The study was undertaken in NE Brazil, including the edges of the Amazon and Atlantic Rain Forest regions and parts of the cerrado savanna, but consisting primarily of the Caatinga Region. In the scientific literature, “caatinga” refers to both the region and various types of vegetation. In the Brazilian lexicon, the Atlantic Forest, Amazon Forest, Cerrado and Caatinga were previously referred to as Domains and are now referred to as Biomes (IBGE 1993; Souza Jr and Azevedo 2017), but here we use the term ‘region’ to not confuse with specific vegetation types, EFGs or biomes. Caatinga vegetation is highly variable, but is generally seasonal with physiognomies ranging from open shrublands to tall seasonally dry tropical forests with a closed canopy (Andrade-Lima 1981; Quieroz et al. 2017). A grassy understorey is absent and fire is rare. Transition zones around the core Caatinga are estimated to be 14% larger than the core itself (Cardoso et al. 2021).

NE Brazil has globally high levels of species endemism and floristic compositional heterogeneity (Santos et al. 2012; DRYFLOR 2016; Bueno et al. 2018). In addition, the region is likely to experience temperature increases and increased aridity with climate change (Pörtner et al. 2022). Interaction between human disturbance and climate change is likely to have a strong negative impact on woody plant diversity (Rito et al. 2017). Furthermore, NE Brazil is the world’s most populated semi-arid land, whilst also one of Brazil’s poorest regions (Melo 2017). Consequently, expected effects of climate change will impact many of the country’s most vulnerable citizens. Caatinga biodiversity has a long and complex relationship with humans and is often considered to be highly anthropogenically disturbed, predominantly as a result of small-scale agriculture, but increasingly through irrigated crop production (Silva et al. 2017, chapter 13).

The characteristics of tropical biomes in terms of attributes describing vegetation expression were identified via a literature review of all biomes potentially present within the region of interest. Remote sensing products were identified and obtained that could describe a subset of these attributes (Table 1).

All data were resampled to 1 km resolution using bilinear interpolation, calculated using the ’resample’ function in the R package ‘raster’ (Hijmans 2022). Resampling allowed analysis to be of fine enough resolution to identify transitions between vegetation types, whilst allowing reasonably fast computation. The burn count data was highly right-skewed and so was transformed by taking the natural logarithm (after adding one) to increase the normality of the data.

Exclusion of non-natural vegetation is important to prevent conflation of human-dominated areas with naturally different structural vegetation groups. We thus removed areas with non-natural land covers using the MapBiomas land cover product for 2015 (along with water covered areas; MapBiomas 2022). However, there remains potential for areas with some human modification to be present in the analysis and, for land use change to have occurred over the time period examined (2000–2022) – see Suppl. material 1: table S1 for a full list of land-cover types removed from the study.

Fuzzy spatial clustering was used to investigate the degree to which pixels belonged to multiple groups. This was implemented by using the geocmeans package in R, version 0.3.3 (Gelb 2021, 2023). Fuzzy clustering was used because it allowed the inclusion of an ‘unclassified’ group where the probability of the pixel being in any identified group was low (< 0.45). This approach has been used in similar studies and is often appropriate ecologically due to the continuous nature of vegetation transitions, particularly in dry tropical regions (Torello-Raventos et al. 2013; Feilhauer et al. 2021). In addition, a fuzzy clustering approach aligns with the IUCN GET theoretically, as the GET acknowledges partial membership of locations to multiple EFGs. Spatial clustering allows consideration of the spatial proximity between observations, improving the suitability of clustering for spatial data (Cai et al. 2007; Zhao et al. 2013; Gelb and Apparicio 2021). Clustering was done using the ‘robust’ function in the geocmeans package, which improves the ability of the clustering to identify groups of different sizes, shapes and density. This is done by normalising the distance between the observations and the centre of the groups (Tsai and Lin 2011). We used the silhouette index, explained inertia and Xie-Beni index to assess cluster quality, select the optimum number of clusters (k) and the parameters of the clustering algorithm (Gelb and Apparicio 2021). The algorithm parameters were: m, which controls the degree of fuzziness, β, which controls the speed of convergence and classification crispness, α, which represents the weight of space in the analysis and the moving window size (Gelb and Apparicio 2021). Selecting the most appropriate number of clusters – and other parameter values – is subjective and here we need to balance multiple indices and parameters. We selected the parameter values sequentially, optimising them one by one. The parameters k and m were first selected, based on classifications using the standard fuzzy c-means algorithm. We aimed to: i) maximise the silhouette index, ii) find the k at which increases in explained inertia were marginal and iii) minimise the Xie-Beni index. Then we used the selected values of k and m in the generalised fuzzy c-means algorithm to explore cluster quality metrics for a range of values of β, selecting the value of β, based on the minimised combined rank of the Xie-Beni and Silhouette indices. Finally, we optimised the value of α and the moving window size for spatial generalised fuzzy c-means clustering, again aiming to minimise the combined rank of the Xie-Beni and Silhouette Index.

