Introduction

Global climate change is one of the greatest challenges of our time, deeply affecting natural systems, societies, and economies1,2,3,4. A key driver of this phenomenon is the increasing atmospheric concentration of greenhouse gases (GHGs), particularly carbon dioxide (CO2), primarily fueled by the combustion of fossil fuels and, in tropical countries such as Brazil, by land-use and land-cover changes, especially deforestation5,6,7,8.

In this context, orbital sensors have emerged as strategic tools for monitoring the atmospheric distribution of CO2 and understanding the global carbon cycle, especially in remote and hard-to-access areas9,10,11,12. Satellites such as GOSAT (Greenhouse Gases Observing Satellite), developed by JAXA, and NASA’s OCO-2 and OCO-3 (Orbiting Carbon Observatory) were specifically designed to quantify the column-averaged dry-air mole fraction of CO2 (XCO2), allowing for the investigation of its spatial and temporal variability12,13,14.

Most XCO2-related studies have focused on the Amazon due to its relevance for global carbon removals and its role in climate regulation15,16. However, other Brazilian regions, such as the semi-arid Northeast, remain underrepresented in such analyses. The Caatinga biome, exclusively Brazilian and covering approximately 10% of the national territory, presents a unique combination of ecological, climatic, and socio-environmental characteristics, making it a strategic region for studying the interactions among carbon, climate, and human activity11,12. Nevertheless, the behavior of XCO2 in the Caatinga remains poorly understood.

Investigating the spatial and temporal variability and anomalies of XCO2 in the Caatinga can provide valuable insights into the regional carbon balance and the biome’s response to natural and anthropogenic disturbances. Decreases in XCO2 levels typically indicate carbon uptake by vegetation, while increases may signal emissions associated with processes such as environmental degradation or climatic variability13,17,18,19. Recent studies have shown that negative XCO2 anomalies are linked to ecological variables such as gross primary production (GPP) and solar-induced chlorophyll fluorescence (SIF), underscoring the potential of these data for biophysical inference14,20.

Moreover, spatial variations in XCO2 can reflect emissions derived from fossil fuel combustion, land use, and urban expansion21,22,23,24, thus serving as valuable indicators for identifying strategic mitigation efforts25,26. These patterns have encouraged the development of inverse modeling approaches to estimate emissions based on atmospheric observations13,26,27.

Despite the advantages of satellite-derived XCO2 observations for global and regional CO2 monitoring, retrieval accuracy can be adversely affected by surface and atmospheric conditions. Reports indicate that uncertainties may increase in regions with high surface reflectance, sparse vegetation, intense aerosol loads, and heterogeneous land cover characteristics commonly found in drylands such as the Caatinga. These factors can lead to spatial and temporal biases in XCO2 estimates, highlighting the need for regional assessments and careful interpretation of satellite-derived data in semi-arid environments28,29.

Additionally, radiometric performance degradation has been documented in long-term satellite-based products used to infer biophysical responses to atmospheric CO2, such as those derived from MODIS. Although MODIS vegetation indices remain some of the most widely used Earth observation products due to their global coverage and temporal consistency, studies have revealed sensor aging effects and calibration drifts that may influence data stability over extended time periods (e.g., Lyapustin et al., 201430. Therefore, integrating XCO2 observations with complementary variables should incorporate considerations of long-term sensor uncertainties to ensure robust environmental assessments.

In light of this context, the present study aims to investigate XCO2 anomalies in the Caatinga biome using data from the OCO-2 satellite, considering its distinct phytoecological zones, here referred to as biozones: Contact areas (Ecotone-Enclave), Deciduous Seasonal Forest, Semideciduous Seasonal Forest, Savanna, Steppe-Savanna, Open Ombrophilous Forest, Pioneer Formation, and Dense Ombrophilous Forest. By mapping and analyzing XCO2 patterns across these biozones, we aim to address important gaps in the understanding of the carbon cycle in Brazil’s semi-arid regions and contribute to the broader knowledge of CO2 emission and sequestration mechanisms under extreme environmental conditions and increasing anthropogenic pressure.

Method

Study area

The biome selected for this study is the Caatinga (Fig. 1), located in the northeastern region of Brazil. It covers approximately 850,000 km2 and spans across nine Brazilian states, accounting for about 10% of the national territory31. The Caatinga is often described as a dry tropical forest dominated by shrubs and herbaceous plants, with a significant number of endemic species32. The prevailing climate is semi-arid, characterized by a prolonged dry season that can last from 5 to 11 months and an average annual precipitation of about 1000 mm33. Temperatures in the region are generally high throughout the year, with annual averages ranging from 24 °C to 28 °C. During the hottest months, maximum temperatures can exceed 40 °C, while the thermal amplitude between day and night is also significant due to low humidity levels33.

Fig. 1
Fig. 1
Full size image

Location of the Caatinga biome and its phytogeographic biozones. Map showing the spatial extent of the Caatinga biome in northeastern Brazil, highlighting the different phytogeographic subzones (phytoecological regions) that characterize its internal heterogeneity. Boundaries are based on official biome delimitations.

The soils of the Caatinga form a complex mosaic that supports the development of diverse vegetation adapted to these conditions. In general, the soil classes found in this biome are shallow, rocky, and of low fertility34. Approximately 70% of the Caatinga terrain is of crystalline origin (hard bedrock that hinders water storage), while the remaining 30% is associated with sedimentary terrains that are more favorable to water retention35. As a result, due to the variability of soils and topography, the Caatinga exhibits a heterogeneous landscape.

Orbiting carbon observatory

The Orbiting Carbon Observatory-2 (OCO-2) was developed by NASA to remotely monitor atmospheric CO2 on a global scale13. OCO-2 measures absorption in the oxygen A-band near the 0.76 μm wavelength and in the strong and weak CO2 bands at 1.6 μm and 2.6 μm, respectively11,36. These parameters are essential for the proper functioning of the ACOS (Atmospheric CO2 Observations from Space) retrieval algorithm, which is used to estimate the column-averaged atmospheric CO2 concentration (XCO2)36,37,38.

