Rootzone storage capacity reveals drought coping strategies along rainforest-savanna transitions

Observation-driven (including remote sensing) data of evaporation and precipitation were selected to derive rootzone storage capacity (see section 2.2). Evaporation is defined here as the total of the evaporative fluxes from soil moisture, interception, transpiration and open water [27]. We selected data that were free from prior assumptions of biome dependent parameterization and soil layer depth. Daily precipitation was obtained from measurements of the Climate Hazards Group InfraRed Precipitation with Station data (CHIRPS) at 0.25° resolution for the years 2001–2012 [28]. In the remainder of the manuscript, we will refer to this data as rainfall because, in the tropics, rainfall is by far the dominant form of precipitation. Monthly evaporation estimates were obtained for the same period from three datasets: (1) Breathing Earth System Simulator (BESS) at 0.5° resolution [29], (2) Penman-Monteith-Leuning (PML) at 0.5° resolution [30] and (3) FLUXCOM-RS [31] at 0.083° resolution. All evaporation datasets were interpolated to 0.25° resolution using nearest neighbor (for BESS and PML) and spatial average (for FLUXCOM-RS). We computed equally-weighted ensemble evaporation to minimize any potential bias. ERA5 daily evaporation [32] at 0.25° resolution was used to downscale monthly evaporation to a daily resolution. These evaporation datasets were either derived from remotely sensed observations or validated against observational evaporation values from FLUXNET sites [29–31].

Remotely-sensed MOD44B (Version 6) annual tree cover (TC) data at 250 m resolution [33] was used to analyze the forest-savanna transitions in South America and Africa. TC represents the above-ground canopy cover in a pixel (%), and is not to be confused with seasonal dynamics of leaf phenology. This TC data was spatially aggregated to 0.25° resolution for the period of 2001–2012 (figure 1(a)). To minimize the human influence on the natural water cycle, we removed grid cells with croplands and pastures greater than 30% based on the data in Foley et al [34] (at 0.083° resolution), as well as human-influenced and non-terrestrial land cover classes from the International Geosphere-Biosphere Program (IGBP) land classification [35] (at 1 km resolution). Removed classes include 'permanent wetlands', 'urban and built-up lands', 'snow and ice' and 'water bodies'. These datasets were also spatially interpolated to 0.25° resolution. Three other datasets related to water table depth [36], rooting depth [26], and ecoregions [37] were used to further support our analysis.

Zoom In Zoom Out Reset image size

Rootzone storage capacity (S_r) is the maximum amount of soil moisture that can be accessed by vegetation for transpiration [23]. Plants can increase S_r by expanding their roots in the soil laterally as well as vertically. We adapted the mass-balance methodology in Wang-Erlandsson et al [25] to estimate S_r from the maximum annual accumulated water deficit, which was calculated using daily estimates of rainfall and evaporation (see section S1 and equations 1–3 in supplementary information). This methodology is based on the assumption that ecosystems do not invest in expanding their storage capacity more than necessary to bridge the water-deficit experienced by the vegetation in dry periods (i.e. periods in which evaporation is greater than rainfall, irrespective of the seasons). Since remote-sensing based time series of evaporation and rainfall were assumed to reflect the actual soil moisture availability [25, 38], we can use them to derive the capacity of ecosystems to store water in their rootzone for use during dry periods. A 20 year drought return period based on the Gumbel extreme value distribution (equation 4 in supplementary information) was used to calculate S_r (figure 1(b); equations 5–7 in supplementary information). We acknowledge that forests typically adapt their S_r to drought return periods of >40 years, savannas for 10–20 years and grasslands to <10 years [23–25]. However, rather than assigning different drought return periods to different land cover types (forest, savanna and grassland), we chose a uniform 20 years drought return period (following [39]) in order to avoid artificially introduced S_r transitions between landscapes.

While being primarily dependent on climatic factors [23], we also acknowledge that S_r of an ecosystem is also dependent on geology such as soil depth to bedrock which could physically limit S_r and soil properties such as field capacity requiring roots in sandy soils to root deeper to achieve the same S_r. However, their strong dependence on climate provides a similar or better representation of hydrological regimes than the soil-derived S_r [23, 24]. Moreover, since S_r derived in this study only represents the hydrological storage capacity of the root zone, the relation with the ecosystem's rooting depth or rooting structure thus depends on several factors. However, maximum rooting depth data [26] and existing literature were used to interpret the subsoil dynamics of the rainforest-savanna ecosystem (see section 3.3).