To understand the determinants of each cluster, hypothesised explanatory variables were selected that have previously been used to explain biome distribution in NE Brazil (e.g. Silva de Miranda et al. (2018); Moonlight et al. (2020); Cardoso et al. (2021)). These were: the climatic variables Mean Annual Temperature (MAT), temperature seasonality (standard deviation of monthly temperature * 100), Mean Annual Precipitation (MAP), precipitation seasonality (coefficient of variation of monthly precipitation), precipitation in the wettest and driest months and precipitation in the warmest quarter (Fick and Hijmans 2017); edaphic factors (soil water capacity, Cation Exchange Capacity (CEC), soil sand content, soil organic carbon content (SOC), soil pH and soil depth, Height Above Nearest Drainage (HAND) (Yamazaki et al. 2019; Poggio et al. 2021), geology, elevation (Yamazaki et al. 2017); and measures of human disturbance (population density, road density and length of time to nearest city (Meijer et al. 2018; Weiss et al. 2018; WorldPop 2018). Climate data were sourced from WorldClim (Fick and Hijmans 2017) and soil data from SoilGrids (Poggio et al. 2021). Although non-natural land cover was removed from the clustering, human-disturbance variables were included in the determinants analysis to account for the fact that intensity of disturbance is experienced as a gradient, rather than a dichotomy. In addition, these measures of human disturbance were included as a check for residual impacts not removed by the removal of non-natural land cover classes. The predictor data were cleaned and processed in the same manner as the vegetation attributes. To check for correlation in the predictors, which can complicate interpretation of random forest outputs, a Pearson’s correlation test was carried out. Where r > 0.8, one of the correlated predictors was removed, based on which variable was considered more mechanistically meaningful. Data sources and justification for their inclusion are found in Suppl. material 1: table S2 and the correlation matrix from the Pearson’s correlation test is found in Suppl. material 1: table S3.

The IUCN GET identifies five types of ecological drivers of biome distribution, three of which are encapsulated in our random forest analysis as variables ecologically relevant to NE Brazil (resource drivers, ambient environment and human activities). Given that only a few drivers shape the properties of ecosystems (Keith et al. 2020b), it is reasonable to select specific drivers for this analysis focused upon dry tropical ecosystems rather than a global analysis. Resource drivers include those related to the availability of water, carbon and nutrients. Ambient environment factors include climate variables and are important for modifying resource availability. Thirdly, human activities are included in the random forest classification. Although disturbance regime is another driver highlighted by the IUCN GET, natural disturbance was not included in the random forest analysis (note that fire was used in the clustering methods as a proxy for the presence of grass, which is difficult to measure using remote sensing techniques).

Random forest classification (Breiman 2001) was carried out using the remaining (non-correlated) explanatory variables with group identity as the response variable. The ‘randomForest’ package in R was used for this analysis (Breiman et al. 2024). We used a random 70:30 train:test split, with model accuracy assessed by calculating the percentage of the pixels in the test group whose vegetation group was correctly predicted by the random forest model. Partial dependence plots were used to illustrate the partial effect of each predictor on the log odds of a pixel being in each group, when all other predictors are held at their mean.