Additionally, the sensors onboard the OCO-2 mission provide a spatial resolution of less than 4 km2, with each measurement frame containing up to eight individual sampling footprints39,40. In this study, we used Level 2, version 10 bias-corrected Lite Files Full Physics data for OCO-241, considering all observation modes: Nadir, Glint, and Target. We included observations that passed cloud screening (quality_flag = 0), as well as those of lower quality (quality_flag = 1) collected between 2015 and 2023 across the entire Brazilian territory. These years were selected because the mission was already operational and providing year-round observations up to the time this study was conducted (January 2025). The OCO-2 Level-2 soundings used in this study represent individual retrievals with an effective revisit frequency of approximately 16 days. To support the spatial-temporal analysis performed here, all valid L2 soundings were aggregated into monthly means for each Caatinga biozone, providing a consistent temporal resolution suitable for detecting seasonal and interannual anomalies in XCO2.

Anomaly model

The XCO2 anomaly model was calculated as described by Hakkarainen et al.13, where positive anomaly values are interpreted as carbon sources and negative values (below zero) as carbon sinks. These two characteristics are associated with anthropogenic activities, such as fossil fuel combustion, as well as biospheric processes, including forest growth and photosynthetic activity.

Accordingly, XCO2 anomalies are computed as the difference between the observed XCO2 value at a given grid cell and the daily median derived from the distribution of all valid OCO-2 observations acquired globally on the same day. This daily median represents a background XCO2 level and is used to highlight relative spatial anomalies, as expressed in Eq. (1):

$$\:{XCO}_{2(i,j)}^{anom}={XCO}_{2(i,j)}-Me\left({XCO}_{2\left(j\right)}\right)$$
(1)

Where \(\:{XCO}_{2(i,j)}\), is the i-th observation of XCO2 on day j, and \(\:Me\left({XCO}_{2\left(j\right)}\right)\) median \(\:\left({XCO}_{2\left(j\right)}\right)\) value for day j.

This model removes the background component of CO2 concentration, reducing biases caused by spatial variations and regional patterns in the data. This approach has been widely applied to investigate emission sources associated with biomass burning and fossil fuel use17,25, as well as to identify natural signals14,19. To explore these natural signals and their relationship with ecosystem functioning, we processed time series of biophysical and climatic variables including the Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), Land Surface Temperature (LST), and cumulative precipitation using the Google Earth Engine (GEE) platform. GEE enables efficient and scalable access, manipulation, and analysis of large volumes of geospatial data, providing an integrated environment for evaluating vegetation dynamics and climate-carbon interactions across broad spatial and temporal scales42.

To ensure consistency before computing correlations, all datasets with different native resolutions were spatially and temporally harmonized. XCO2 measurements from OCO-2, which present irregular spatial sampling, were aggregated into monthly means for each biozone. MODIS-derived vegetation indices (250 m, 16-day) and LST (1 km, daily) were composited into monthly averages, matching the aggregation period used for XCO2 and precipitation. Precipitation data were also summarized monthly and spatially aggregated to the same biozone extent. These harmonization procedures minimize scale-related biases, although we acknowledge that uncertainties may remain due to differences in the original resolutions of each product.

Vegetation indices (NDVI and EVI)

The NDVI (Eq. 2) and EVI (Eq. 3) indices were obtained from the MODIS MOD13Q1 product, Collection 6.1 (MODIS/061/MOD13Q1), with a spatial resolution of 250 m and a temporal resolution of 16 days. Time series data were collected from January 1, 2015, to December 31, 2022. Both NDVI and EVI were considered in this study because they provide complementary information on vegetation dynamics in dryland ecosystems such as the Caatinga. NDVI is widely used as an indicator of vegetation greenness, but it can saturate under moderate-to-high canopy density and be strongly influenced by soil background reflectance, which is common in semi-arid landscapes. EVI, in turn, incorporates corrections for soil and atmospheric effects, offering improved sensitivity in sparse vegetation conditions. Thus, using both indices allows a more robust assessment of vegetation responses to climatic variability in the region. NDVI is used as an indicator of photosynthetic activity and vegetation density, whereas EVI provides greater sensitivity in areas with dense canopy or high biomass, reducing the saturation effects commonly observed in NDVI43.

$$\:NDVI\:=\:\frac{NIR\:+\:RED}{NIR-RED}$$
(2)

Where: NIR = near-infrared reflectance (MODIS band 2); RED = red reflectance (MODIS band 1)

$$\:EVI=G\cdot\:\:\frac{NIR+{C}_{1}\cdot\:RED}{{C}_{2}\cdot\:BLUE+L.NIR-RED}$$
(3)

using the coefficients recommended by Huete et al.44: G = 2.5 (gain factor); C1 = 6.0 (red coefficient); C2 = 7.5 (blue coefficient); L = 1.0 (canopy background adjustment factor); BLUE = blue reflectance (MODIS band 3).

Land surface temperature (LST) and precipitation

Land surface temperature data were obtained from the MODIS MOD11A2 product, Collection 6.1 (MODIS/061/MOD11A2), which provides average LST values at a spatial resolution of 1 km and a temporal resolution of 8 days. This variable is useful for analyzing thermal patterns associated with ecological processes such as water stress and fire risk45.

$$\:LST=\:{T}_{4}+a\left({T}_{4}-{T}_{5}\right)+b{\left({T}_{4}-{T}_{5}\right)}^{2}+c\left(1-\epsilon\:\right)+d\left(\varDelta\:\epsilon\:\right)$$
(4)

Where: T4 and T5: Brightness temperatures measured in thermal infrared bands 4 and 5 (typically centered at 11 μm and 12 μm, respectively); a, b, c, d are empirical coefficients determined through calibration based on atmospheric and surface data; ε: Mean surface emissivity.

