The southern Amazon is one of the regions that has undergone substantial deforestation in the 21st century. The region investigated in this study, 5–25° S, 50–70° W (Fig. ) covers the majority of the “arc of deforestation”, which forms the epicentre of Amazon deforestation . The area includes parts of the southern Amazon, as well savannas and grasslands such as the Cerrado.

All observational data were re-gridded to a 1° × 1° horizontal resolution using linear interpolation to provide spatially consistent datasets for analysis and inter-comparison. The atmospheric composition and burned area data were all analysed on a monthly temporal resolution, while the land cover and LAI data were annual, with the exception of calculating the LAI seasonal cycle. To better understand the impacts of seasonality, we separated the monthly data into wet (February, March, and April) and dry (August, September, and October) seasons. The 3-month intervals were chosen based on seasonal variations in precipitation and compatibility with previous definitions . The dry-season data were further separated into high-burned-area (monthly burned areas ≥ 0.04 % grid cell area) and low-burned-area grid cells. The threshold for defining the high burned area of ≥ 0.04 % was chosen to ensure sufficient sample sizes in each category, while identifying areas with clear fire signals (see the Supplement). There is a 6 km difference in elevation within the region with maximum elevations in the south west (SW). As some of the satellite retrievals are associated with high errors over the high altitudes of the Andes, a small portion of the domain > 1000 m a.s.l. was not included in our analysis.

Whenever data for different domains are compared (e.g. difference in isoprene over regions of low and high burned area at a given forest cover), the mean across all available data satisfying the domain conditions was calculated. First, the relevant datasets were subset in time, so the only years included are those for which land cover, burned area, and composition data are all available. Next, the data were split based on the burned area threshold. Each subset (high and low burned area) was then sorted into land cover bins, and then the mean and standard error for each bin were calculated.

Regression analysis was used to quantify the relationship between the following surface variables: land cover, LAI and burned area, and the atmospheric constituents. A Spearman rank correlation , 1r=1-6∑i=1ndi2nn2-1, where r is the Spearman rank correlation coefficient, n is the sample size, and di represents the difference between the ranks of the ith values from each sample. The ordinary least squared (OLS) and Theil–Sen regression methods were utilised. The OLS regression aims to minimise the sum of the squared differences between the observations and the model (residual sum of squares) when fitting a linear model. The Theil–Sen regression is more robust to outliers than OLS regression, as it fits a slope and intercept based on the spatial median of these parameters calculated on subpopulations of the data. Where a clear relationship (OLS coefficient of determination (R2) ≥ 0.25) was identified between an atmospheric constituent and land cover or burned area, the data were binned by the explanatory variable, and a weighted least-squares (WLS) regression was used to account for the variance of the data in each bin. Throughout the text, we use standard errors to represent uncertainty ranges.

Over the early 21st century (2001–2019), forest cover averaged over the southern Amazon region decreases by 3.8 % from 52.0 % to 48.2 % of the study region (Fig. a). The deforestation is greatest in the north (N) of the domain at around 10° S (see Fig. ). Generally across the region, forest cover reduces substantially from 2001 to 2013, with a smaller decline thereafter. Over the same period, savanna and grassland expand from 46.7 % to 50.5 % (Fig. b) (see the Supplement for the separated savanna and grassland time series). The two land cover categories, broadleaf forest and savanna and/or grassland, display opposite trends through time (Fig. a, b), reflecting that broadleaf forest cover is being replaced by the savanna/grassland modal land cover type.

Mean LAI and total annual burned area averaged over the study region do not have consistent trends over 2001–2019 and 2001–2016, respectively (Fig. c, d). Average annual LAI values in the region fluctuate around a value of 3.1, with a standard deviation of 0.06. As the regional estimate of LAI is dependent on all vegetation types in the domain, any decrease in LAI due to decreases in broadleaf forest (high LAI values) may be of a lesser magnitude than year-to-year variability in the vegetation overall due to, e.g., weather or disease. Burned area also exhibits substantial year-to-year variability, with maximum values observed in 2007 and 2010, while the least burning occurs in 2009. Since 2011, the monthly average burned area has remained below 200 km2 in the peak burning months, and the inter-annual variability has decreased.

