Introduction

As a major force in the evolution of many ecosystems globally, fire is often critical to the maintenance of ecological structure and function. Knowledge and application of fire science is especially important in the Anthropocene, in which many pre-historic disturbance regimes have been supplanted by human management (Bowman et al. 2011, Coughlan and Petty 2012). In general, mesic rangelands (savanna and grassland ecosystems) are both particularly fire prone and heavily influenced by human impact (Sauer 1950, Ellis and Ramankutty 2008). Mesic rangeland in sub-Saharan Africa is especially fire prone and has a long history of human interactions with fire that continues in the modern era (Shaffer 2010, Laris 2011, Trollope 2011).

While many studies report fire effects in rangeland, much of this work occurs in a wildfire context, and little has addressed temporal and species-specific trends within fuelbeds to support proactive fire management practices and policies. A common trope throughout sub-Saharan Africa—that fire causes rangeland degradation—supports anti-burning policies (Laris and Wardell 2006, Nyamadzawo et al. 2013) and sets a wildfire context in which observations of fire effects are based on fully cured fuels and extreme fire weather and behaviour. The general emphasis on fire effects, positive or negative, often focuses on specific outcomes such as primary productivity, plant community composition, or biodiversity conservation (Snyman 2004, Uys et al. 2004, van Wilgen et al. 2011) and overlooks potential effects of the plant community itself on the nature of the fire being studied.

Globally, fire ecology has begun to consider composition and structure within grassland fuelbeds, but most studies in sub-Saharan Africa regard grass-dominated fuelbeds as homogeneous. Fuelbed heterogeneity can arise from spatially explicit disturbance patterns that drive spatial patterns of fire spread and behaviour (Kerby et al. 2007, Davies et al. 2010, Leonard et al. 2010). Likewise, introducing substantially different fuel types can alter the spatial distribution and intensity of wildland fire (D’Antonio 2000). Similar work in sub-Saharan Africa is lacking.

In this study, we compared seasonal trends in fuel moisture of common rangeland grasses at two locations in South Africa. Sampling began prior to the conventional burning period and outside the range of conditions associated with wildfire, affording an opportunity to study species-specific differences and associations between fuel curing and soil moisture. Variation in fuel moisture can account for high variation in fire behaviour (Govender et al. 2006), and live fuel moisture can vary with soil moisture (Qi et al. 2012) but, to our knowledge, the relationship has not been tested for curing grassland fuels. It is possible that environmental data can serve as proxies for fuel moisture to support prescribed fire planning, assess wildfire risk, and predict or explain variability in fire effects. Within this context, we addressed the following questions: (1) Do grassland fuels cure differently among species or across locations? (2) Do differences in curing meaningfully affect fire behaviour? (3) Is fuel moisture associated with soil moisture?

Methods

Study Area

The study included two locations in KwaZulu-Natal, South Africa: a mid-altitude site (840 m a.s.l.) at the University of KwaZulu-Natal Ukulinga Research Farm in Pietermaritzburg (30° 24′ S, 29° 24′ E; hereafter Ukulinga) and a high-altitude site (1850 m a.s.l) at Highmoor in the central region of the uKhahlamba-Drakensberg Park (20° 32′ 07″ S, 29° 62′ 16″ E). Ukulinga is characterised by hot summers, mild winters, and 838 mm mean annual precipitation, mostly from summer rainstorms (Buis et al. 2009). Highmoor is characterised by cold winters, mild summers, and 1400 mm mean annual rainfall (Mentis and Bigalke 1981) occurring over a long, foggy summer rainy season. The Soil and Terrain Database for southern Africa classifies soils at Ukulinga as Dystric Regosols and soils at Highmoor as Eutric Regosols (ISRIC 2008). Regosols are a taxonomic “rest group” of soils that do not fit other major soil groups due to a general lack of diagnostic horizons characteristic of eroded land and mountainous regions (FAO 2001). In our observations, soil at Highmoor tended to have higher proportions of organic matter than soils at Ukulinga.

Despite climate and elevation differences both locations are characterised by grassland dominated by warm-season, C4 grasses including Themeda triandra Forssk., Alloteropsis semialata (R.Br.) Hitchc., and Aristida junciformis Trin. & Rupr. Because of the higher elevation, Highmoor sits at a transition between sub-montane grassland dominated by warm-season species and mixed Themeda-Festuca montane grassland (Killick 1963). Thus, large patches of the cool-season, C3 Festuca costata Nees occur among warm-season grassland at Highmoor.