\(\:\varDelta\:\epsilon\:={\epsilon\:}_{4}-{\epsilon\:}_{5}\) Difference in emissivity between the two thermal bands.

Monthly precipitation data were obtained from the TerraClimate database, available in the IDAHO_EPSCOR/TERRACLIMATE collection, with a spatial resolution of ~ 4 km. The time series covers the period from 2015 to 2022, encompassing both seasonal and interannual events. This database combines observational data with climate reanalyses and is widely used in studies of environmental dynamics and climate change at regional scales46.

Statistical analyses

The statistical analyses aimed to identify patterns, trends, and relationships between atmospheric CO2 concentration (XCO2) and biophysical and climatic variables across the Brazilian territory from 2015 to 2023.

Initially, the Local Moran’s I index (LISA) was applied to detect significant spatial clusters of XCO2 anomalies within the Caatinga biome, considering its biozones. This allowed the identification of hotspots (areas with high anomalous CO2 concentrations) and coldspots (areas with low concentrations), enabling a regionalized analysis of carbon sources and sinks and their relation to land use and cover.

Subsequently, Pearson correlation analyses were conducted between XCO2 anomalies and environmental variables such as precipitation, LST, NDVI, and EVI. To address the limitations of Pearson correlations in the presence of multicollinearity, partial correlation analyses were further performed to isolate the independent relationship between XCO2 and each predictor. Additionally, multicollinearity was diagnosed using the Variance Inflation Factor (VIF), which indicated severe redundancy between NDVI and EVI (VIF > 36), while the remaining variables remained within acceptable thresholds. These procedures ensured more robust inference regarding the dominant environmental predictors of XCO2 variability.

Additionally, Pearson correlation analyses were conducted between XCO2 anomalies and environmental variables such as precipitation, LST, NDVI, and EVI. Statistically significant correlations (t-test, p < 0.05) allowed the inference of possible associations between carbon fluxes and biophysical factors such as photosynthetic activity, water stress, and rainfall seasonality.

To assess the normality of residuals and homogeneity of variances (homoscedasticity) in the statistical models, the Shapiro–Wilk test and the Levene test were applied at a significance level of 1% (p > 0.01). Testing for normality and homoscedasticity was essential to validate the assumptions of regression analyses and subsequent mean comparisons. To evaluate statistically significant differences between distinct groups (e.g., regions, seasonal periods, or climatic regimes), an analysis of variance (ANOVA) based on the F-test was performed at a 5% significance level.

Furthermore, to evaluate trends and seasonal patterns throughout the time series of XCO2 anomalies, the ARIMA (AutoRegressive Integrated Moving Average) model was employed. This model enabled the detection of cyclical behaviors and temporal persistence in the data, assisting in identifying structural or continuous changes in atmospheric CO2 dynamics. Prior to modeling, all-time series were subjected to the Augmented Dickey–Fuller (ADF) test to assess stationarity. When necessary, differencing (integration) or the inclusion of seasonal terms was applied to the models, based on the Akaike Information Criterion (AIC) and residual diagnostics. All statistical analyses and modeling were performed using R software (R Core Team, 2025)47.

In the ARIMA modeling step, three model structures were evaluated for each bioclimatic zone: ARIMA (0,1,0), ARIMA (1,1,0), and ARIMA (1,1,2). Model selection was based on the lowest Akaike Information Criterion (AIC) and diagnostic checking of residuals (Shapiro–Wilk and Ljung–Box tests). The prediction intervals displayed in Fig. 5 (colored shaded areas) correspond to the forecast horizon generated by the best-fit ARIMA model for each bioclimatic zone, and do not represent observational uncertainty. All forecasts were produced for a 12-month projection beyond the observed period.

Results

Temporal and seasonal variability of XCO2 anomalies

The intra-annual analysis of monthly XCO2 anomalies throughout the time series reveals a consistent cyclic pattern, characteristic of the strong seasonality of the Caatinga biome. In all evaluated years, positive anomaly peaks are observed between January and March, corresponding to the late dry season and the onset of the rainy period, when photosynthetic activity remains reduced and CO2 uptake is not yet fully reestablished. As the rainy season progresses, between April and September, there is a systematic decrease in the anomalies, with values close to zero or negative in most years.

Although the annual cycle is recurrent, the amplitude of this oscillation exhibits interannual variation, indicating differences in ecosystem behavior over time. This is evident in years such as 2015 and 2017, which show more pronounced seasonal peaks, whereas 2019 and 2021 display a more attenuated seasonal variation (Fig. 2a). Additionally, in some years, a delay in the transition to negative anomalies is observed, with positive values extending into April, reflecting shifts in the timing of vegetation recovery among different years.

Fig. 2
Fig. 2
Full size image

(A) Interannual trends of XCO2 anomalies by vegetation type, where the solid line represents the median XCO2 anomaly and the shaded area indicates the variability range of anomalies over time (reflecting the dispersion of positive and negative values). (B) Seasonal cycle of monthly XCO2 anomalies across the study period.

The temporal trends of XCO2 anomalies differed among vegetation types (Fig. 2b). Dense Ombrophilous Forest showed the largest positive trend, with XCO2 anomalies increasing at 0.099 ± 0.020 ppm per year (p < 0.001), indicating a consistent upward trend over the study period. Semideciduous Seasonal Forest (0.035 ± 0.005 ppm yr−1, p < 0.001), Savanna (0.028 ± 0.006 ppm yr−1, p < 0.001), Pioneer Formation (0.013 ± 0.006 ppm yr−1, p = 0.031), and Contact (Ecotone and Enclave) (0.0033 ± 0.0013 ppm yr−1, p = 0.013) also exhibited positive trends, although of smaller magnitude.