We reviewed the literature to identify the vegetation types that we expected the classification to detect. In order to describe the vegetation groups found by the clustering, these structural descriptions of differences between vegetation groups were cross referenced with additional sources of information: i) previous descriptions of the geographical distribution of the vegetation types compared to our groups (Fig. 1A), ii) the mean and spread of the soil, climate and human disturbance variables used in the random forest analysis (Suppl. material 1: fig. S3, tables S5–S7), iii) quantitative descriptions of the vegetation properties of each group (Fig. 2, Suppl. material 1: table S4) and iv) floristic information (Suppl. material 1: table S8). We used data from Silva de Miranda et al. (2018) to match floristic groups in the region to our groups, based on spatial co-location. Silva de Miranda et al. (2018) used hierarchical clustering to assign 4,103 sites across lowland South America and neighbouring subtropical areas into biomes, based upon tree species composition (Silva de Miranda et al. 2018). The floristic groups corresponded to: Amazon Forest, Atlantic Forest, savanna, seasonally dry tropical forest (SDTF) and chaco. All of these groups are present in our study region, except chaco vegetation. Sites with an assigned floristic group were joined to the structural map using qgis (QGIS Team 2021) and the percentage of each floristic group present in each vegetation group which we identified was calculated.

Figure 1. 

Comparison of the vegetation structural groups identified by c-means clustering of vegetation attributes (A) with the IUCN Global Ecosystem Typology 2.0 Ecosystem Functional Groups in NE Brazil (B). The maps of the EFGs in B is produced by the aggregation of the structural groups in A, as per the lower panel, rather than the GET indicative maps. Scrubby caatinga, heterogeneous caatinga and hyper-seasonal caatinga represent an additional EFG not yet described by the IUCN GET which we term ‘non-pyric shrublands’ as they consist of vegetation that does not burn, but which structurally resembles a shrubland or short-statured forest rather than a tall canopy, closed forest.

Figure 2. 

Attributes of the vegetation in each group in NE Brazil. The scaled mean vegetation attribute for each vegetation group is shown, as a percentage of the maximum of each vegetation attribute. The vegetation attributes were (clockwise from top) aboveground woody biomass, aboveground woody biomass heterogeneity (bio.cv), fire frequency (burn), mean Leaf area index (LAI) in the dry season (LAI.dry), mean LAI in the wet season (LAI.wet), NDVI seasonality (seas) and canopy height (height).

The structural groups identified in the clustering were compared to the descriptions of EFGs in the IUCN GET. We focused on ‘T1.1 Tropical subtropical lowland rain forests’, ‘T1.2 Tropical subtropical dry forests and thickets’, ‘T3.1 Seasonally dry tropical shrublands’ and ‘T4.2 Pyric tussock savannas’. We considered the ecological equivalence between our vegetation groups and the IUCN EFG descriptions, with a particular focus upon the links between vegetation structure and ecosystem processes emphasised in each EFG. Furthermore, we considered the results of IBGE’s analysis of EFGs in Brazil (IBGE 2021). Comparisons were also made to other vegetation and biome maps, including MapBiomas (MapBiomas 2022) and the Terrestrial Ecoregions of the World datasets (Olson et al. 2001); these comparisons can be found in the Suppl. material 1 (Suppl. material 1: figs S4, S5).

The silhouette index, explained inertia and Xie-Beni indices gave varying results for the optimum values of k, the number of clusters (Suppl. material 1: fig. S1). The Xie-Beni index suggested that five or six clusters is the optimum, depending on the value of m. Contrastingly, the silhouette index suggested seven or nine clusters, slightly depending on the value of m. The explained inertia values improved as k increased, although there were diminishing returns beyond k = 7 or 8 clusters, depending on the value of m. As such, it was not possible to fulfil all three of our conditions for selecting the optimum value of k and m. In balancing the different criteria, we considered k = 7 as being the best value and focus our analyses and discussion on seven groups. This set of seven includes multiple caatinga vegetation groups besides the expected divergent vegetation groups (namely, rain forest and cerrado). Based on the silhouette score suggesting seven or nine clusters, we compared the clustering results for k = 7 and k = 9 in Suppl. material 1: appendix S1 and found that the change in k does not alter our main arguments, except that the definition of groups of caatinga vegetation is sensitive to the number of caatinga-type clusters. After selecting k = 7, we selected m = 1.4, β = 0.6, α = 0.7 and a 3 × 3 pixel moving window, following sequential optimisation which minimised the combined ranks of the silhouette and Xie-Beni indices (Suppl. material 1; Fig. 1).

