activity were associated with higher temperature (r=0.42) and BLH (r=0.76), indicating that warmer conditions promote both fire occurrence and vertical distribution of NH₃. Although the relationship with monthly precipitation was weak (r=-0.07), the seasonal transition to the rainy period coincided with a substantial reduction in both fire activity and atmospheric NH₃, consistent with enhanced wet removal during the we season. Years characterized by extreme temperature conditions (2016, 2020, 2023) from Table S6, were consistent reported by INSIVUMEH (2023). These years occurred during the episode of El Niño Southern Oscillation (ENSO), which modulates temperature and precipitation conditions in the region of Central America (Anderson et al., 2023). Similar patterns were reported for Amazonia, where NH₃ emission from biomass burning exhibit strong spatial and temporal variability, with interannual peaks linked to severe droughts and late fire-season activity, highlighting the influence of ENSO events (Anderson et al., 2023; Whitburn et al., 2016). Similarly, in Southeast Asia, particularly Indonesia, widespread human-induced fires for land clearing burned extensive tropical forest and peatlands (Chang et al., 2021; Whitburn et al., 2016).

The mechanisms discussed above converged during April 2020, producing the most intense atmospheric NH₃ episode observed during the study period. This event provides an illustrative example of how a combination of persistent agricultural activity, biomass burning, and favorable meteorological conditions can produce extreme NH₃ levels. This month recorded the highest NH₃ total column observed during the study period (2.3 × 10¹⁶ molecules cm⁻², Fig. 9), despite occurring shortly after the implementation of COVID-19 lockdown measures in March 2020. Similar increases in atmospheric NH₃ during lockdown conditions have been reported in previous studies, which attributed this trend to the persistence of agricultural activities and biomass burning (Evangeliou et al., 2025; Lovarelli et al., 2021; Viatte et al., 2021; Xu et al., 2022). The spatial distribution indicates distinct NH₃ hotspots concentrated in the northern (Petén), north-central (Quiché) and southern (Escuintla), with the highest values observed in the department of Petén, reaching up to (6.4 × 10¹⁶ molecules cm⁻²). Wind vectors were characterized by strong easterly winds (∼ 3 meter per second, m s⁻¹) that dominated the northern region, potentially facilitating the transport of NH₃, while central areas of Guatemala experienced moderate to stagnant wind conditions.

The vertical distribution of atmospheric NH₃ further supports the interpretation that the April 2020 event was dominated by near surface NH₃ as evidenced by the averaged vertical profiles of the mass mixing ratios (kg kg⁻¹). The zonal cross-section of NH₃ shows the highest values below approximately model level 56, located between 14 and 15° N latitude (Fig. 10a). This indicates that the enhanced columns originated primarily within the lower troposphere rather than from long-range transport. This vertical confinement is consistent with the values register from agricultural activities and biomass burning which remained concentrated near the surface during the event. Furthermore, the area average vertical profiles in Fig. 10b demonstrate the exceptional character of the April 2020 episode. Mean NH₃ mixing ratios (∼ 1.0 × 10⁻⁹ kg kg⁻¹) exceeded both the spring average (∼ 0.9 × 10⁻⁹ kg kg⁻¹) and the annual 2020 mean (∼ 0.5 × 10⁻⁹ kg kg⁻¹), indicating that the event represented a substantial enhancement beyond normal seasonal conditions rather than typical dry season variability.

The concurrent fire activity strongly supports biomass burning as the principal driver of this exceptional event. April 2020 recorded the highest monthly number of detected fires across Guatemala throughout the entire study period (957 fires; Figs. 11, S5). Most detections occurred in northern Guatemala (Petén) accounting for 674 active fires (Table S7, Fig. S6), alongside widespread burning across the north-central (Quiché, Alta Verapaz, and Izabal), and southern areas (Escuintla, Santa Rosa, and Jutiapa). The spatial correspondence between fire hotspots, burned areas, and NH₃ maxima provides additional evidence that biomass burning substantially amplified atmospheric NH₃during this month.

The concatenated dataset integrated a total of 228 months of observation data across the 108 months of the study period, providing stable timeseries for analyzing NH₃ variability over Guatemala. A comparison during the overlapping period shows a strong correlation between sensors, with R2 above 0.9 and regression slopes close to 1 (Fig. S7). This indicates that the inter-sensor differences are relatively small compared to the observed variability over time. Furthermore, the biases represent only a small fraction of the mean NH₃ column, supporting the use of a concatenated dataset for long-term analysis. This consistency is further supported by the monthly L3 total column from individual IASI sensors (Fig. S3 and Table S3), aligning with previous time-series studies for the three sensors and demonstrating robustness over 10 different remote regions (Clarisse et al., 2023; Van Damme et al., 2014).

The observation sampling density remains relatively stable over time, ranging from 1.85–2.65 obs. grid cell per month with mean annual pixel counts between 978 and 1492 (Table S10). The year 2020, which exhibits the highest NH₃ is associated with a lower observation density (1.97 obs. grid cell per month), indicating that the observed maximum is not driven by increased sampling effort. This is further supported by the absence of a significant relationship between sampling frequency and NH₃ levels (R2=0.028, Fig. S8), confirming that the observed variability is not an artefact of sampling bias. Moreover, the standard errors (Table S9) remain small relative to the magnitude of interannual variability. This indicates that the observed differences between years exceed the measurement uncertainty and robustness of the NH₃ column patterns. On a seasonal scale, a clear pattern emerges: peak columns coincide with months of high sampling density (MAM), while lower columns during the rainy season occur under a 17 % reduction in observation density, likely due to increased cloud cover. The lower sampling density detected is located over high altitude regions such as Sierra de las Minas (Fig. S9), which is associated with persistent cloud forests (Chaluleu, 2020; Holder, 2003, 2006).

Although topography varies considerably across Guatemala (Fig. 1), the major NH₃ hotspots are more closely associated with agricultural land use than with elevation gradients. The observed spatial distribution of NH₃ does not simply follow elevation patterns. Elevated NH₃ columns are observed over low- and mid-elevation agricultural areas (Petén, Alta Verapaz, Quiché, Izabal and Escuintla). In contrast, some low-elevation areas exhibit relatively low NH₃, despite having some of the highest satellite observation densities (Fig. S9). This pattern is consistent with the limited extent of agricultural activities in these specific low-yielding zones, rather than indicating an elevation-related retrieval artifact. These observations indicate that the spatial variability of NH₃ primarily reflects differences in total column sources rather than terrain-induced biases.

Finally, the use of the L3 dataset, based on averaging L2 retrievals, ensures that the observations retain the original retrieval signal. Guatemala's tropical location also provides favorable thermal contrast conditions for infrared NH₃ retrievals, supporting robust NH₃ detection despite a relatively moderate background total column.