Climate, fire, and anthropogenic disturbance determine the current global distribution of tropical forest and savanna

Tropical forest and savanna biomes are pivotal in the functioning of the Earth system. Both are biodiverse and under increasing threat due to land clearing and anthropogenic climate change, and play important roles in the global carbon cycle, through maintenance of a large carbon pool in tropical forests, and exchange in savannas through extensive landscape fires. Reliable mapping of tropical forest and savanna is essential to understand how the current distribution of these vegetation types is controlled by climate land clearing and fire. Using Google Maps satellite imagery, we manually classified 24 239 random points as forest, savanna, or anthropogenic landscapes within the tropics and applied this novel dataset to defining the climatic zone where forest and savanna exist as alternative states. Because fire and climate are correlated, we developed separate geospatial models to rank the importance of climate, topography, and human influence on vegetation present. This modeling confirmed that those areas with more fires had lower probabilities of tropical forest, that forest was most likely in areas with high mean annual rainfall with little seasonal variation in precipitation, and that anthropogenic factors disrupt this environmental predictability. We also identified areas where tropical forest and savanna both co-occur, but these were relatively uncommon. These relationships suggest that future drier climates projected under anthropogenic climate change, combined with clearing and burning that have reduced tropical forest extent to a subset of its theoretical distribution, will lead to irreversible loss of tropical forests. Our modeling provides global mapping that can be used track further changes to distribution of tropical forests.