In contrast, Open Ombrophilous Forest exhibited a significant negative trend (−0.040 ± 0.013 ppm yr−1, p = 0.002), suggesting a gradual decrease in XCO2 anomalies. Deciduous Seasonal Forest (0.0027 ± 0.0020 ppm yr−1, p = 0.171) and Steppe Savanna (−0.00095 ± 0.00073 ppm yr−1, p = 0.190) did not show statistically significant trends. These results highlight that the temporal evolution of XCO2 anomalies is strongly dependent on vegetation type, with some biomes acting as increasing sources or sinks over the analyzed period.

Spatial and temporal distribution of XCO2 anomalies (2015–2023)

The spatial distribution of XCO2 anomalies across the Caatinga biome between 2015 and 2023 (Fig. 3a) reveals recurrent spatial patterns accompanied by marked interannual variability. Quantitatively, negative XCO2 anomalies consistently dominated the biome, accounting for approximately 55–70% of the total area in most years (Fig. 3b), particularly between 2015 and 2017, when moderate negative anomalies were spatially widespread across central and southern regions.

From 2018 onward, the proportion of area exhibiting positive anomalies increased, reaching up to 30–35% of the biome in 2019 and 2020 (Fig. 3b). These years also showed the emergence of spatially clustered hotspots in the northern and northeastern portions of the Caatinga (Fig. 3a), suggesting enhanced CO2 source behavior potentially linked to land-use change, wildfire activity, or prolonged water stress.

Fig. 3
Fig. 3
Full size image

(a) Spatial distribution of XCO2 anomalies (ppm) across the Caatinga biome between 2015 and 2023, derived from OCO-2 satellite observations. The maps were produced by interpolating monthly XCO2 anomalies and are presented by phytophysiognomic biozones in a faceted layout. Negative anomalies (blue shades) indicate carbon sink behavior, whereas positive anomalies (red shades) represent potential CO2 emission sources, highlighting spatial heterogeneity among vegetation types and the occurrence of localized hotspots linked to climatic and anthropogenic drivers. (b) Interannual variation in the proportional area (%) of the Caatinga biome classified by XCO2 anomaly classes over the same period, summarizing the relative contribution of sink- and source-dominated conditions across years.

In 2021 and 2022, a partial recovery toward sink-dominated conditions was observed, with negative anomalies again representing approximately 60% of the biome area, particularly in the southern sector (Fig. 3b). Nevertheless, a persistent fraction of 10–20% of the area continued to exhibit positive anomalies, highlighting the heterogeneous response of the Caatinga to climatic and environmental pressures.

Spatial patterns of local XCO2 autocorrelation across Caatinga biozones

The spatial analysis of XCO2 anomalies based on the local Moran’s I index (Fig. 4) revealed distinct autocorrelation patterns within the various phytophysiognomic biozones of the Caatinga. The biozones exhibited significant variability in the distribution of positive spatial clusters (high-high), negative clusters (low-low), and spatial outliers (high-low and low-high), underscoring the complexity of CO2 emission and absorption dynamics within the biome.

Fig. 4
Fig. 4
Full size image

Local spatial autocorrelation (LISA – Local Indicators of Spatial Association) of XCO2 anomalies in the Caatinga biome between 2015 and 2023. Clusters classified as high–high (red) indicate areas with consistently elevated anomalies, while low–low clusters (blue) represent persistent sinks. Outliers are shown as high–low (light red) and low–high (light blue), capturing local contrasts in carbon flux dynamics. The results highlight spatial dependence and heterogeneity among phytophysiognomic biozones, reflecting the combined influence of ecological characteristics and anthropogenic pressures.

The Steppe-Savanna biozone, which covers a broad area in the southeastern Caatinga, presented the largest and most significant high-high clusters, particularly concentrated in the central-eastern portion (Fig. 4). In contrast, zones such as the Seasonal Deciduous Forest and Contact areas exhibited a more mixed distribution, with the simultaneous occurrence of both positive and negative clusters.

Smaller biozones, such as the Pioneer Formation and the Dense Ombrophilous Forest, showed a lower density of significant anomaly points. This may reflect both data resolution limitations and the lower spatial variability in XCO2 dynamics in these regions. Nonetheless, isolated points of both positive and negative autocorrelation were observed, indicating that even in these areas, hotspots and coldspots of CO2 emission and absorption may occur (Fig. 4).

The local Moran’s index scatterplots (Fig. 5) reinforce these observations. All biozones displayed a general tendency toward positive spatial correlations between XCO2 values and their neighboring values, indicating significant spatial dependence. The regression slope varied among biozones, with a steeper inclination observed in the Steppe-Savanna (slope = 0.58), confirming a higher degree of spatial clustering in that domain. The slope values ranged from 0.09 in Dense Ombrophilous Forest to 0.58 in Steppe-Savanna, indicating variability in the strength of spatial association across vegetation types. The slopes for each vegetation type are shown in the respective panels.

Fig. 5
Fig. 5
Full size image

Moran scatterplots of XCO2 anomalies by Caatinga phytophysiognomic biozones from 2015 to 2023. Each plot shows the relationship between standardized XCO2 anomalies (x-axis) and their spatial lag (y-axis). The regression line indicates the slope corresponding to the local Moran’s I value, reflecting the degree of spatial autocorrelation within each vegetation type. Steeper slopes, as observed in the Steppe Savanna and Contact areas, indicate stronger spatial structure and clustering, while flatter slopes (e.g., Dense Ombrophilous Forest) suggest weaker spatial dependence.

Analysis of XCO2 anomalies by vegetation type with ARIMA forecasts

Figure 6 presents the time series of XCO2 anomalies for the different biozones of the Caatinga from 2015 to 2023, together with forecasts generated using ARIMA models. The temporal dynamics of the anomalies differ considerably across the biozones, reflecting distinct ecological regimes. These variations suggest that each biozone exhibits a unique pattern of atmospheric CO2 fluctuation, potentially associated with differences in emission and removal processes over time.