We also conducted the clustering analysis using a hard clustering approach, but found little difference between fuzzy and hard clustering results. In addition, we also tested whether using PCAs of the vegetation attributes as the clustering input altered our results and found that it did not. Details of these analyses are found in Suppl. material 1: appendix S2.

The seven groups are mapped in Fig. 1A (and individually in Suppl. material 1: fig. S2), with the vegetation attributes summarised in Fig. 2. In the section, Naming the groups and linking to previous vegetation descriptions, we name the groups, based on our interpretation of the literature and, for simplicity, we use these names henceforth. The groups are named scrubby caatinga, hyper-seasonal caatinga, heterogeneous caatinga, transition caatinga, arboreal caatinga, cerrado and rain forest. The seven groups are primarily distinguished by the amount of the vegetation in terms of canopy height, biomass and LAI. The presence of fire is the main defining factor separating cerrado from the other vegetation groups (Fig. 2). Most vegetation groups are identified by a particularly high value of at least one vegetation attribute, for example, rain forest has high biomass, LAI and canopy height and heterogeneous caatinga has high biomass heterogeneity. Only 0.24% of the clustered area had a probability < 0.45 of being in any group and was grouped as unclassified.

Random forest classification was highly accurate at predicting cluster membership, based on the predictors listed in section Random forest models to quantify the determinants of vegetation groups. The random forest model had an accuracy value of 81% for the held-out ‘test’ data set. The per group error rate from the random forest confusion matrix was below 24% for all groups, except the unclassified group, which had an error rate of 95%. Based on the importance values (which quantify the mean decrease in Gini coefficient when that variable is randomised; Fig. 3), the most important predictor of group membership was soil pH (mean decrease in Gini coefficient = 65,000). Following this were climate and soil variables (47,000 to 17,000) with travel time to nearest city the most important human use variable (27,000). The least important variables included all the other human disturbance related variables, soil cation exchange capacity, soil depth, Height Above Nearest Drainage, soil water content and geology. Surprisingly, geology was the least important variable by some distance – and repeating the analysis without soil variables (which might be somewhat correlated) gave the same result.

The partial dependence plots (PDPs; Fig. 4) for the three most important explanatory variables showed substantial, mostly non-linear effects. Overall, rain forest and arboreal caatinga groups show similar trends, being located in wetter areas. For rain forest, there is a strong pH and precipitation seasonality threshold, being more likely to be located on soil with pH < 5.5 and in areas with precipitation seasonality < 90. The caatinga groups are located in drier areas, with high soil pH, with heterogeneous and hyper-seasonal caatinga, in particular, having a strong affinity with for soil pH > 5.8 and hyper-seasonal caatinga an affinity for precipitation seasonality > 95. The rain forest and hyper-seasonal caatinga groups showed opposite associations with all three of the most important environmental variables.

The hyper-seasonal caatinga group generally was more sensitive to environmental variables compared to heterogeneous caatinga, although they exhibit similar trends. This could suggest the hyper-seasonal caatinga group has stricter environmental requirements and is less likely to be found in unfavourable conditions.

Figure 3. 

Variable importance plot illustrating the importance of explanatory variables in determining the structural groups. Values indicate the mean decrease in Gini Index after removal of each variable from the random forest classification. Shapes indicate the type of variable; cross = soil variables, circle = climatic variables, square = human-related variables, triangle = geology.

Figure 4. 

Partial dependence plots showing the effect of the three most important variables predicating vegetation type in northeast Brazil. A. Soil pH; B. Precipitation in the wettest month; C. Precipitation seasonality (coefficient of variation of monthly precipitation). The y axis shows the log odds of a pixel being in each vegetation type, given varying levels of the focal predictor variable, with all other predictors held at the mean of the dataset.