The year-to-year variability for annual regional mean CO, NO2, and AOD is similar to that for the burned area (minima in 2009 and maxima in 2007 and 2010; Fig. h, i, j). These atmospheric constituents also show reduced year-to-year variability from 2011 onwards, with the exception of elevated values in 2015 for AOD and CO. The 2007 and 2010 maxima are also observed in the HCHO record, although the year-to-year variability does not reduce for this trace gas (Fig. g). The similarities in the temporal record suggest that the burned area has an influence on the domain-averaged AOD, CO, NO2, and HCHO.

Methanol and isoprene are also characterised by substantial inter-annual variability, and neither shows a clear trend through time for the region average (Fig. e, f). Consequently, other annually varying factors affecting these constituents, such as meteorology or changes in other sources and sinks , may have a greater impact than the decrease in broadleaf forest cover, which is limited in its spatial extent, over this time period. It is noted that the annual mean isoprene and methanol datasets do not extend as far back in time as the land cover dataset, and the change in broadleaf forest cover is smaller during the period covered by these datasets.

The regional average seasonal cycles of the six atmospheric constituents are found to be similar with a peak in the dry season (August to October) for all species (Fig. ). Uniquely, isoprene shows a secondary peak earlier in the calendar year in March/April, which is when the isoprene column densities reach 80 % of the annual maximum (August regional mean of 8.6 ± 0.5 × 1015 molec. cm-2).

Several of the constituents also have co-occurring minima. Isoprene, methanol, and AOD all have the lowest values of 4.7 ± 0.1 ×1015 molec. cm-2, 0.1 ± 0.01 ppbv, and 0.11 ± 0.004, respectively, around June. HCHO, CO, and NO2 remain more stable between December and June.

The amplitudes of the seasonal cycles of the different constituents vary. Isoprene, HCHO, and CO have relatively low intra-annual variation, as their column amounts remain above 50 % of their maxima throughout the year. In comparison, methanol, AOD, and NO2 exhibit a more extreme seasonality, with values dropping below 30 % of their annual maxima. This pronounced dry-season peak in the atmospheric constituents is consistent with the burned area seasonal cycle, which rises from 0.02 ± 0.002 % of the study region in July to 0.06 ± 0.009 % in September, before decreasing to 0.02 ± 0.002 % again in October. Burned area values remain below 10 % of the September peak for the rest of the year.

The LAI seasonality is substantially different, as values remain above 70 % of the maximum throughout the whole year. LAI is elevated at > 3 (> 90 % of the maximum value) between April and October. The regional monthly mean drops slightly (< 2.7 or < 80 %) from December to February. The lower-percentage amplitude of the LAI seasonal cycle may be driven by different phenologies of the vegetation in the domain; e.g. the Cerrado savanna grassland flowering is relatively consistent throughout the year, while Amazon Rainforest vegetation tends to flower in the dry season . Other research in Brazil has found that while seasonal changes in forests are associated with solar radiation, they are driven by rainfall for savannas and grasslands, resulting in opposite cycles for the different land cover types . A comparison of six LAI datasets consistently shows that in the southern Amazon the LAI in regions of broadleaf forest cover is higher in July, compared to January, while the savanna/grassland region has the opposite LAI signal . Consequently, phenology-driven LAI values may peak at different times, depending on the land cover types, resulting in full/partial vegetation coverage of the region throughout the year.

The magnitudes of all atmospheric constituents examined in this study increase in the dry season when the burned area is largest, suggestive of changes to the atmospheric chemistry due to a substantial pyrogenic source. In particular, the seasonal variability in NO2 and AOD closely matches the seasonality in the burned area, while methanol does so with a lag of 1 month. As outlined above, vegetation cover, as represented by the LAI, shows more consistent values throughout the year. Therefore, constituents with lower seasonal amplitudes are potentially more strongly linked to the seasonality of biogenic emission sources, although other factors such as atmospheric lifetime will have an important impact on the respective atmospheric constituent concentrations.

Across the southern Amazon, broadleaf forest, savanna, and grasslands represent the dominant vegetation types. Annual average broadleaf forest cover dominates in the northwest (NW), with land surface coverage typically between 80 % and 100 %, while savannas and grasslands represent the main land cover (80 %–100 % coverage) in the southeast (SE) (Fig. ). In the centre of the domain, a transitional region exists, where the competing phenologies typically have 40 %–60% coverage. The corresponding LAI spatial distribution highlights larger values (3.5 to > 5.0) in the NW and lower values (2.0–3.5) in the SE. This is consistent with broadleaf forest having a larger biomass (per unit area) in comparison to savanna/grassland vegetation types.