Fig. 6
Fig. 6
Full size image

Temporal dynamics of XCO2 anomalies in Caatinga phytophysiognomic biozones from 2015 to 2023 with ARIMA model forecasts. Solid lines represent observed anomalies, while shaded areas denote fitted ARIMA predictions and confidence intervals. Distinct model structures (e.g., ARIMA (0,0,0), ARIMA (1,0,0), ARIMA (1,1,1)) highlight the contrasting carbon dynamics among vegetation types, with stationary regimes in savanna formations, seasonal patterns in pioneer vegetation, and upward trends in ombrophilous forests.

The time series for the Contact areas biozone (ecotone and enclave) was modeled with an ARIMA (0,0,0) with a non-zero mean, indicating no trend or significant autoregressive structure, but persistence of positive anomalies, especially between 2021 and 2022. The forecasts suggest the continuation of this pattern in the short term. A similar pattern was observed in the Deciduous Seasonal Forest biozone, whose series showed moderate variations and episodes of increasing anomalies in recent years. The ARIMA (0,0,0) model with a non-zero mean adjusted to this biozone also indicated a statistically stationary regime around a distinct non-zero mean, reflecting stability, though with signs of recent increase (Fig. 6).

In the Pioneer Formation biozone, a well-defined seasonal behavior was observed, with regular peaks that justified the application of an ARIMA (0,0,0) (1,0,0)12 model, incorporating an annual seasonal component (Fig. 6). This pattern suggests that variations in XCO2 anomalies in this biozone are strongly associated with seasonal cycles, possibly related to vegetation phenology or the alternation between growth and senescence periods. In the case of the Open Ombrophilous Forest biozone, the series exhibited significant oscillations and an upward trend in recent years. The most appropriate model was ARIMA (0,0,1) with a zero mean, which captured short-term dependencies and low-order fluctuations, indicating a possible increase in anomalies. The Dense Ombrophilous Forest biozone, in turn, showed a growing trend between 2019 and 2022, though with smoother fluctuations. The best fit for this biozone was the ARIMA (1,1,1) model with a non-zero mean, evidencing that the series responds to past shocks with short-term memory, but with a consistent trend of increasing anomalies.

The Dense Ombrophilous Forest class exhibited behavior marked by sharp peaks and abrupt variations over time. This pattern was best fitted by the ARIMA (1,1,1) model, which combines differencing, autoregressive, and moving average components being the most complex among the applied models (Fig. 6). This adjustment indicates that the series has a long memory and requires transformations to achieve stationarity, possibly reflecting intense seasonal disturbances such as caused by fires or recent deforestation. For the Savanna biozone, stationary behavior was observed around a slightly negative mean until 2021, followed by stability in subsequent years. The ARIMA (0,0,0) model with a non-zero mean was considered sufficient to represent the series, indicating no significant trends or dependence structures. Finally, the Steppe-Savanna biozone stood out for presenting a marked seasonal pattern, with recurring annual peaks. In this case, an ARIMA (1,0,0) model with a non-zero mean was adjusted, whose projection indicates the persistence of seasonal cycles and a slight upward trend in anomalies in future periods.

Comparison of environmental covariates among vegetation types

For EVI (Fig. 7a), a general pattern similar to that of NDVI was observed, but with more pronounced differentiation between groups. The highest EVI means were found in the Dense Ombrophilous Forest and Ombrophilous Forest, both statistically superior to the other biozones. On the other hand, Savanna, and Steppe-Savanna had the lowest EVI values, clustering into significantly distinct categories from the forest formations. Thus, EVI due to its higher sensitivity to dense vegetation and lower saturation in areas with high canopy cover more clearly highlighted the structural contrasts among the different biozones.

Fig. 7
Fig. 7
Full size image

Climatic variability across Caatinga phytophysiognomic biozones. Results of ANOVA and Tukey’s post hoc test (p < 0.05) applied to annual mean temperature (°C) and annual accumulated precipitation (mm) for the period 2015–2023. Boxplots show statistically distinct groups among vegetation types, with higher precipitation in ombrophilous forests and pioneer formations, and higher mean temperatures in savanna and steppe regions.

For NDVI (Fig. 7b), higher values were observed in the Dense Ombrophilous Forest, and Open Ombrophilous Forest biozones. These biozones differed significantly from the more open formations such as Steppe-Savanna and Pioneer Formation which showed the lowest average NDVI values. These results reflect a greater degree of vegetation cover and biomass density in the forest formations compared to the Savanna and Pioneer formations, which are more sparse, seasonal, or subject to disturbance.

Although these vegetation structure patterns are well established in the literature, their inclusion here is crucial because they provide the ecological basis supporting our atmospheric findings: biozones with denser canopy cover and higher carbon assimilation capacity are the same areas where we observe stronger negative XCO2 anomalies. Thus, NDVI and EVI are key to interpreting how vegetation functional differences modulate carbon dynamics across the Caatinga.

Therefore, the coherence between the two indices reinforces the robustness of the analyses and underscores the relevance of these indicators in monitoring the vegetation structural conditions, especially in the context of carbon dynamics and XCO2 anomalies studies.

Variation in temperature and precipitation among vegetation types

Figure 7c and d presents the results of the analysis of variance (ANOVA), followed by Tukey’s multiple comparison test (p < 0.05), applied to the environmental variables of mean precipitation (mm) and temperature (°C), considering the different biozones. Based on the results, it was found that biozones such as Open Ombrophilous Forest, and Pioneer Formation exhibited the highest average precipitation values, differing statistically from the drier biozones, such as Steppe-Savanna, which formed a group with the lowest precipitation values.