Our review of the literature identified nine vegetation groups we would expect to find in this region: rain forest, cerrado savanna, restinga, campos rupestres, cerradão and four groups of caatinga vegetation: crystalline, sedimentary, karst and arboreal (Fig. 5). Arboreal caatinga is forest-like, but crystalline, sedimentary and karst caatinga have distinct forms, determined by their underlying geology (Queiroz et al. 2017; Fernandes et al. 2022). Crystalline caatinga is the most typical caatinga phytophysiognomy, with deciduous, spinescent trees, a high diversity of herbaceous climbers and non-woody plants and few grasses (Quieroz et al. 2017). Crystalline caatinga is commonly found in the Sertaneja Depression, with shallow, fertile, stony soils and flat to slightly hilly topography (Queiroz 2006; Moro et al. 2016; Queiroz et al. 2017). Sedimentary caatinga – which has a more scrub-like formation – is found in Mesozoic sedimentary basins, with sandy, deep, oligotrophic soils (Rocha et al. 2004; Queiroz 2006; Moro et al. 2016; Queiroz et al. 2017) which may have a greater capacity to retain water. Thirdly, karstic caatinga is found on karst outcrops in small islands within sedimentary basins and has the richest soils in the region, distinct physiognomies and many endemic species (Queiroz et al. 2017; Fernandes et al. 2020).

Figure 5. 

Comparison of vegetation groups in NE Brazil according to the literature and structural groups identified in this work by unsupervised classification of vegetation attributes. Solid lines indicate clearly identified parallels between the literature definition and our structural groups and dotted lines identify more tenuous relationships between the structural groups and literature definitions.

We found the following links between the previous descriptions of vegetation in NE Brazil and our vegetation structural groups. One group clearly represented rainforest (of both the Atlantic and Amazon Regions). Cerrado and campos rupestres were represented together in the cerrado group and the remaining five groups represented different formations of caatinga. These caatinga groups were named arboreal caatinga, scrubby caatinga, transition caatinga, hyper-seasonal caatinga and heterogeneous caatinga. Arboreal caatinga is equivalent to the caatinga type of the same name as described in literature. Scrubby caatinga is equivalent to sedimentary caatinga and restinga. The hyper-seasonal caatinga and heterogeneous caatinga may collectively comprise the crystalline caatinga, given their geographical distribution, characteristics of high seasonality and shallow soils with high soil pH, which are typical of crystalline caatinga (Quieroz et al. 2017). Finally, transition caatinga seems to be a type of caatinga that is transitional between the more typical caatinga vegetation groups and arboreal caatinga. We did not identify a structural group corresponding to cerradão or karst caatinga vegetation.

Detailed descriptions of the vegetation groups can be found in Suppl. material 1: appendix S3; below we briefly outline each group.

The vegetation groups were assigned to the IUCN GET EFG groups present in the study region by comparing the descriptions of the EFGs (Keith et al. 2020b) and the vegetation descriptions above.

The clustering method identifies seven ecologically meaningful groups with different physiognomies. These structural groups align with previously recognised vegetation groups in the region – categorised using a variety of methods including floristic classifications (Queiroz 2006; Moro et al. 2016; Queiroz et al. 2017) (Fig. 5). Although we do not identify a single group corresponding to the previously described crystalline caatinga, our hyper-seasonal caatinga and heterogeneous caatinga together may comprise this formation. This suggests that our work may have generated more subdivisions within the crystalline caatinga, due to the emphasis on vegetation structure as opposed to floristic composition and evolutionary history.

The number of groups identified is suitable for capturing the heterogeneity of the region without being overly complex for interpretation at a regional scale. For further discussion of the effectiveness of our method, see Suppl. material 1: appendix S4 for an example of regions of topographic complexity; the Chapada Diamantina and Serra do Araripe. We used an unsupervised clustering approach, enabling us to categorise the vegetation as quantified by remote sensing products, rather than based upon preconceived opinions on vegetation in the region. However, there are limitations to this method, including the difficulties associated with labelling resulting groups. This interpretation may introduce biases in a similar manner to supervised classifications. Furthermore, quantifying the accuracy of an unsupervised approach to check whether the labelled groups match existing ecosystems requires extensive field effort; this may be a particular challenge in human-modified areas, which add complexity within an already complicated region. For example, we suggest the heterogeneous caatinga group may, in part, represent human-modified caatinga or may result from high levels of encroachment of invasive species (Nogueira et al. 2019). However, verifying this would take considerable on-the-ground effort. Finally, groups identified by unsupervised classification may not align with previous classification systems, as we found here for some caatinga groups, which may reduce the applicability of the research to some users. For example, we did not locate a group aligned to karst caatinga. However, we have made considerable efforts to relate our retrieved groups to previously described vegetation, in order to accurately place them within existing frameworks from literature.

