We utilized active fire observations from the Visible Infrared Imaging Radiometer Suite (VIIRS) sensors carried by the Suomi NPP and NOAA-20 platforms. These measurements capture combustion-related thermal signals at approximately 375 m resolution. To facilitate large-scale analysis, individual fire detections were spatially aggregated onto a 0.1° × 0.1° global grid, and the number of detections within each cell was used as a metric of fire occurrence. Daily wildfire activity was characterized using the combined daytime and nighttime VIIRS observations, accounting for both fire occurrence and intensity. Fire detections were spatially overlaid with a forest land-cover mask to extract events occurring within forested pixels, and only grid cells with more than one active fire detection during 2012–2024 were retained to ensure data robustness. To enhance data reliability, fire pixels flagged as “low confidence” (confidence level “l”) were excluded from the analysis.

Land cover information was characterized by using the MODIS land cover type dataset at 500 m resolution, available through NASA's Land Processes Distributed Active Archive Center (LP DAAC). Topographic characteristics were represented using the General Bathymetric Chart of the Oceans (GEBCO) 2024 gridded data, which provides global elevation and bathymetry data at 15 arcsec resolution (∼ 500 m) (Harper and Sandwell, 2024). To ensure spatial consistency, land cover and elevation datasets were resampled to match the analysis grid. Forest categories and elevation values at fire locations were then extracted using nearest-neighbor interpolation. Forested regions were grouped into five major types: evergreen needleleaf (ENF), evergreen broadleaf (EBF), deciduous needleleaf (DNF), deciduous broadleaf (DBF), and mixed forest (MF). Elevation was stratified into 20 zones using 100 m intervals below 2000 m. Given the relatively low number of fire detections at higher elevations, all regions above 2000 m were combined into a single category to ensure sufficient sample sizes and enhance the robustness of subsequent trend analyses.

We employed the Canadian Forest Fire Weather Index (FWI) system to characterize weather conditions that favor wildfire development and to assess climate-driven variability in fire danger (Su et al., 2025; Liu et al., 2023; Van Wagner, 1987). The index is calculated from standard meteorological inputs, including near-surface air temperature, wind speed, relative humidity, and accumulated total precipitation over the preceding 24 h (Van Wagner, 1987). Daily FWI data with a resolution of 0.25° × 0.25° from 2012 to 2024 are obtained from the historical fire danger index archives. Supplementary meteorological parameters, such as 2 m air temperature (2mT), 10 m wind speed (10mW), and the max daily total precipitation, were extracted from the ERA5 reanalysis produced by the European Centre for Medium-Range Weather Forecasts (ECMWF). Monthly means of FWI and meteorological variables were calculated for subsequent analyses with forest fire activity.

Size-resolved aerosol optical depths at 500 nm, including the fine (FAOD) and coarse (CAOD) fractions, were estimated using SIDN (Simultaneous FAOD–CAOD Inversion Deep Neural Network) algorithm. The inversion is physically constrained by incorporating interactions between the two aerosol size classes, thereby enhancing retrieval reliability (Luo et al., 2024; Chen et al., 2025). SIDN integrates multi-source satellite observations with meteorological reanalysis data and is trained against high-quality AERONET measurements using a shared representation learning strategy based on the EntityDenseNet architecture. By dynamically balancing FAOD and CAOD contributions during training, SIDN improves retrieval accuracy and internal consistency across diverse aerosol regimes. A comprehensive description of the model architecture, training procedure, and validation is provided in the above references and is not repeated here. Extensive evaluations against independent ground-based observations demonstrate that SIDN outperforms conventional satellite products, yielding reduced uncertainties and improved spatial completeness, particularly under cloudy and complex surface conditions. In addition, the deep-learning framework trained with AERONET observations substantially improves retrieval accuracy and stability across different aerosol regimes, making the dataset particularly suitable for investigating wildfire–aerosol interactions on a global scale. The resulting daily global FAOD and CAOD dataset spans 2012–2024 at a spatial resolution of 0.5°. Because the spatial resolution of the AOD data is coarser than that of the forest fire dataset, a nearest-neighbor approach was used to match the datasets. This method preserves the original aerosol observations without introducing artificial smoothing or interpolation effects and has been widely applied in satellite data matching studies. Although the spatial resolution mismatch may introduce some uncertainties, its impact is expected to be limited for the large-scale analyses conducted in this study.