For the dry-season burned area data, there are sporadic hotspots peaking at > 0.3 %. The most coherent spatial structures are a filament of burned area values between 0.2 % and 0.3 %, stretching along the Bolivian eastern border down into Paraguay in the south (S) of the region, and the cluster of western Brazilian burned area values at > 0.1 %, peaking at 0.3 %–0.5 %. This latter feature is generally consistent with the reported “arc of deforestation” in the Amazon . The hatching in Fig. d represents at least a 2.5 % decrease in broadleaf forest between 2001 and 2019, which coincides with the burned area patterns reported here. Therefore, the fire activity may be related to regions undergoing deforestation and, to a certain extent, the land cover classifications. For instance, the burned area filament in Fig. d closely follows the high (> 90 %) or low (< 20 %) savanna/grassland and broadleaf forest structure in Fig. a and b. Thus, it is suggestive of transitional regions between biomes driven by predominately anthropogenic pyrogenic activity. Overall, the spatial distribution of burned area exhibits more localised maxima than vegetation cover.

The spatial distributions of multi-annual mean values of total column isoprene are similar to those of broadleaf forest cover and LAI for both the wet and dry season (cf. Figs. a, b and b, c). Values decrease from the far NW (maximum of 2.7×1016 molec. cm-2 in the dry season) to < 0.5×1016 molec. cm-2 in the S, following the NW–SE transition from forest to savanna/grassland as evident in the dry-season total column isoprene values (see Figs. b and b). Over the broadleaf forest region, isoprene is elevated in the west (W), where the maximum forest cover occurs, compared to the east (E). The spatial similarities between broadleaf forest cover and total column isoprene amounts suggest that the broadleaf forest is the dominant source of this BVOC for this region, especially in the dry season. This finding is also consistent with the short lifetime of isoprene confining the peak concentrations to the source region. The W–E gradient in mean total column isoprene over the forested region itself suggests variability in the isoprene emissions within the broadleaf forest biome due to different plant species or local environmental conditions, as found by , who identify an elevation gradient in isoprene emissions in the tropical forest N of the region studied here, probably driven by plant species distributions. In this study, the highest total column isoprene values occur at elevations of around 200 m a.s.l., with lower isoprene column densities in the N of the forest, where elevations decrease to below 100 m a.s.l.

Two other atmospheric constituents, namely CO and AOD, have similar distributions in the wet season to that of broadleaf forest cover. CO is elevated in the NW (> 16×1017 molec. cm-2; maximum of 18.2×1017 molec. cm-2), compared to the SE (< 14×1017 molec. cm-2). AOD is similarly elevated in the N (> 0.2, maximum of 0.35) and reaches a minimum in the SE of the study region (< 0.1) (Fig. a, c). The spatial patterns of CO and AOD during the wet season are consistent with the broadleaf forest acting as a source of biogenic precursors of CO and aerosols as values increase over more densely forested areas (compare with Fig. b). This is consistent with biogenic sources having a greater relative impact on CO and aerosols in the wet season, which is when pyrogenic emissions are minimal .

In contrast, wet-season HCHO and NO2 increase from the NW (HCHO and NO2 column values of ∼ 0.8×1016 and < 0.6×1015 molec. cm-2, respectively) to the SE, where HCHO reaches a maximum of 3.2×1016 molec. cm-2 and NO2 is 1.36×1015 molec. cm-2. This spatial pattern is not consistent with biogenic emissions from the forest being the dominant controlling factor (Figs. e and e; compare to Fig. b). Furthermore, biomass burning is minimal in the wet season (see Sect. ), suggesting that another source or sink is driving total column HCHO and tropospheric column NO2 in this season. The wet-season methanol concentrations are also different to that of vegetation cover. The highest methanol concentrations of ≥ 0.4 ppbv are recorded in the far E of the domain between 10 and 16° S (Fig. c).