Random forest classification showed that soil pH was the most important determinant of vegetation structural groups. Precipitation in the wettest month and precipitation seasonality were the next most important variables. These results broadly agree with the argument that climate is not the main determinant of vegetation type in NE Brazil and align with the findings of several authors who describe edaphic factors as important drivers of vegetation in NE Brazil (Murphy and Bowman 2012; Terra et al. 2018; Castro Oliveira et al. 2019; Maia et al. 2020; Souza et al. 2020).

The importance of soil pH as a determinant of vegetation type, may be due to its relationship with soil fertility. Hyper-seasonal caatinga and heterogeneous caatinga (collectively comprising crystalline caatinga) were associated with higher soil pH (Fig. 4), which agrees with previous findings of crystalline caatinga vegetation having higher soil pH (Castro Oliveira et al. 2019). On the other hand, rain forest and arboreal caatinga (to a lesser extent), are more likely to be found in areas of lower pH (Fig. 4). The shapes of the pH partial dependence plot for rain forest and hyper-seasonal caatinga suggests there is an abrupt threshold pH value for these structural groups, either side of which one is unlikely to be found. Other structural groups do not have such a strong association with a particular pH value (Fig. 4) and it is likely that different environmental variables determine their presence. However, it is also possible that pH may not be the key driver; other, correlated but unmeasured, environmental variables may also play a role. In particular, the use of climate variables to model pH in the SoilGrids product used here (Poggio et al. 2021) makes causality hard to infer. Furthermore, previous work has demonstrated uncertainty in SoilGrids variables in dry ecosystems (Cramer et al. 2019; Dandabathula et al. 2022), particularly for chemical as opposed to physical soil properties (Miller et al. 2024). Although some previous studies have found climatic, fire and anthropogenic variables to be of greater importance than soil in determining vegetation type, for example, in tropical Africa, (Bond et al. 2005; Greve et al. 2011; Pausas and Ribeiro 2017), others acknowledge the importance of including soil properties in studies on vegetation distribution – particularly in savanna ecosystems (Campo-Bescós et al. 2013; Arruda et al. 2015; Arruda et al. 2017; Oliveira et al. 2021).

Variables describing precipitation were the most important climatic determinants of vegetation structural group. This makes sense in a dry region where water availability is important for plants and where elevational variation is relatively limited. The IUCN EFG descriptions for the groups present in the region all describe water availability as an important driver for these groups (Keith et al. 2020b; Keith and Russell-Smith 2020; Lehmann et al. 2020; Pennington et al. 2020), which matches the importance of the precipitation regime in the random forest results. In addition, precipitation regime interacts with edaphic and other environmental variables. Maia et al. (2020) found a negative impact of precipitation seasonality and a positive impact of precipitation in the driest quarter on tree species richness in Caatinga and Cerrado sites. Both of these effects were mediated by soil sand content, being stronger in soils with less sand. Interactions between environmental variables increases the complexity in understanding determinants of vegetation structure. Hyper-seasonal caatinga has a particularly strong relationship with precipitation (Fig. 4), being much more likely to be found in areas with wettest month precipitation below 200 mm and precipitation seasonality above 95. Finally, it is important to note that these drivers display interactions including feedback loops, which makes their interpretation in structuring vegetation groups complex (see Suppl. material 1: appendix S2, p. 6 of Keith et al. (2022) for further discussion of interactions).

A difficulty in the analysis is that SoilGrids data do not include some relevant variables, including phosphorus and aluminium content which are known to be important determinants of tree species composition (Bueno et al. 2018). For example, although correlated to soil pH, aluminium content is considered a key feature differentiating caatinga and cerrado soils (Castro Oliveira et al. 2019) and, consequently, would be particularly relevant for this study. Furthermore, a caveat in using SoilGrids is that these modelled products utilise a number of covariate datasets, including climate variables, vegetation indices such as NDVI and raw MODIS bands (Poggio et al. 2021). These data are also utilised in our vegetation structural attributes, which means that they are not entirely independent, leading to some potential circularity in the results of the random forest classification. As such, further work would be required using regional sampling and modelling to improve the accuracy of this analysis, as carried out by Cramer et al. (2019) for the Greater Cape Region, South Africa.