Regarding mean temperature, significant differences were also observed among the biozones. The highest temperature values were recorded in the Steppe-Savanna, Savanna, Pioneer Formation, and Contact areas (ecotone and enclave), which were statistically distinct from the Dense Ombrophilous Forest and Open Ombrophilous Forest biozones. These latter zones exhibited the lowest average temperatures, highlighting the influence of vegetation structure and canopy density on local microclimatic conditions.

Independent effects of vegetation and climate drivers on XCO2 anomalies

Partial correlation analysis revealed that vegetation activity is the primary independent driver of XCO2 anomalies (Table 1). NDVI showed a strong negative association with XCO2 (r = − 0.625, p < 0.05), indicating enhanced biogenic uptake during periods of higher vegetation greenness. EVI also presented a significant correlation (r = 0.593, p < 0.05), but with an inverted signal compared to the Pearson coefficient, reflecting strong redundancy between vegetation indices. Conversely, temperature and precipitation exhibited weak and non-significant partial correlations (r = 0.187 and r = − 0.180, respectively; p > 0.05), suggesting limited independent influence on XCO2 anomalies after accounting for vegetation–climate interactions. Variance Inflation Factor (VIF) values confirmed severe multicollinearity for NDVI and EVI (VIF > 36), while the other predictors remained within acceptable thresholds.

Table 1 Artial correlations and multicollinearity diagnostics for environmental predictors of XCO2 anomalies.
Full size table

Discussion

Intra-annual variability of XCO2 anomalies in the Brazilian Caatinga biome

The seasonal patterns observed in XCO2 anomalies reflect a behavior consistent with the phenology of the Caatinga biome. Enhanced biogenic CO2 assimilation occurs in response to leaf regrowth and increased foliage during the wet season. With the onset of rainfall and higher soil moisture, vegetation resumes growth, expands leaf area, and intensifies photosynthesis, resulting in greater CO2 uptake. This explains the lower (or negative) anomaly values observed in the wetter months48. Such patterns are supported by eddy-covariance and remote sensing studies, which show that gross primary productivity (GPP) peaks during the rainy season alongside increases in vegetation indices (e.g., NDVI, EVI), demonstrating the strong link between foliar phenology and photosynthetic capacity48. Consequently, our results reinforce that the Caatinga functions as a net carbon sink during the wet months, when biogenic assimilation outweighs respiration and soil emissions49.

Conversely, the interannual variability in the amplitude of XCO2 anomalies years with pronounced positive peaks versus years with more attenuated seasonality highlights spatial and temporal heterogeneity in ecosystem responses. This heterogeneity is likely driven by variability in rainfall intensity, timing, and distribution, as well as edaphic conditions, vegetation structure, and land-use history. These observations align with prior studies indicating that carbon flux dynamics in the Caatinga are strongly influenced by rainfall seasonality and local environmental factors49. Our seasonal cycle analysis further reveals clear differences in temporal trends among vegetation types: Dense Ombrophilous Forest exhibited an increase of ~ 0.099 ppm yr−1, whereas Open Ombrophilous Forest showed a negative trend of ~−0.040 ppm yr−1. Such results indicate that even within the same biogeographic domain, vegetation cover responds differently to the drivers of atmospheric CO2, likely reflecting differences in phenology, productivity, leaf area, and carbon sequestration capacity. Denser vegetation communities appear to retain or accumulate CO2 or at least reduce relative emissions more effectively than more open formations, a pattern consistent with evidence that terrestrial carbon sinks in tropical biomes are increasingly spatially heterogeneous and uncertain50.

Overall, these findings support the premise that climatic controls, particularly hydrological regimes and rainfall timing, play a decisive role in shaping the carbon–vegetation cycle in the Caatinga. Phenology, modulated by rainfall, directly drives the XCO2 anomaly signals, emphasizing the importance of capturing intra-annual variability to accurately assess the biome’s carbon balance. Furthermore, our results have important implications for carbon quantification and biogeochemical modeling in dry biomes: relying solely on annual means or raw atmospheric CO2 data without accounting for seasonality and phenology may under- or overestimate the Caatinga’s carbon sequestration potential. Using monthly anomalies, as in this study, provides a robust approach to capture the pulses of carbon assimilation and emission associated with phenology, climate, and hydrological regimes.

Patterns of XCO2 anomalies in the Caatinga (2015–2023)

The predominance of negative and neutral XCO2 anomalies (Fig. 2) across much of the Caatinga biome is consistent with its characterization as an efficient carbon sink. Mendes et al.51 demonstrated this behavior even under extreme drought conditions, reporting an annual mean flux of − 1.86 µmol m−2 s−1 during wet years and − 0.81 µmol m−2 s−1 during severe droughts. This pattern is further supported by findings indicating that the Caatinga ranks among the most effective carbon-sequestering biomes in Brazil, accounting for approximately 45–60% of the CO2 absorbed nationwide52. These results are also in agreement with previous assessments summarized in Table 2, which collectively highlight the biome’s strong tendency toward net carbon uptake despite environmental variability.

Table 2 Summary of the methods employed in the analysis of emissions and environmental variables.
Full size table

The intensification of negative anomalies between 2015 and 2017, especially in the central-southern and southwestern regions, reflects the strong seasonal control of vegetation activity in the Caatinga. At the onset of the rainy season, vegetation rapidly resumes photosynthesis after prolonged drought, leading to an abrupt increase in CO2 uptake and consequently more pronounced negative XCO2 anomalies. This behavior is consistent with eddy covariance studies reporting enhanced gross primary productivity (GPP) and ecosystem respiration (Reco) following the first rains, reinforcing the biome’s temporary role as a carbon sink during this transition period32,57.