Introduction
Understanding the distribution of tropical forest and savanna remains a core concern of global biogeography.Tropical forests constitute one of the largest carbon pools in the biosphere (Post et al 1982, Berenguer et al 2014), whereas tropical savannas, the most fire prone ecosystems on Earth, drive seasonal flux of carbon between vegetation and the atmosphere (Van Der Werf et al 2003, Grace et al 2006, Duvert et al 2020).Both biomes are rich in biodiversity (Pennington et al 2018), which is being diminished by human land uses, particularly tree clearance for agriculture and pastoralism (Almeida de Souza et al 2020, Lopez-Carr 2021).Furthermore, climate change is likely to transform tropical forest through increased occurrence of drought and associated fires (Corlett 2016).Without effective economic incentives combined with sustainable management programs it is likely that the extent and ecological integrity of these systems will continue to decline (Andersson et al 2018, Cuenca et al 2018).Accurate determinations of forest and savanna distributions provide essential baselines to understand the rate, geographical foci and scale of forest and savanna loss (Aleman andStaver 2018, Zimbres et al 2020).Such baselines are also important in validating Earth system models that chart the likely fate of carbon fluxes and climate change (Tang et al 2019).
Historically, climate has been considered the primary factor determining the distributions of tropical forest and savanna (Whittaker 1974, Bowman 2000, Woodward et al 2004, Langan 2017).It is no coincidence that tropical forests are often called 'rainforests' and that savannas and associated fires are associated with monsoonal or seasonally dry climates.There remains debate and uncertainty around the environmental controls of these tropical biomes.Existing maps of the extent of tropical forest and savanna, such as the widely used World Wildline Fund biome map, rely on compilation and subjective synthesis of various geographic thematic data streams (Olson et al 2001), placing uncertainty over their reliability (Ocón et al 2021).Remote sensing products provide globally consistent and contemporary maps of tree cover that have been used to delineate tropical forest and savanna.For example, the Moderate Resolution Imaging Spectroradiometer (MODIS) tree cover product has been interpreted as distinguishing forest from savanna (Hirota et al 2011, Staver et al 2011).Such use of remote sensing to delineate forest and savanna has been subject to criticism because the algorithms that produce these global coverages can artificially reinforce discontinuities in canopy cover or overestimate actual tree cover (Hanan et al 2014, Staver et al 2015).
Fire is often implicated as a factor limiting forest distribution in the tropics (Bowman 2000, Staver et al 2011, Beckett and Bond 2019).This is a complex relationship, however, because pyrogeographic analyses have shown that fire activity is controlled with both climate and vegetation productivity.This creates a unimodal relationship, with maximum fire activity in savanna environments that have seasonal aridity (dry season) and productivity (wet season).Fire activity is constrained in both arid, low productivity environments where fuel biomass is limiting, and humid, high productivity environments biomass where fuel is typically too wet to burn (Bowman et al 2014).
Geospatial models of the environmental controls of forest and savanna routinely mask out landscapes drastically anthropogenically modified through land clearing and urbanisation (e.g.Zeng et al 2014, Aleman and Staver 2018).Nonetheless this approach cannot exclude the influence of humans in shaping forest and savanna distribution through localised impacts (e.g.swidden agriculture in forests, and pastoralism in savannas) (Wuyts et al 2017).For instance, various studies have demonstrated that human activities are drivers of shifts in vegetation states (Rietkerk et al 2004), and also of land area burnt (Archibald et al 2009).In more mesic areas, Veenendaal et al (2018) found fire maintains open savanna grasslands under deliberate human-mediated, high-frequency, late season fire regimes.It must be acknowledged that vegetation and fire patterns are also shaped by terrain that can alter the spread of fire (Bowman 2000, Ondei et al 2017), as well as creating environmental gradients that affect the distribution of forests such as rain shadows, although global analyses fire and forest and savanna distributions typically do not consider terrain variables (Hirota et al 2011, Staver et al 2011).
Global empirical analysis and modeling show that the predictability of savanna and forest biomes from environmental drivers is imperfect, with tracts of savanna occurring in environments seemingly suited for forest and vice versa (Swaine et al 1992, Moreira 2000, Russell-Smith et al 2003, Bond 2008, Murphy and Bowman 2012).For example, Pausas and Bond (2020) (2011) show that in areas with mild seasonality (SI) and intermediate levels of precipitation (1000-2500 mm y −1 ), alternative states of forest and savanna tree cover predominate depending on fire activity.Although mathematical models that result in bistability have been formulated (e.g.Staal et al 2016), it is notable that a range of definitions and means of detecting bistability have been proposed in landscape ecological analyses.Varied understandings of vegetation bistability range from those requiring experimental evidence of both multi-generational stability and state-switching of species suites at a given location (e.g.Connell andSousa 1983, Pausas andBond 2020), through to broader definitions that encompass bimodal forest cover distribution within climate bounds (e.g.Staver et al 2011).Similarly, a range of mechanisms, feedback components and climate constraints have been posited as driving bistability.We summarise some key definitions of bistability and likely drivers in table 1.In our analysis of predictors of tropical savanna and forest cover, we use a broad definition of bistability, following Staver et al (2011) to identify the existence of both high-and low-vegetation cover communities within a climatic envelope.
Here we undertake an analysis to evaluate the degree to which climate, fire, topographical and anthropogenic factors predict the extent of tropical forest and savanna in the early 21st century, and map the extent of the zone where they exist as alternative states.Rather than using a remote sensing tree cover product, we manually sample and classify publicly available global true-colour satellite imagery.Recent studies using a similar approach have highlighted anomalous areas of forest cover in arid States defined by stability over generational timescales, co-occurrence in the same environment, with demonstration of evidence for potential dynamic occupation of one site by both states.
Pausas and Bond (2020) Fire-vegetation-soil feedbacks determine forest and non-forest states in western Tasmania.
Wood and Bowman (2012) Local forest probability of between 40%-60%, derived from binomial model of precipitation total and seasonality.
This study landscapes (Bastin et al 2017).We analyzed the distribution of tropical forest and savanna with relation to climate and topography from global datasets.
Like prior studies, we also include fire as a predictor variable (Bowman 2000, Staver et al 2011, Beckett and Bond 2019), however we modeled climate-fire and climate-biome relationships individually to separate the highly correlated relationships between fire and vegetation, based on a methodology developed by Murphy et al (2010).Finally, our model output is compared to the widely used WWF biome map (Olson et al 2001), and satellite-derived classification of tropical broadleaf forest (Hansen 2000).