We have identified both coherence and challenges with the existing IUCN’s GET EFGs, with the key issues surrounding the complexities of classifying dry forest, thicket and shrubland. We find that the transition caatinga and arboreal caatinga groups align well with the ‘T1.2 Tropical subtropical dry forests & thickets’ EFG, a finding supported by IBGE (2021), which demonstrated full equivalency of ‘forested caatinga’ to T1.2 (IBGE 2021). However, the placement of scrubby caatinga, hyper-seasonal caatinga and heterogeneous caatinga into existing EFGs is not clear. These groups might be considered to fall under either ‘T3.1 Seasonally Dry Tropical Shrublands’ or ‘T1.2 Tropical-subtropical dry forests and thickets’, but neither are a complete fit. This finding of non-alignment is also in agreement with IBGE (2021) which suggested that the caatinga vegetation groups have only partial equivalence to multiple EFGs (IBGE (2021) considered the caatinga partially equivalent to ‘T1.2 Tropical-subtropical Dry Forests and Thickets’, ‘T5.2 Thorny deserts and semi-deserts’ and ‘T3.1 Seasonally dry tropical shrublands’).

In the IUCN GET, dry forest and thickets are described as closed canopy forest, with seasonally high LAI and deciduous or semi-deciduous phenology and absence of grasses (Pennington et al. 2020). In contrast, T3.1 shrublands are low, open forests, shrublands and shrubby grasslands, which are sometimes evergreen and have a canopy height below 6 m. C4 grasses may be co-dominant, but not continuous and there are recurrent fires (Keith and Russell-Smith 2020). Scrubby caatinga, hyper-seasonal caatinga and heterogeneous caatinga are more or less shrubby in terms of remotely-sensed vegetation attributes (Fig. 6) and fit the main IUCN GET structural description of T3.1. However, according to our remotely-sensed data, scrubby caatinga, hyper-seasonal caatinga and heterogeneous caatinga do not burn, which is a key ecosystem process of T3.1. This lack of fire means they are, in some ways, a better fit for ‘T1.2 Dry forests and thickets’, but, in turn, classifying these three caatinga groups as dry forest seems inappropriate as they are distinct in height, biomass, LAI etc. from the more forest-like arboreal and transition caatingas (Suppl material 1: appendix S3; Fig. 6). Whilst it is possible that there are misrepresentations in the remote-sensing data or that, recently, fire has been suppressed in these vegetation types, on balance, we think the difficulty in alignment with the GET is because it is missing an EFG. This conclusion agrees with IBGE (2021), which suggests a new EFG be added to the GET to accommodate the predominant vegetation formations in the caatinga.

Figure 6. 

Scaled mean vegetation structural attributes for the IUCN Global Ecosystem Typology’s Ecosystem Functional Groups in NE Brazil. Axes show the percentage for that structural vegetation group of the maximum data point for each vegetation attribute. The vegetation attributes were (clockwise from top) aboveground woody biomass, canopy height (height), mean Leaf area index (LAI) in the dry season, mean Leaf area index (LAI) in the wet season, burn count (burn), aboveground woody biomass heterogeneity (bio.cv) and NDVI seasonality (seas). Scrubby caatinga, heterogeneous caatinga and hyper-seasonal caatinga are placed in the Non-pyric shrublands group (blue). Arboreal caatinga and transition caatinga are placed in T1.2 Tropical-subtropical Dry Forests EFG (red) and the cerrado and rain forest structural groups in T4.2 Pyric tussock savannas (yellow) and T1.1 Tropical-subtropical Lowland rain forests (green) EFGs, respectively.