We chose six representative transects following the longest possible transition line from the densest part of the rainforests all the way into the savannas and grasslands (figure 1(a)). To reduce any local signal noise and regional spatial heterogeneities, we applied a second-order polynomial based on the Savitzky–Golay smoothing technique [40] over a window of seven grid cells along the transects. Transect methods have proven to be able to clearly distinguish spatial trends in what seem to be heterogeneous ecological patterns at first sight (e.g. [41, 42]).

We classified the drought coping strategies of the ecosystem based on the transitions along the transects, defined using states and trends therein of empirical and statistical proxies. The empirical characteristics were based on the inspection of the trend and magnitude of TC and S_r. The statistical characteristics were based on the correlation coefficient, covariance (equation 11 in supplementary information) and confidence interval (CI; 95% CI across time) between TC and S_r. A moving window of seven grid cells was selected for statistical characterization along the transect.

The western part of the Amazon has the lowest S_r (<100 mm) with a high TC (>75%) (figure 1). In contrast, an increasing trend of S_r can be observed from the eastern part of the Amazon until the north-eastern and central-western region of Brazil with a consecutive decrease in TC estimates. However, the magnitude of S_r starts to decrease again when moving towards the far eastern part of South America or south of 15°S. In Africa, the southern extent of the Congo rainforest has the lowest S_r (<100 mm) with the highest TC (>70%). Moreover, S_r is higher at lower TC towards the Southern African savanna (around 15°S). Further southwards (<20°S) or northwards (>10°N), the S_r estimates are, however, lower again. These patterns indicate that there is no monotonic relationship between TC and S_r, as will be analyzed in further detail in section 3.2.

The transects generally reveal non-linear and non-monotonic relationships between S_r and TC in South America (figure 2) and Africa (figure 3). At the start of transects #1–6, at TC > 70%, S_r tends to be around 100 mm and remains unchanged along the transect. Following the transects towards the relatively lower TC regions (somewhat climatologically drier), S_r starts to increase strongly relative to only a minor decrease in TC. Only after the increase of S_r, we observe a sharp decline in TC with only a small increase in S_r . After a certain point beyond which S_r, however, does not further increase. Instead, both TC and S_r start to decline simultaneously (figures 2 and 3). This point indicates the region of forest-savanna transition.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

In South America (figure 2), S_r of both transects maximizes just before or at the forest-savanna transition zone. The maximum S_r is around 750 mm for the Maranhão Babaçu forests (in transect #1), whereas, for the Madeira-Tapajós moist forests (in transect #2), it was around 550 mm. These differences are likely to be caused by the higher seasonality in rainfall (figure S2) in the Maranhão Babaçu forests (transect #1). S_r and TC, both decrease after the forest-savanna transition, which happens gradually towards Caatinga (S_r until about 500–600 mm in transect #1), and quite abruptly towards Dry Chaco (S_r until about 125–200 mm in transect #2). The higher S_r in Caatinga can be explained by the vegetation responding to a higher temperature during the peak dry-season, which causes high potential evaporation rates and ∼75% of the annual rainfall is being evaporated [43]. The rainfall seasonality is also much stronger in Caatinga compared to Dry Chaco (figure S2), which is caused by the larger influence of the Intertropical Convergence Zone (ITCZ) and Atlantic sea surface temperatures [44].

In Africa (figure 3), the changes along the transect are similar to those observed in South America. Starting again from high TC (>70%), we find that it corresponds with a minimum S_r of about 100 mm. S_r increases to around 400 mm before the forest-savanna transition (transects #3–6). The highest S_r of around 550 mm is, however, found in the Central Zambezian wet Miombo woodlands (transect #5), which is probably due to high evaporative demand and high rainfall seasonality (figure S3) due to the ITCZ (e.g. [45]).

Before the forest-savanna transition, a gradual increase in maximum rooting depth with increasing S_r is observed. However, it remains mostly unchanged (on average with a lot of spatial variabilities) when S_r starts to decrease after the forest-savanna transition (figures 2 and 3). The low values of S_r in savannas are paradoxical as savanna trees are known to have deep roots [46]. However, this analysis shows that those deep roots are not representative of the subsurface water accessible to the ecosystem per unit area (i.e. S_r), which will be further discussed in section 3.3.