Methods
Commonly utilized global satellite tree-cover datasets have been identified as potentially displaying biases that may confound understanding of tree-cover discontinuities (Hanan et al 2014, Staver andHansen 2015).To establish a dataset of vegetation type independent of automatic satellite classification, we manually classified 24 239 randomly selected points within the tropics in areas expected to comprise forest-savanna gradients, defined as a 100 km buffer around the Tropical (A) and hot arid steppe (BSh) Köppen climate regions (figure 1), to ensure adequate sampling of the arid end of the productivity gradient around savanna regions.We excluded sampling from any area with elevation greater than 2500 m which corresponds to the expected altitude of treelines in the tropics (Rehm and Feeley 2015) using the EarthEnv global topographic dataset at a 5 km resolution (Amatulli et al 2018).Land cover type was determined using a webbased vegetation mapping tool developed for this study based on publicly available satellite imagery utilizing Google Maps imagery collected between 2014 and 2020.We took advantage of developments in geospatial tools within the last decade (Belward andSkøien 2015, Johansen et al 2015), with Google providing access to aerial images with very high spatial resolution (pixel size ∼1 m), which are sufficiently resolved to allow visual identification of individual tree crowns (Karlson et al 2014).Visual assessment of vegetation based on Google Maps data has been demonstrated previously (Bey et al 2016), and our method employs a similar but simpler single-point context-dependent classification system.Random points to be classified within the selected climate zones were generated and presented sequentially to a member of our team with expertise in vegetation classification, who manually labeled the area at the selected point as belonging to as either forest, savanna or urban landscapes on the basis of visual interpretation of the imagery, considering it within a surrounding 500 m × 500 m visual context.Archetypal images (figure S1) were supplied to the classifier as a visual guide, with these comparison images selected from confirmed examples of savanna and forest vegetation from across their global range.The classifier did not gauge an estimated tree cover percentage, but rather selected from the available fixed classification types.We took particular care to examine if the tree cover comprised plantations-areas where tree canopies were arranged in regular rows were excluded from the analysis.We also checked for any human influence (e.g.roads, houses, intensive agriculture, or industry) associated with point.Points were classified as urban landscapes when the predominant cover constituted roads, buildings, or gardens.In total, we classified 11 493 savanna, 7045 forest and 5701 urban points.We conducted an independent test of our classification by comparing our classified points with percent tree cover data from the 500 m MOD44B collection 5 product (DiMiceli et al 2011).This MODIS product has been shown to provide percent canopy cover at a scale and resolution appropriate for discriminating regional (Bartholomé and Belward 2005) and even global patterns (Van Nees et al 2018).Our validation against the MODIS data shows that our vegetation classes correspond broadly to MODIS tree cover values (figure S2).
The precipitation SI and mean annual precipitation (MAP) for each cell were derived from the WorldClim2 climate data set (Fick et al 2017).For each cell, the human influence index (HI) was estimated from the Last of the Wild v3 dataset (Venter et al 2018) at a 1 km resolution and resampled to the 5 km grid consistent with the other datasets.Topographic position index (TPI) for each cell was extracted from the EarthEnv global topographic dataset at 5 km resolution (Amatulli et al 2018).For fire modeling, we extracted the mean hotspot count per 5 km cell from the MODIS Aqua and Terra (product MCD14DL; Giglio et al 2016) active fire hotspot archive from 2002 to 2019, excluding fixed and low-certainty hotspots.

Data analysis 2.2. Occurrence of tropical forest
We constructed predictive models of tropical forest occurrence using binomial generalized additive models (GAMs).These are well suited for modeling presence-absence data (Hastie andTibshirani 1990, Guisan et al 2002).Additionally, our use of GAMs, as opposed to generalized linear models, is more appropriate as the assumption of a linear relationship between the response and explanatory variables are void, and because GAMs allow for nonparametric smoothing splines to be fitted to the covariates which more accurately represents the response curve (Elith and Burgman 2003).GAM models were built in R with the 'mgcv' package (Wood 2001), using a binomial probability distribution and logit link function.Detailed summaries of model structure and model fit diagnostics are available in supplementary methods.
To identify the drivers of vegetation type, we fitted GAMs with vegetation class as the response variable against 4-knot thin-plate splines of MAP, precipitation SI, human influence, and topographic position.A total of 18 538 data points were used in the final analysis (17 045 forest and 1493 savanna) with urban points excluded.The suite of models included a global model with all four predictor variables, climate plus topography, climate plus human influence, a model excluding the two climate variables, and a model including and climate only.We also ran separate models with the same response and explanatory variables for classified samples from each of the four biogeographical realms (5202 data points for the Neotropics, 9258 for the Afrotropics, 2106 for Australasia, and 1335 for Indomalaya).The most supported set of predictors was inferred using AIC-based model selection (Burnham and Anderson 2002) corrected for sample size, and the most parsimonious models were selected as those with the lowest AICc score, and greater than 2 AICc difference from other models.To obtain a measure of relative support of each model combination, we also computed their Akaike weight w, which refers to the probability of a model combination being the best supported one among a given model set.
Predictions from the best model were used to plot a SI vs. MAP surface chart showing forest occurrence probability for climate areas where sample points were present.To visualize patterns on each of the four main biogeographic realms of focus, we used models trained on data subset for each realm to further graph individual SI vs. MAP forest probability surfaces, based on the best performing model in each case.To visualize the probabilities of savanna and forest on these charts, predicted forest probabilities across a gradient of MAP and SI, with HI and TPI held at their median values, were generated for points in the climate space.These forest probabilities were also mapped globally within the tropical domain and transformed to quartiles for construction of bi-choropleth map overlaying forest probability with fire activity and human influence.