The emergence of intense localized positive XCO2 anomalies in 2019 and 2020 supports previous findings linking extreme droughts and fire events to temporary CO2 emission peaks27,34, potentially reversing the Caatinga typical carbon sink behavior. This interpretation is consistent with the observations of Sayedi et al.49, who reported pronounced interannual variability in CO2 fluxes within Caatinga ecosystems, driven largely by reductions in gross primary productivity (GPP) and enhanced soil respiration during drought years. These episodic shifts, also documented for other tropical seasonal biomes, reflect the biome’s sensitivity to hydroclimatic extremes and biomass loss, reinforcing the idea that extreme droughts and fire disturbances can temporarily weaken or even invert the region’s carbon uptake capacity49. Such dynamics highlight the importance of monitoring short-term anomalies to better understand the resilience and vulnerability of the Caatinga carbon cycle under increasing climate variability.

The partial return to negative anomalies observed in 2021–2022, especially in the southern region, indicates ecosystem resilience and the recovery of carbon removal capacity following the stress events of prior years. Although spatially heterogeneous, this recovery is consistent with dry forest systems that tend to resume sink behavior when water availability is restored58. In a global context, the interannual variability observed in the Caatinga resembles that of other tropical biomes, as indicated by OCO-2 satellite data and models such as GONGGA, which report intermittent sink behavior with temporary source pulses during extreme events59.

In summary, the spatial and temporal patterns of XCO2 anomalies in the Caatinga reflect a combination of climatic drivers (precipitation distribution) and extreme events (droughts, fires), in agreement with the literature that highlights the sensitivity of tropical seasonal forests to water and temperature variability51.

Spatial patterns of local XCO2 autocorrelation in Caatinga biozones

The analysis of local spatial autocorrelation patterns of XCO2 anomalies across the phytophysiognomic biozones of the Caatinga biome revealed robust evidence of spatial dependence, with the presence of significant clusters (high–high and low–low) and local outliers (high–low and low–high). These results suggest that CO2 absorption dynamics in the biome are not random or spatially homogeneous but rather structured in response to regional environmental and anthropogenic drivers.

The predominance of positive clusters (high–high) in the Steppe-Savanna biozone, especially in its central–eastern region, indicates areas of potential accumulation or systematic emission of XCO2, likely associated with intense anthropogenic pressures such as agricultural expansion, deforestation, and fire use—factors widely documented for this region60. This pattern may also reflect intrinsic ecological characteristics of the biozone, including lower vegetation cover density and greater soil exposure, which limit carbon sequestration and favor net emissions.

In contrast, biozones with denser vegetation or mixed ecological structures, such as the Seasonal Deciduous Forest and Contact areas, exhibited more heterogeneous spatial patterns. The simultaneous presence of positive and negative clusters in these areas may reflect complex interactions among vegetation, soil, and atmosphere, as well as edge effects and transitional ecological conditions. This heterogeneity is consistent with the ecotonal nature of these zones, where small variations in topography, water availability, or land use can lead to significant differences in carbon fluxes61.

Smaller biozones, such as Pioneer Formations and Dense Ombrophilous Forest, showed a lower density of significant local autocorrelation points. This may be due to limitations in the resolution or coverage of satellite-based XCO2 measurements, especially in areas with dense vegetation, which can hinder the precision of orbital sensors36,62,63. Alternatively, it may indicate lower intrinsic variability in carbon fluxes in these zones, suggesting greater ecological stability or reduced influence from recent disturbances.

The local Moran’s I scatterplots support these findings, showing positive correlations between XCO2 values and their neighbors in all biozones analyzed. The steeper slope observed in the Savanna–Steppe biozone indicates a stronger spatial structure in this domain, reinforcing its vulnerability to changes in the carbon biogeochemical cycle. These results are consistent with previous studies reporting increasing environmental degradation in the region, with direct impacts on greenhouse gas dynamics15,27.

Analysis of XCO2 anomalies by vegetation type with ARIMA forecasts

The temporal analysis of XCO2 anomalies across the different phytophysiognomic biozones of the Caatinga biome revealed distinct patterns in atmospheric carbon behavior, reflecting ecological and functional contrasts among vegetation types. Modeling with ARIMA (AutoRegressive Integrated Moving Average) captured stationary structures, seasonal dynamics, and long-term trends, offering a quantitative interpretation of XCO2 fluctuations from 2015 to 2022.

Biozones characterized by structural stability, such as Contact areas (ecotones/enclaves), Deciduous Seasonal Forest, and Savanna, were adequately modeled with ARIMA (0,0,0) with a non-zero mean. In these cases, the absence of autoregressive and moving-average components indicates stationarity around a constant mean, which nonetheless reflects persistent positive anomalies in recent years potentially linked to sustained anthropogenic pressures or a gradual decline in sequestration capacity15,63.

Biozones with pronounced seasonal modulation displayed recurring peaks likely associated with alternating periods of vegetative growth and senescence. The Pioneer Formation showed a well-defined annual cycle, best captured by a seasonal ARIMA (0,0,0) (1,0,0)12, highlighting the influence of annual climatic forcing. The Steppe- Savanna also presented marked seasonal behavior with recurring annual peaks; its dynamics were satisfactorily summarized by an ARIMA (1,0,0) with a non-zero mean, indicating persistence and a slight upward tendency in anomalies, consistent with a seasonally modulated regime64,65.

In wetter forested biozones, XCO2 anomalies exhibited increasing trends, suggesting recent shifts in carbon dynamics. The Open Ombrophilous Forest was best fitted by ARIMA (0,0,1) with a zero mean, capturing short-term dependencies and low-order fluctuations amidst a recent rise in anomalies. The Dense Ombrophilous Forest required a more complex ARIMA (1,1,1) with a non-zero mean, indicating response to past shocks with short-term memory and a consistent upward trend since ~ 2019, alongside occasional sharp oscillations that may reflect disturbances such as fire or edge effects from land-use change36,60.