Temporal trends in wildfire activity were assessed using linear regression applied to annual fire pixel counts aggregated at 0.1° × 0.1° resolution. For each grid cell, a linear regression was fitted between the natural logarithm of annual fire pixel counts and time, with statistical significance determined at p<0.05. The regression slope was exponentiated to estimate the annual percentage change in fire occurrence. For the global and forest-type–specific assessments, annual records were further processed with a 3-year centered moving average, which filters out short-term interannual noise. To verify the stability of the detected trends, we employed the Mann–Kendall non-parametric test in combination with the Theil–Sen estimator, which provide outlier-insensitive assessments of trend direction and magnitude. The Theil–Sen slope was then derived for grid cells showing significant changes, expressed as SFC=medianFCj-FCij-i,2012≤i≤j≤2024 where SFC denotes the median rate of change, and FC_i and FC_j represent fire pixel counts at moments i andj, respectively.

Trends in AOD over 2012–2024 were analyzed using the same logarithmic linear regression framework to ensure comparability with fire pixel count trends. AOD values were restricted to the range 0–5 to exclude extreme outliers.

To quantify the influence of forest fires on aerosol loading, linear regressions of the form AOD = a+k×log⁡10[fires] applied separately to FAOD and CAOD. The regression slope (k) represents fire sensitivity, with larger values indicating stronger aerosol responses to fire activity. The log₁₀[fires] was adopted as the explanatory variable, following common practice in wildfire research (Zhao et al., 2024; Abatzoglou and Williams, 2016). We applied linear regressions to individual 1° × 1° grid cells that contained no fewer than 10 fire observations. Where statistically significant associations were identified (p<0.05), both the slope parameter (k) and the corresponding R2 values were mapped to illustrate their spatial variability. Differences in regression slopes among grid cells were evaluated using two-tailed t tests.

For the elevation-based analysis of forest fire activity, all monthly fire, meteorological, vegetation, and topographic variables were spatially resampled to a 1° × 1° grid. For each grid cell, multiple linear regression model was conducted with log₁₀[fires] as the response variable and FWI, leaf area index (LAI), 2mT, 10mW, and daily maximum total precipitation as predictors. Year and month were additionally included to represent long-term trends and seasonal variability. We employed ordinary least squares (OLS) models to obtain baseline relationships, and statistical inference was conducted using heteroskedasticity-robust standard errors (p<0.05). To mitigate multicollinearity among predictors, ridge regression with L2 regularization was additionally applied after standardizing all variables, with the optimal regularization parameter selected via cross-validation. Variance inflation factors (VIFs) were calculated to diagnose collinearity, and partial correlation analysis was applied to isolate the independent effects of each driver after accounting for temporal influences. Regression coefficients and model performance metrics were subsequently summarized within elevation bins to characterize altitude-dependent fire–environment relationships.

Analysis of satellite-derived fire pixel counts across global forested regions from 2012 to 2024 reveals a modest but statistically significant increasing trend. Spatially, 51.1 % of forest grid cells exhibit positive trends, exceeding the 48.9 % showing negative trends, with a global mean linear increase of 2.89 % yr⁻¹ (p<0.05, Fig. 1A). Our results indicate a growing number of forest fires, in agreement with the previously reported global increase in forest burned area and mean fire pixel counts over last two decades (Zheng et al., 2021; Yu and Ginoux, 2022). Geographically, significant positive trends in fire pixel counts are concentrated in North America, northern South America, and central Africa, whereas predominantly negative trends occur in Southeast Asia, western and central Siberia, and eastern Europe (Fig. 1A). The spatial patterns derived from linear regression are consistent with those identified using the Mann-Kendall test (Fig. 1B). Notably, regions exhibiting increasing trends largely overlap with areas of enhanced burned area and fire-driven carbon emissions in the tropical moist broadleaf forest pyrome identified by Jones et al. (2024) for the period 2001–2023. Increasing fire activity has also been reported in selected extratropical forest ecoregions, including Canada (Wang et al., 2025) and Northeast China (Liu et al., 2023).