Campos rupestres is a vegetation type in eastern Brazil that does align closely with EFG T3.1 (shrublands). This vegetation type has a different floristic and functional composition from the vegetation in our three caatinga groups. Most importantly, plant species in campos rupestres show many fire-resistance and fire-tolerance traits (Figueira et al. 2016) and there is floristic overlap with fire-prone savanna vegetation in the Cerrado. In contrast, many species, which are characteristic of the caatinga, do not feature fire adaptations, notably members of the Cactaceae family and other stem succulent species, such as the large Malvaceous tree Cavallinesia umbellata (Pennington et al. 2009; Oliveira-Filho et al. 2013; Queiroz et al. 2017). Furthermore, plants typical of caatinga vegetation often have thorns and are deciduous, with leaves having a low LMA, whereas campos rupestres plants are often evergreen, unarmed, with a high LMA (Queiroz et al. 2017; Mariano et al. 2021). Finally, the caatinga woody flora is more similar to that of tall dry forests scattered through the Cerrado and in the Chiquitania dry forest region of Bolivia, as opposed to campos rupestres (Silva de Miranda et al. 2018).

Given the discussion above, we suggest a new EFG should be incorporated into the GET to describe scrubby caatinga, hyper-seasonal caatinga and heterogeneous caatinga. However, it is beyond the scope of this paper to definitively determine the name of this new EFG and its location within the higher biome level of the Typology, given the regional nature of this study in comparison to the international scope of the IUCN Typology. Solely from our NE Brazil perspective, we suggest that the new EFG should be ‘T1.5 Tropical-subtropical non-pyric thickets and shrublands’. This is because of the lack of fire and the short stature, but closed canopy of our caatinga groups (Fig. 2; Dong et al., in review). This new EFG would correspond to the ‘T1 Tropical-subtropical forests’ biome and not the current T3.1 shrublands, as the latter are defined as open ecosystems (Keith et al. 2020a).

The new EFG which we propose would occupy the drier end of the seasonally dry tropical forest spectrum (sensu Pennington et al. (2000)) and be functionally distinct from ‘T3.1. Seasonally dry tropical pyric shrublands’. It would be analogous to the “succulent biome” sensu Schrire et al. (2005), describing a non-fire-adapted, succulent-rich, grass-poor biome, mapped by Ringelberg et al. (2020) using distribution of stem succulents. This splitting of dry forests from thickets affirms a division suggested by other authors (Schrire et al. 2005; Oliveira-Filho et al. 2013; Ringelberg et al. 2020; IBGE 2021) and provides an alternative to the broad definition of dry forest suggested by Murphy and Lugo (1986) and Pennington et al. (2000), which was adopted in the current definition of the T1.2 EFG. Our work suggests that splitting T1.2 into two EFGs is necessary to accurately describe vegetation in NE Brazil.

Overall, we find that the remote sensing-driven approach which we develop here can be used to identify structural groups in a complex region. Using an understanding of the drivers of EFG distribution including disturbance regime, soil and climatic factors, the structural groups from an unsupervised clustering can largely be grouped into EFGs as described by the IUCN GET. Therefore, this approach is suitable for fulfilling the fifth criterion of the GET (Keith et al. 2020b, 2022), that of providing spatially explicit maps, based upon an understanding of structure and vegetation ecology, using remote sensing. By going one level deeper into the thematic legend to map structural groups within EFGs, we have addressed the trade-offs between hierarchical levels in the IUCN GET – increased local applicability and realism at the local level, while methodologically and ecologically linking with the global EFGs. It is useful to understand highly complex regions with multiple interdigitated biomes and, as such, starting at the unsupervised structural level allows detailed information on the structural heterogeneity within EFGs to be obtained.

This work, therefore, adds to the growing body of work implementing the IUCN GET in a range of ecosystems (Murray et al. 2019; Karger et al. 2021). In addition, it extends the use of NE Brazil as a model system for biome delineation and mapping and demonstrates improvement in the ability of remote sensing since the efforts of Beuchle et al. (2015). Further work could aim to incorporate other strands of information into biome mapping in order to develop more ecologically meaningful maps. This would align with other parts of the ‘Ecosystem Properties’ method of defining biomes described by Keith et al. (2020b). For example, this could include the distribution of important functional traits such as CAM photosynthesis, C4 grasses, leaves with drip tips, deciduousness and spinescence (Conradi et al. 2020).