Occurrence of fire
Hotspot density was modeled with GAMs using the same variable suite and model structure as the best performing forest cover model but with a quasipoisson distribution to account for the continuous but skewed fire distribution.Predicted hotspot density was then mapped within the domain and transformed to quartiles for construction of a bichoropleth map overlaying fire activity and forest probability.
We then developed separate generalized linear models (Gaussian distribution) of fire occurrence (log-transformed MODIS hotspot count) for forest and savanna points using the global suite of variables.This analysis follows the method of Murphy et al (2010) and was intended to disentangle the influence of fire on vegetation community (forest vs. savanna) or vice versa.Given that forest presence is expected to suppress fire, the effects of these climate variables on fire activity are modeled in the two vegetation communities separately.If the same variables are important in both models, but with opposite relationships with fire, it can be concluded that the distribution of forest is driven by the inherent distribution of fire, and areas of low fire frequency.Alternatively, if the relationship between fire and the climate variables is in the same direction, we could conclude that vegetation patterns are controlled by both fire and climate, which introduces some indeterminacy more consistent with bistability.Marginal effect plots of fire occurrence against MAP and SI were plotted for each vegetation type.

Results
Nearly a quarter of our random sample points (23.5%) were classified as belonging to anthropogenic landscapes, and were discarded from our analysis of forest cover.We classified 7045 points as forest, which had an interquartile range of 1569 mm-2368 mm MAP and a precipitation SI interquartile range of 33-70.Savanna vegetation (11 493 points) had an interquartile range of 419 mm-1188 mm MAP and a precipitation SI of 75-118.Among the points classified as forest, 8% occurred in sheltered topographic positions at the 5 km scale (TPI < −0.5) while only 4% of non-forest points occurred in topographic depressions.
The most supported model in terms of predictions of forest probability was the global model (table S1).The model that excluded the climate variables performed poorly, while excluding human influence and topography had negligible impact on model performance.Globally, forest was predicted to occur in areas of low SI (and precipitation >1000 mm) (figure 2), which suggests that annual precipitation lower than 1000 mm and long dry seasons may limit forest establishment.The indeterminantzone had an interquartile range of 1311 mm-1702 mm MAP and a precipitation SI interquartile range of 48-83.Such 'bi-stable' areas were common in the Afrotropics (figure 3(a)), the Neotropics (figure 3(c)) and in Indomalaya (figure 3(d)), but less so in Australasia (figure 3(b)).When the model outputs were projected unto the global map, areas with high forest probability (>60% probability) covered 1501 Mha (32.7% of the projected area); areas with low forest probability (<40% probability) comprised 2674 Mha (58.2% of the projected area) and intermediate areas covered 418 Mha (9.1% of the projected area) (figure 4).GeoTIFF rasters of model projections and classified data are available in a public data repository (Williamson and Tng 2023).These projections are also associated with a map of model variance (figure S3) showing low variance in most areas except in south Brazil, India, and some areas in southeast Asia.These areas correspond to areas where predicted forest extent based on climate is disconnected from satellite-based mapped extent of forest (Hansen et al 2000) or theoretical tropical forest biome (Olson et al 2001) (figure 5).In savanna locations, fire frequency was positively associated with precipitation and SI (figure 6(a)).By contrast, forest showed the opposite trend where high precipitation and low SI resulted in low fire incidence (figure 6(b)).The inverse relationship of driving variables in these two fire models in contrasting vegetation types indicates a negative relationship between tree cover and fire.Reinforcing this, we also found that across all biogeographic realms fire probability decreases as forest probability increases (figure 7).In our graphic model represented as choropleth maps (figure 8), we show the trivariate maps of forest probability, human influence, and fire activity (figure 8(a)).Forest probability is lowest where fire activity is high (figure 8(b)), and where human activity is high (figure 8(c)).