Considerable heterogeneity in fire activity is found across forest types. EBF accounts for the largest proportion of fire pixel counts globally, whereas ENF exhibits the strongest positive trend in fire occurrence (Fig. 1C). By comparison, no statistically significant changes are observed for DNF and MF. These forest types are primarily distributed across western and central Siberia (Fig. 2A), regions characterized by declining fire activity (Fig. 1A). Fire occurrence and severity are strongly mediated by vegetation characteristics, including forest type and fuel structure (Wang et al., 2025). High fire frequency and density are commonly associated with evergreen needleleaf forests, such as those dominated by Chir pine (Pinus roxburghii), which are characterized by high-intensity fire regimes (Kumar and Kumar, 2022). This elevated fire susceptibility is attributed to needle-shaped foliage, high resin content, and ladder-like branch architecture, which collectively enhance fuel flammability and vertical fire spread (Alexander and Cruz, 2011; Wang et al., 2025).

Aerosols influence the atmospheric radiation balance through the scattering and absorption of radiation (Andreae et al., 2004; Blanchard-Wrigglesworth et al., 2025). Owing to their distinct optical properties, different aerosol types exert varying climatic effects (Lin et al., 2021), particularly between fine-mode and coarse-mode particles (Li et al., 2025). To better investigate aerosol responses to forest fire activity, aerosol loading is therefore characterized using FAOD and CAOD. A consistent positive association between FAOD and fire pixel counts is observed across global forest regions from 2012 to 2024, whereas no comparable trend is detected for CAOD (Fig. 3). It should be noted that FAOD and CAOD may also be influenced by aerosol sources unrelated to wildfire emissions; therefore, the comparison with fire activity should be treated as a first-order diagnostic of their potential linkage rather than a strict attribution of aerosol sources. The fire–FAOD relationship was quantified using linear regression of the form (FAOD = a+kF×log⁡10[fires]) applied to each 1° × 1° grid cell. Widespread statistically significant positive kF values indicate that fine-mode particles dominate wildfire-related aerosol loading (Fig. 4A), consistent with previous evidence that biomass burning predominantly contributes to the fine-mode portion of the aerosol population (Yan et al., 2024; Yang et al., 2021b; Kaskaoutis et al., 2014; Pei et al., 2025).

The kF slope values exhibited variation across forest types, with the highest values magnitude observed in DNF (0.25 ± 0.24), followed by ENF (0.13 ± 0.18), EBF (0.09 ± 0.14), MF (0.06 ± 0.13) and DBF (0.03 ± 0.08) (Fig. 4b). This ranking is consistent with the spatial pattern of average FAOD across these forest types (Fig. 4a). Such differences can be attributed to species-dependent emission strengths, as dominant taxa contribute unevenly to wildfire-derived aerosols. For example, within ENF regions, Pseudotsuga and Picea account for 15 % and 71.1 % of total FAOD contribution from wildfire, respectively; among broadleaf forests, the genus Quercus alone contributes 50.6 % (Chen et al., 2025). These patterns underscore that aerosol emissions from forest fires depend not only on burned areas but also on fuel type and combustion efficiency (Keywood et al., 2013). In contrast, no significant positive kC values are detected for CAOD (Fig. 4C). The average of kC values in different forest types was close to 0 with minimal variation (Fig. 4d), indicating the absence of a systematic forest fire influence on coarse-mode aerosol loading. In contrast to the large aerosol particles associated with large fire events, such as pyroCb emissions (Li et al., 2025) or coarse dust produced by post-fire dust outbreaks caused by reduced vegetation cover and soil moisture (Yu and Ginoux, 2022), the results presented here mainly represent the global mean response within forest fire regions. Although large fire events may generate unusually high aerosol loading or coarse-mode dust, such episodic events are not fully captured in the large-scale averages analyzed in this study.

To evaluate how topography shapes forest fire activity globally, we examined the elevational dependence of fire occurrence from 2012 to 2024. A clear and statistically significant decline in mean annual fire pixel counts with increasing elevation was identified (Fig. 5A). Forest fires were predominantly concentrated at low elevations, with 66.6 % of total fire detections occurring below 600 m, followed by 19.2 % between 601 and 1000 m and 13.3 % between 1001 and 2000 m. Only 0.9 % of fires were detected above 2000 m (Fig. 5A). This inverse relationship between elevation and fire activity agrees with previous regional studies conducted in the Canadian Rocky Mountains (Rogeau and Armstrong, 2017) and across forested regions of China (Tian et al., 2013; Ma et al., 2020). Similar elevational dependence is also evident in the interannual trends. Both the multiyear mean fire pixel counts, and the corresponding annual trends exhibit a pronounced decrease from low to high elevations (Fig. 5A, B). However, Xu and You (2022) reported increasing burned areas in global high-mountain regions above 3000 m, primarily focusing on high-mountain ecosystems that include alpine and other non-forest environments. In contrast, the present study examines wildfire activity within global forest regions, where most fires occur at low and mid elevations. Differences in vegetation type, fuel availability, and fire regimes between forested landscapes and alpine environments may partly explain these contrasting patterns.