The dry-season spatial patterns of HCHO, CO, and AOD differ somewhat from their wet-season distributions, suggesting a change in dominant sources/sinks e.g. from vegetation to fires, transport, and/or atmospheric chemistry processes. CO appears more well-mixed compared to the other trace gases, as expected, with its relatively longer lifetime compared to the other species (average CO tropospheric lifetime of 1–3 months; ; compare with Table 1) (Fig. b). HCHO, CO, and AOD reach maximum values (up to 3.7×1016 molec. cm-2, 32.7×1017 molec. cm-2, and 0.88, respectively) over the transition zone between forests and other land cover classes, although the exact locations of these peaks vary. This region is most strongly affected by deforestation and proximate to the highest average dry-season burned areas (compare with Fig. ). Total column HCHO decreases beyond this zone, consistent with a significant pyrogenic source , both further into the broadleaf forest (1.2 to 1.6 ×1016 molec. cm-2) and to the S, where values drop to < 0.6 ×1016 molec. cm-2 (Fig. f). Minima in the AOD and CO data are also found in the S at the foot of the Andes and in the SE (< 0.2 and < 18 ×1017 molec. cm-2, respectively) (Fig. d). Peak total column CO is found slightly further S to the HCHO maximum (10–12° S compared to 8–11° S). AOD is characterised by two maxima (> 0.8), within the same region as CO, which occur in areas of mixed land cover (forest cover 10 %–100 %) that have experienced substantial deforestation (> 2.5 % area deforested) and some burning (dry-season monthly mean burned area ∼ 0.01 %) (compare Figs. d and ). However, this does not correspond to the locations of maximum dry-season burned areas. The co-location of these maximum values with regions of deforestation and biomass burning suggests that a pyrogenic source, potentially emissions from deforestation fires, is important for these constituents. These constituents were also elevated to the N of this region, where the forests are dominant, suggesting the presence of a biogenic source for this region or the transport of pyrogenic emissions. Particularly in the case of HCHO, the presence of NOx associated with biomass burning could affect the HCHO yield from BVOC emissions , potentially increasing HCHO formation from isoprene oxidation over the broadleaf forest.

Methanol dry-season concentration values are similarly elevated (> 0.6 ppbv) in the N over the forest to savanna–grassland transition, but they are also consistently high further into the Amazon Rainforest, as well as in the NE. Some of the maxima are located over areas of mixed vegetation, which experienced substantial deforestation over 2001–2019, as found for HCHO, CO, and AOD (compare Fig. d with Figs.  and ), but other regions do not reflect any of the studied land surface variables.

The dry-season NO2 column densities had several distinct maxima > 2.7×1015 molec. cm-2, with the highest being 3.5 ×1015 molec. cm-2, which are all located 9–16° S. Some of these overlap with the two AOD maxima. Unlike AOD, NO2 is clearly elevated over the maximum burned areas recorded at 15° S, 65° W and in the E of the study region (Figs. d, f and d). In regions of minimal fire activity, such as deeper into the Amazon Rainforest and the edge of the Andes, tropospheric column NO2 is < 0.9×1015 molec. cm-2. The spatial pattern of the tropospheric column NO2 closely resembles that of burned area, highlighting the close relationship between the trace gas and fire activity, owing to its much shorter lifetime (around 1 d for NOx at the surface; ) compared to CO and aerosols.

In this section, statistical relationships between land cover, LAI, and burned area and the atmospheric constituents for the dry season are explored to quantify the extent to which land cover variables drive atmospheric composition over the southern Amazon for the time period during which both datasets are available. The Spearman rank correlation (Table ) and OLS regression (Table ; see Sect. ) are utilised. The results from both methods are very similar, with the squared Spearman rank correlation coefficients slightly higher than the equivalent OLS R2 values.

There is a strong significant positive relationship between the two vegetation variables of broadleaf forest cover and LAI and total column isoprene over the region (Spearman's r = 0.77; OLS R2 = 0.59 and 0.54 for broadleaf forest cover and LAI, respectively), as expected based, on Sect. –. However, there is a weak negative relationship between isoprene and burned area. In contrast, tropospheric column densities of NO2 exhibit a moderate positive relationship with burned area (Spearman's r = 0.46; OLS R2 = 0.25) but a weak relationships with land cover variables over the region.