Discussion
Our analyses confirm the universality of MAP and dry season length as key factors driving the distribution of tropical forest.Areas with high probability of forest (>60%) occur in the Neotropics, Afrotropics and Indomalayan tropics, closely matching current core areas of tropical forest distribution.Our analyses identified areas where the probability of forest and savanna were approximately the same (regarded as the zone of 40%-60% probability).Separate analyses demonstrated that these climate factors also control the frequency of fire.Our analysis found a minor effect of coarse-scale topography and human influence in shaping forest distribution.

Global tropical forest and savanna mapping
Reliable vegetation mapping is essential for meaningful earth system research; hence an important practical finding of this study is the support for the WWF biome mapping of tropical broadleaf forests (Olson et al 2001).Remarkably, despite having been cited over 8000 times in Google Scholar (as from February 2022), this product has not been previously validated using independent data.We found that our predictive models closely match the boundaries of broadleaf tropical forest biomes, as well matching automated    and Jansen 1998).Indeed, the mismatch between the potential and actual distribution of tropical forest highlights the extraordinary impact humans have had on the biosphere.Our approach can be used as a baseline to track further erosion of tropical forest cover in response to deforestation, fire, and climate change.

Climate, fire, and anthropogenic disturbance
Tropical forest species are renowned for luxuriant plant growth and high productivity where drought stress is rare.Droughts in tropical forests have been known to substantially impact primary productivity and can cause die-back with substantial loss of carbon (Phillips et al 2009, Corlett 2016).Tropical forest drought is often associated with fire in an ecosystem that otherwise does not support landscape fires.Contrastingly, savanna trees are typically more tolerant of drought stress, often associated with seasonal largely rain-free dry seasons.The oscillation of wet and dry seasons favours frequent landscape fires to which savanna trees are adapted to and recover readily from fire disturbance.
Fire is an important control of the macroecological distribution of forest and savanna (Murphy andBowman 2012, Lasslop et al 2016).The effect of fire on shaping forest and savanna distributions is difficult to separate from climate given their correlation with each other (e.g.Murphy et al 2010).By modeling fire separately from the probability of forest we demonstrate that areas of high forest probability occur in corresponding areas with low fire probability.Our analysis also found that a broad index of topographic variation played a significant, albeit small, additional role in predicting forest probability and fire activity.Regional topographic settings such as plateaus, gorges and mountain ranges are known to provide fire refugia (e.g.Ash 1988, Bowman et al 2010) but at a global scale these terrain effects are subsumed by climate.A future direction for research is local-scale analysis of topographic effects within the transition zone, to clarify the role of topography-driven moisture regime in favouring forest or savanna.We also found that our index of human influence shapes the occurrence of forest distribution.For example, our modeling shows areas, such as the Indian subcontinent and southern Brazil, are suitable for forest but have high human footprint.Such anthropogenic landscapes have contrasting patterns of fire activity compared to natural areas.For example, tropical forest conversion to pasture is associated with increased burning, but by contrast savanna area converted to agriculture has substantially reduced fire activity.

Bistability of tropical forest and savanna
Several satellite-based studies have suggested that there are large areas of the tropics than can support either tropical forest or savanna (Hirota et al 2011, Staver et al 2011).It has been argued the breakdown of the predictability of forest and savanna biomes from environmental drivers is the product of a fire-driven 'alternative stable state' dynamic, where forest is fire-excluding and savanna is fire-promoting (Staver et al 2011).We identified a much smaller geographic extent of such areas where alternative stable state model is likely to explain equal probabilities of forest and savanna occurrence, particularly identifying areas in central Brazil and around the Congo in Africa.It is possible such areas lacking predictability from environmental drivers may be capturing tropical dry forests (Sunderland et al 2015), which can have closed canopies during the wet season or vary considerably in tree cover and occur in areas with high SI (Sunderland et al 2015, Dryflor et al 2016).Indeed, much of the area where we have >40% forest probability lies where others have mapped dry forest (David et al 2022).In interpreting model output of forest probability, a probability range of 40%-60% was interpreted as indicative of a zone where forest presence was indeterminant, and it should be noted that this interpretation is somewhat arbitrary, and wider bounds would result in a conclusion of greater area of indeterminant vegetation.However, our zone of indeterminacy does concur with recent work identifying a limited zone of ASS in Africa based on climatic suitability for various plant growth forms (Higgins et al 2023).Furthermore, a wide cover range definition of bistability would undermine the concept by artificially emphasizing the continuous cover gradient between the distinct forest and savanna biomes.Our results lead us to suspect that truly bistable and ecologically transitional states may exist only in narrow ecotonal vegetation between wet forest types and other arid forest types.The existence of savanna vegetation in climatic zones that are more likely to support forest can be attributed in part to interactions between fire and humans, which can drive plant communities to a large-scale dieback of forests through a savannization process (Flores et al 2016, Staal et al 2018).Conversely, changing burning practices and land uses such as pastoralism, as well as increased atmospheric CO 2 due to anthropogenic activities have resulted in widespread encroachment of forest into savannas (Hecht and Saatchi 2007, Aráoz and Grau 2010, Tng et al 2012, Nanni et al 2019).