Notably, while the largest absolute number of fires were observed between 100 and 200 m, the strongest positive trend in fire activity occurred near ∼ 300 m before declining at higher elevations. This discrepancy suggests that recent fire activity has increasingly concentrated around the ∼ 300 m elevation range, potentially reflecting the combined influence of climate variability and fire management practices. In contrast, a weak negative trend in fire occurrence was observed below 100 m, which may be associated with more effective control of ignition sources in densely managed lowland regions (Wang et al., 2022b).

In contrast to the monotonic decline in fire occurrence, FAOD exhibits a distinct elevational pattern. While FAOD trends below 600 m broadly track those of fire activity, its multiyear mean values peak at mid-elevations (600–1400 m) rather than at low elevations (Fig. 5C, D). This divergence indicates that FAOD is not solely controlled by local fire occurrence but is also influenced by forest composition and topographically modulated by smoke transport processes. Around 600 m elevation, forest composition shifts toward a higher fraction of ENF (Fig. 2B), which exhibits stronger FAOD sensitivity to fire activity (kF=0.14±0.18) than EBF (kF=0.09±0.14). In addition, smoke particles released by low-elevation fires may be carried to higher levels of the atmosphere via aerosol self-lifting processes associated with radiative heating and convective motions (De Laat et al., 2012; Ohneiser et al., 2023; Xu et al., 2025), thus elevating FAOD levels at mid-elevations. Above 600 m, the strengthening FAOD trends with elevation diverge from the concurrent decline in fire occurrence, implying the involvement of additional processes beyond local fire emissions. In addition, with fewer than 50 grid cells above 1500 m exhibiting statistically significant trends, the representativeness and accuracy of the computed global FAOD trend may be compromised. Collectively, these results suggest that mid-elevation regions (600–1400 m) are disproportionately influenced by smoke aerosols from forest fires at a global scale.

To explain the contrasting elevational patterns of forest fire occurrence identified in Sect. 3.2, we next examine the meteorological and ecological drivers governing fire activity along with the elevation gradient. Vegetation fire activity is governed by multiple interacting factors, including ignition sources, fuel supply and moisture, fire-conducive weather, and forest management practices (Wang et al., 2022a; Clarke et al., 2025). However, the relative contributions of these drivers change with elevation. Notably, the FWI – often used to characterize the meteorological fire danger – exerts little explanatory power, with regression coefficients near zero at all heights (Fig. 6). Partial correlation coefficients were negative between 300 and 2100 m, indicating a weak association between FWI and fire occurrence after accounting for elevation. The weak agreement observed here aligns with earlier reports (Su et al., 2025; Wang et al., 2025) and arises primarily from mismatches in spatial and temporal resolution. FWI primarily represents short-term surface meteorological conditions conducive to wildfire occurrence, but it does not explicitly account for atmospheric instability, which can play an important role in the development of large fires (Pinto et al., 2020). In addition, our analysis focuses on the influence of topographic effects. The complex regional topography may strongly modulate atmospheric instability (Santos et al., 2023), which is closely associated with lightning activity and has been shown to be an important factor in fire growth at the local scale (Haines, 1988).