The relationships for the other atmospheric constituents are mixed and much weaker. HCHO, CO, and AOD all show positive weak to moderate relationships with both broadleaf forest cover and LAI (r values around 0.4; OLS R2 = 0.08 to 0.18), suggesting some influence of vegetation on these atmospheric constituents, particularly for HCHO. The relationships between these atmospheric constituents and the burned area are considerably weaker but also positive (r values from 0.17 for CO to 0.23 for HCHO; OLS R2 between 0.01 and 0.03). Methanol showed a similarly weak positive relationship with both land cover and fire (r values from 0.15 with burned area to 0.22 with LAI; OLS R2 between 0.01 and 0.04). These weak results, compared to the isoprene relationship with vegetation, highlight that with the longer lifetimes of these trace gases and aerosols, especially CO and methanol (average lifetimes of 1–3 months and 5 d, respectively; ), there is more transport and mixing.

Multiple OLS regression was also performed to test whether the combination of biogenic and pyrogenic sources improved their explanatory power for the variation in atmospheric constituents. The R2 value is unchanged for isoprene, while for NO2, the R2 values are minimally affected, suggesting one main emission source of vegetation and fire, respectively, for these two atmospheric species. For the other atmospheric constituents, the combination of broadleaf forest cover or LAI and burned area increased the R2 values (compared to the OLS linear regression results) modestly. The change is most notable for HCHO (R2 value of 0.23 compared to 0.18) and then for AOD (R2 value of 0.12 compared to 0.09). While the overall R2 values are still much lower than the R2 value for the linear relationship between isoprene and broadleaf forest cover, for HCHO the multiple regression R2 values are only slightly lower than for the NO2–burned area relationship. This illustrates the relevance of both a biogenic and pyrogenic source to HCHO column densities, as previously observed elsewhere in Brazil .

The more robust relationships between broadleaf forest cover vs. isoprene and burned area vs. NO2 were studied in more detail (Figs. a and ). The dry-season composition data were binned based on broadleaf forest cover (for isoprene) or burned area (for NO2). We used a bootstrapping approach to test the significance of the isoprene (NO2) increase with broadleaf forest cover (burned area) and found that the results are significant at the 95 % confidence level (not shown). Over the region, both binned (at 10 % broadleaf forest cover intervals) and non-binned data suggest that isoprene increases by 1.1 ×1015 molec. cm-2 for every 10 % increase in broadleaf forest cover in the dry season (OLS R2 = 0.59; WLS R2 = 0.97) (Fig. ). Consequently, the values for total isoprene column densities over completely (100 %) forested regions are on average 4 times greater than over non-forested regions (0 %). In non-forested regions, isoprene concentrations reflect the local background arising from emissions from non-forest species, as well as mixing and transport of forest-related emissions on short timescales. Tree species composition, in addition to forest dynamics and environmental conditions affecting the plant's emission efficiency, could explain the variability in the isoprene column amounts within each forest cover bin.

The linear burned area–NO2 relationship was found to exhibit different sensitivities, depending on vegetation cover (Fig. ). Over regions of high forest cover and high LAI values, tropospheric NO2 increases with burned area more than over locations with low forest cover and lower LAI (i.e. those dominated by savannas and grasslands; see Sect. ). These different sensitivities are particularly clear up to around 0.5 % burned area (Fig. ; see Fig. f for average conditions for burned areas from 0 % to > 0.1 %). Extremely high burned areas tend to occur in less forested regions, though the spatial distributions (Fig. ) suggest that these are savanna/grassland fires proximate to the broadleaf forest in the central part of the region. The dominance of savanna/grassland at extremely high burned areas could decrease the variation in NO2 explainable by the forest cover over the whole dataset (Tables  and ).

Hence, to represent this more complex burned area–NO2 relationship, we also explored the log–log relationship for NO2 (log⁡e(NO2)) and burned area (log⁡e(BA)) (R2 = 0.32 for non binned data; WLS R2 = 0.90 for binned data (at 0.25log⁡e(BA) intervals); Fig. ). This log–log relationship, approximately a fifth root power law, indicates that the change in tropospheric column NO2 is slightly greater at low burned area percentages compared to at high burned areas. This is consistent with the findings for the different vegetation types outlined above. The relationship was found to remain relatively consistent through time by testing the relationship between the natural logarithm of burned area from GFED5 and natural logarithm of the NO2 tropospheric column for the later time period 2017–2020 (not shown). For both NO2 and isoprene, the WLS linear regression results were largely unaltered when the bin sizes were halved or doubled (not shown).