Conclusions
Using a random sampling of manually classified global satellite imagery, we mapped the distribution of tropical forest and savanna in the global tropics and subtropics and used statistical models to identify the environmental drivers of forest distribution.Our results confirm that rainfall is the overarching predictor of forest and savanna distribution where fire activity is low in the former and high in the latter.In contrast to previous studies, we suggest that there are only small areas where climate does not reliably predict forest and savanna.Comparison of our classified sample sites to previous attempts to map land suitable for tropical forests suggest large areas have transitioned away from forest, likely due to anthropogenic disturbance.Anthropogenic climate change combined with land cover clearance is likely to further reduce the cover of forest in coming decades.Our maps provide a useful baseline to gauge the rate and geographic location of the major Earth system transformation.

Figure 1 .
Figure 1.Distribution of classified points within Tropical (A) and hot arid steppe (BSh) Köppen climate zones, colored by class.Map lines delineate study areas and do not necessarily depict accepted national boundaries.

Figure 2 .
Figure 2. Predicted forest probability within the global sampled climate space of Mean Annual Precipitation and Precipitation Seasonality Index based on the best performing generalized additive model, with human influence index and topographic position index held at their median values.Colors denote the predicted probability of forest in 10% intervals.

Figure 3 .
Figure 3. Predicted forest probability within (a) Afrotropics (AT); (b) Australasia (AA); (c) Neotropics (NT), and (d) Indomalaya (IM) climate space of Mean Annual Precipitation and Precipitation Seasonality Index based on the best performing generalized additive model, with human influence index and topographic position index held at their median values.Colors denote the predicted probability of forest in 10% intervals.

Figure 4 .
Figure 4. Predicted distributions of forest and non-forest vegetation types across the tropics based on best performing generalized additive model fitting forest occurrence to climate, topographical and human influence.Grey colour indicates areas outside of sampling area, based on climate type or elevation mask.Map lines delineate study areas and do not necessarily depict accepted national boundaries.

Figure 5 .
Figure 5. (a) Within the analyzed climate zones, boundary of the modeled contour representing 50% forest occurrence probability (red line) overlaid on the Hansen (2000) global satellite-based classification of extant evergreen broadleaf forest (dark green), and the Olson et al (2001) biome boundary of tropical and sub-tropical moist broadleaf forest (light green).(b) 50% forest probability area shown in red.

Figure 6 .
Figure 6.Effect plots for generalized linear models fitting the fire frequencies to environmental variables.Effect plots using the (a) Non-forest savanna grid cells (n = 9904; 29.3.7%deviance explained) and (b) forest grid cells (n= 6208; 25.5% deviance explained).Fuzzy points indicate density of training data.

Figure 8 .
Figure 8.(a) Trivariate choropleth map of forest probability, human influence, and fire activity.(b) Bivariate choropleth map of forest probability quartile and fire activity quartile, and (c) forest probability quartile and human influence index.Map lines delineate study areas and do not necessarily depict accepted national boundaries.
describe stable non-forested (open) vegetation in climates that are warm and wet enough to support forests.Conversely, forest vegetation can occur in dry or highly seasonal climates (<1000 mm annual precipitation) (Bastin et al 2017, Schepaschenko et al 2017, Maestre et al 2021).

Table 1 .
Review of definitions and of bistability and alternative stable states as they relate to forest and savanna systems.Local coexistence of forest and savanna states only possible very near critical climate threshold points, although with hysteresis driven by fine-scale edaphic/topographic variability.Use a value of 50% tree cover to separate basins of attraction.