To further disentangle elevational controls on fire activity, we examined individual climatic and ecological variables associated with fuel aridity, ignition potential, and fuel availability. Observed declines in LAI and wind speed with increasing elevation (Fig. 7) are consistent with a reduced probability of fire occurrence at higher altitudes. LAI was selected as a proxy for fuel availability (Yu et al. 2020). Multiple linear regression analyses reveal pronounced elevation-dependent relationships (Fig. 6). At high elevations (>1500 m), fire occurrence shows no significant relationship with LAI. In contrast, fire activity shows clear positive correlations with LAI at both low (200–300 m) and mid-elevations (600–1400 m), highlighting the importance of vegetation density as a proxy for available fuels. While relatively strong regression coefficients are observed in the low-elevation band (200–300 m), the corresponding R2 values are comparatively low, indicating that LAI explains only a limited fraction of the variability in fire occurrence at these elevations. Consequently, the relationship between vegetation and fire activity is more robust at mid-elevations. Although ridge regression produced smaller coefficients than OLS, the coefficient for LAI remained significantly higher than those of the other predictors. These observations support earlier studies showing that vegetation structure and forest composition strongly regulate fire dynamics (Parks et al., 2018; Wang et al., 2022c). In particular, evergreen needleleaf species such as Pinus densiflora and Pinus roxburghii possess resin-rich needles that enhance flammability and promote high-intensity fire behavior (Baek et al., 2022; Kumar and Kumar, 2022).

Air temperature, closely linked to climate change, influences fire activity by modulating fuel drying and convective processes (Dastour et al., 2024; Duane et al., 2021). However, temperature exhibited negligible effects, with both OLS and ridge coefficients remaining close to zero across all elevations under both nighttime and midday conditions. The weak elevational dependence of 2 m air temperature (Fig. 7) suggests that temperature alone cannot explain the observed decline in fire occurrence. At mid-elevations, nighttime temperature exhibits a weak negative partial correlation with fire occurrence, while daytime temperature shows a positive association, highlighting the complex and diurnally dependent influence of temperature under topographic constraints. In contrast, precipitation emerges as a robust suppressor of fire activity across all elevation bands, with the strongest effects observed at mid-elevations. The consistently negative regression coefficients indicate that precipitation deficits are a primary determinant of fire occurrence and spatial distribution, consistent with previous findings linking fire activity to precipitation amount and rainy-day frequency (Kim et al., 2025). Relative humidity likely reinforces this effect by further modulating fuel moisture, particularly in boreal forest systems (Veraverbeke et al., 2017; Dastour et al., 2024).

Wind speed also exhibits strong elevational dependence. Nighttime wind speed shows a positive association with fire occurrence below 1800 m in both OLS and ridge models, with the strongest influence at mid-elevations (Fig. 6). Daytime wind speed is positively related to fire activity at low elevations but shows no significant relationship at mid-elevations, likely reflecting differences between daytime up-valley winds and nighttime downslope flows. In mountainous terrain, downslope flows may advect warm and dry air toward lower elevations, creating fire-favorable conditions when fuels are sufficiently dry (Abatzoglou et al., 2023, 2021). Partial correlation analyses indicate that, when isolated from other factors, wind speed remains positively associated with fire activity across elevations below 2000 m, particularly at mid-elevations. Elevated wind speeds enhance fire activity by increasing evaporative drying, promoting sustained combustion, and facilitating ember transport (Ma et al., 2020; Jones et al., 2022).

Together, these results identify precipitation, wind speed, and vegetation abundance as primary drivers of the elevation-dependent distribution of forest fires. At mid-elevations, fire activity is jointly controlled by sufficient fuel availability and aridity, the latter mediated by reduced precipitation and elevation-specific wind regimes. In addition, enhanced FAOD associated with fire-emitted smoke aerosols – dominated by light-absorbing black carbon (Yu et al., 2019) – may further contribute to atmospheric heating and fuel drying at seasonal to interannual timescales (Pei et al., 2025), potentially reinforcing fire-favorable conditions. Several potentially important drivers, including anthropogenic ignitions, lightning frequency, soil moisture, and regional fire management practices, were not explicitly considered in this analysis. In addition, the limited number of grid cells (<10) between 1300 and 2100 m reduces statistical robustness, suggesting that fire controls at these elevations may be more complex at the global scale. This threshold differs from that used in the FAOD trend analysis because the two analyses are conducted at different spatial resolutions. Specifically, the FAOD trend analysis is based on a 0.1° grid with a large number of global grid cells, for which a threshold of 50 grid cells is applied to ensure statistical representativeness. In contrast, the elevation-dependent regression analysis is performed on a coarser 1° grid, where the number of available grid cells is substantially smaller; therefore, a lower threshold of 10 grid cells is adopted to maintain sufficient statistical robustness. These unaccounted factors likely contribute to the residual variability in fire occurrence and warrant further investigation.