Overall, these results demonstrate that the different Caatinga vegetation types respond in distinct ways to climatic and environmental controls most notably rainfall seasonality, soil moisture dynamics, canopy structure, and phenological strategies. Preserved areas and mature forest formations tend to exhibit more stable and predominantly negative XCO2 anomalies, reflecting stronger carbon uptake associated with deeper root systems, higher biomass, and greater functional resilience. In contrast, remnant fragments, open shrublands, and anthropogenically impacted zones show more variable and often positive anomalies, driven by heightened sensitivity to hydrological stress, reduced canopy cover, and the influence of land-use activities. Together, these patterns highlight how the interplay between ecological integrity and climatic forcing shapes the spatial heterogeneity of atmospheric CO2 behavior across the Caatinga biome66. The diversity of fitted models reinforces the need for spatially and ecologically differentiated monitoring of carbon fluxes across the biome. The coexistence of increasing trends and seasonal patterns in several biozones points to heightened vulnerability to climate change and land-use pressures, underscoring the importance of evidence-based mitigation strategies15,67.

Influence of climatic variables in Caatinga biozones: temperature and precipitation

Figure 7 shows the ANOVA results, complemented by Tukey’s multiple comparisons test (p < 0.05), applied to annual mean temperature (°C) and annual accumulated precipitation (mm) across Caatinga biozones. The tests indicate statistically significant climatic differences among vegetation types.

In terms of precipitation, the Open Ombrophilous Forest and Pioneer Formation biozones exhibited the highest mean values, forming a statistically distinct group (p < 0.05) from the Savanna and Steppe-Savanna, which had the lowest precipitation levels. This outcome aligns with expected eco-physiographic traits: biozones with denser vegetation or located in regions of greater elevation and slope tend to receive higher rainfall interception and maintain more humid microclimates65,68.

In contrast, savanna biozones are mostly located in flatter areas with lower water input, reflecting the typical water limitation of the northeastern Brazilian semiarid region30. This variation in precipitation among vegetation types may strongly influence primary productivity, CO2 fluxes, and ecological regeneration dynamics, especially in response to disturbances such as fire and deforestation33.

Regarding annual mean temperature, significant differences were also observed between biozones. The highest temperatures occurred in the Steppe-Savanna, Savanna, and Contact areas (ecotones/enclaves), consistent with their location in more exposed areas, characterized by low vegetation cover and high solar radiation incidence. These conditions tend to favor surface heat accumulation and reduced relative humidity, which may enhance carbon emissions through soil respiration and organic matter degradation15,65. On the other hand, the Dense Ombrophilous Forest and Open Ombrophilous Forest biozones exhibited the lowest mean temperatures, statistically differing from the others. This pattern is associated with the shading effect, higher canopy density, and the presence of deeper and moister soils, which function as local microclimatic buffers62. The lower mean temperature in these zones suggests a greater capacity for thermal moderation and potentially higher resilience to regional climate change.

These climatic distinctions between biozones help explain part of the variability observed in XCO2 anomalies and underscore the importance of considering local climatic contexts when interpreting carbon emission and removal patterns. Moreover, the results provide valuable input for identifying priority areas for mitigation and climate adaptation strategies in the Caatinga biome.

Correlation patterns between XCO2, vegetation, and climate

Partial correlation analyses between XCO2, vegetation indices (NDVI, EVI), and climate variables (temperature and precipitation) revealed biome-wide patterns in the Caatinga (Table 3). When statistically controlling for the shared variance among predictors, XCO2 exhibited only a weak and non-significant association with temperature and vegetation indices (p > 0.05), indicating that short-term phenological or thermal fluctuations do not directly explain atmospheric CO2 variability at the biome scale. In contrast, XCO2 maintained a significant positive relationship with precipitation (p < 0.05), consistent with rainfall-induced shifts in ecosystem carbon exchange via growth stimulation or moisture-driven respiration pulses.

Variance Inflation Factor (VIF) diagnostics revealed extreme multicollinearity between NDVI and EVI (VIF > 36), while climate variables remained within acceptable thresholds. This redundancy among vegetation indices indicates that both metrics capture largely overlapping phenological patterns; however, after accounting for shared variance, NDVI retained the strongest independent relationship with XCO2 anomalies, reinforcing its importance as the dominant vegetation-based predictor in the biome.

Overall, the strong dependence of vegetation indices on hydrological seasonality shown by their positive relationship with precipitation and negative association with temperature highlights the overarching role of water availability in regulating ecosystem functioning across the Caatinga. These statistical patterns suggest that the response of atmospheric XCO2 to biome processes is subtle and likely modulated by regional atmospheric transport and broader-scale carbon dynamics, rather than being driven solely by local vegetation activity.

Table 3 Synthesis of relationships among NDVI, EVI, and environmental variables in the Caatinga.
Full size table

Conclusion

This study revealed that XCO2 anomalies in the Caatinga biome were predominantly positive, exhibiting markedly heterogeneous spatial and temporal patterns modulated by climatic, structural, and phenological factors specific to each phytophysiognomy. The combination of spatial autocorrelation analysis, ARIMA time series modeling, and correlations with environmental variables showed that zones such as the Steppe-Savanna and Pioneer Formation are particularly prone to fluctuations in atmospheric carbon concentrations, with concerning signs of recent increases. The strong dependence of these biozones on hydrological and thermal seasonality, coupled with increasing anthropogenic pressure, suggests that the biome responds acutely to environmental and climatic disturbances, compromising its role as a potential CO2 sink.

The presence of emission hotspots and the instability observed in some forest formations point to the growing vulnerability of this ecosystem. In light of intensifying climate change and the historical neglect of Caatinga conservation, the findings presented here reinforce the urgency of targeted actions for monitoring, protection, and ecological restoration of this biome. Protecting the Caatinga is not only essential for preserving its biodiversity and ecosystem services but also for mitigating carbon emissions and contributing strategically to national and global climate goals.