The two continents show a difference in S_r along the transects. This difference is smaller for the denser part of the forest (i.e. at >65% TC, the maximum S_r is around 350–400 mm for South America and 300–350 mm for Africa) compared to forests with lower TC (i.e. maximum S_r around 550–750 mm for South America and 400–600 mm for Africa at 30%–40% TC). These differences can be explained by the ecosystem's adaptability to seasonal moisture carryover capacity (i.e. high wet-season rainfall in South America (figure S1) allows for more water storage for the dry-season) [47] and high water use efficiency (i.e. carbon uptake per unit water loss through transpiration) of African forest [48]. We think that the abundance of the high water use-efficient C4 grasses in African forest and savanna [49] further reduces the competition for moisture uptake in the African ecosystems. Similar to the findings of Guan et al [47], these S_r differences indicate a high buffer capacity of the Amazonian rainforest and a low buffer capacity of the African rainforest, making the latter more sensitive to droughts [50].

In order to define distinct drought coping strategy classes, we analyzed the drought-induced forest dynamics (above- and below-ground) for different TC and S_r relations. We categorize the drought coping strategies into four classes (table 1) based on the empirical (table 1; figures 2 and 3) and statistical analysis (table 1; figures S4 and S5). Furthermore, we synthesize responses between TC and S_r not only in terms of absolute magnitude but also based on the spatial trends of TC and S_r (figure 4). Moreover, we link our findings with the existing literature in the following paragraphs to provide support for our classifications.

Zoom In Zoom Out Reset image size

3.3.1. Lowly water-stressed forest

3.3.2. Moderately water-stressed forest

3.3.3. Highly water-stressed forest

3.3.4. Savanna-grassland regime

Our classification and identified regions of forest-savanna transition (red dashed lines and red hatched regions in figures 2 and 3) correspond well with the ecoregions [37]. For transects #1 and #2, the forest-savanna transition falls in the Maranhão Babaçu forests and Chiquitano dry forests, respectively. Both these ecoregions are very close to Caatinga and Dry Chaco characterized by a much drier climate and fire stress. This climate-fire feedback on vegetation may push and maintain the ecosystem in a savanna state [67, 68]. However, a clear transition could not be defined for transect #3 due to lack of a strong statistical signal, but for transect #4 a significant portion of the transition region lies on the Western Congolian forest-savanna. This is identical for transects #5 and #6, where the transition zones lie in the Central Zambezian wet miombo woodlands and Northern Congolian forest-savanna ecoregions, respectively.

In the absence of longer time-series of observed historical climatological and ecological data for rainforests, the 'space-for-time' assumption is often used to infer temporal ecological trajectories from available spatial patterns [69]. An application of the 'space-for-time' assumption on a bundle of transects (figures 2 and 3), each representing a unique climatological and ecological succession, could help us infer potential changes in S_r, TC and drought coping strategy class with changes in climate. Given future intensification of droughts and lower access to water availability, ecosystems could be pushed into drier drought coping strategy classes (figures 2–4). Most relevant in this context are the regions of highly water-stressed forests as those regions are most likely to exhibit a transition to a savanna state. The observed moderate-high variance in S_r and TC for highly water-stressed forest (figures 2 and 3, table 1), further supports the interpreted vulnerability towards a savanna transition [70]. Such a climate-induced vulnerability may be aggravated by deforestation [4] and local water usage for irrigation [71], which may trigger an even earlier transition. Moreover, the simultaneous occurrence of all observed factors may further reinforce a forest loss-drought feedback [4] into self-amplified forest mortality [7] and thus trigger abrupt large-scale regional changes. The fatality of such changes might be amplified when it appears that these changes cannot easily be reversed [55, 67, 72].

The present study uses an ensemble product of three independent evaporation datasets (BESS, FLUXCOM-RS, PML; see section 2.1) to minimize the single product bias. Conducting a sensitivity analysis (figures S6 and S7) resulted in a concurring magnitude of S_r along the lowly water-stressed, moderately water-stressed and highly water-stressed forest classes. In contrast, significant differences were observed in the savanna-grassland regime: S_r ^FLUXCOM-RS is the highest, while estimates from S_r ^BESS and S_r ^PML were found to be similar to S_r. Since soil moisture is dependent on evaporation (see section 2.2), the different evaporation datasets used reflect a different soil moisture constraint. S_r from FLUXCOM-RS corresponds to a higher soil moisture availability (high FLUXCOM-RS evaporation) than BESS and PML in arid and semi-arid regions (figures S6 and S7), thus signifying a high uncertainty of the S_r estimate in the savanna-grassland regime. Jung et al [31] suggested that the reason for high FLUXCOM-RS evaporation fluxes (15%–20% greater than multi-tree ensemble (MTE) and LandFlux-EVAL) could be the choice of machine learning methods or poor constraints by flux tower stations in the arid and semi-arid regions. Despite these differences, we found that when we use any of the three individual evaporation products, it does not have any or very little influence on the location/region where we found the forest-savanna transition. The main conclusions drawn from the study do, therefore, not depend on the dataset used.