Data Collection

We sampled fuel and soil moisture five times at approximately monthly intervals at both locations in 2014 (Highmoor, 8 March through 26 July; Ukulinga, 25 March through 2 August). Sampling was timed to span the end of the growing season into mid-winter when grassland fuels are typically fully cured (Everson et al. 1988). All three C4 grasses (Themeda, Alloteropsis, Aristida) were collected at each location in addition to the C3 Festuca collected only at Highmoor. We used a spatially nested design wherein two samples per sample type (three to four species plus soil) were collected at two sites per location during each sampling event (n = 2 true replicates per sample per event per location).

We sampled fuel and soil moisture content by clipping all aboveground biomass from a randomly located grass clump at ground level for each species, and by digging a ∼6 cm3 soil core, respectively. For both biomass and soil, moisture content was determined as mass lost after drying at 60 °C for 48 hr and expressed on a dry-weight basis (mass of water in sample divided by dry mass of the sample, multiplied by 100).

Data Analysis

Prior to each hypothesis test, we assessed model assumptions—specifically, independence of observations and distributions. When high variance and low sample size precluded the application of statistical models to observed data, we employed a moment-matching technique in which we simulated data based on the mean, variance, and distribution of observed data (Ricci 2005, Hobbs and Hooten 2015).

We assessed independence of observations among sample sites within locations with a variance partitioning technique that calculates the proportion of variance attributable to each sampled scale (Winter et al. 2012, McGranahan et al. 2015). When data were to be simulated and a substantial (>10%) proportion of variance arose from the site factor, we used site-specific moments in data simulation; otherwise, location-level mean and variance were used in simulations. When statistical models were applied to observed data without simulation, we used mixed-effect regression models that also fit site as a random variable. All analysis was performed in the R statistical environment (R Development Core Team 2013). Linear mixed-effect regression models were fit using maximum likelihood with the lmer function in the lme4 package (Bates et al. 2013). In addition to model comparison via analysis of deviance, which returns a P-value based on a χ 2 statistic, we estimated 95% confidence intervals from the quantiles of 1000 model simulations (Nakagawa and Cuthill 2007). When factors with more than two levels, such as species, were significant terms in the overall model (P ≤ 0.05), we employed post hoc Tukey comparisons using the glht function in package multcomp (Hothorn et al. 2013).

Fuel curing rates. We simulated a dataset (N = 10) based on the mean and variance of fuel moisture for each species at each location for each sampling event based on a log-normal distribution. Moments were calculated at the location level per species per sampling event as differences among sampling sites accounted for approximately 7% of variance in fuel moisture.

To determine if fuel moisture differed across study locations, we used analysis of deviance to compare two linear mixed-effect regression models: a null, intercept-only model, and one with a location term. For each of these models, log-transformed, simulated fuel moisture values were fit as response variables. Sampling date was included as a random effect to control for the repeated measures component of the sampling design.

To determine if species differed in fuel moisture content within study locations, we used analysis of deviance to compare two linear mixed-effect regression models: a null, intercept-only model, and one with a species term. As above, log-transformed, simulated fuel moisture values were fit as response variables, sampling date was included as a random effect, and models were fit with maximum likelihood using lmer.

Fire behaviour. To predict whether differences in fuel moisture accounted for differences in fire behaviour, we modeled fire intensity for each species at each location for each sampling event. We used a fire intensity equation based on a linear regression model developed from fuel, weather, and fire behaviour characteristics recorded from 200 fires in South African grass-dominated fuelbeds (Trollope and Trollope 2002). The Trollope equation,

$$\begin{array}{*{20}c}{FI = 2729 + 0{.}8684\; \cdot \;{x_1} - 530\; \cdot \quad \quad \quad} \\ {\surd{x_2} - 0{.}1907\; \cdot \;{x_3}^2 - 596\; \cdot \;{1 \over {{x_4}}},} \end{array}$$
(1)

returns fire intensity (FI, kJ sec−1 m−1, converted for presentation here as kW m−1) from fuel load (x 1), percent fuel moisture on a dryweight basis (x 2), relative humidity (x 3), and wind speed (x 4). We used our observed fuel moisture data and set fuel load at 5000 kg ha−1 for both locations based on published primary production values (Everson et al. 1985, Buis et al. 2009), relative humidity at 25%, and wind speed at 3.5 m sec−1. Three fuel moisture scenarios were modeled, with the following sets of inputs for x 2: a low-moisture, species-specific scenario in which we used the second quantile of fuel moisture percentage per species per sampling event; a high-moisture, species-specific scenario in which we used the fourth quantile of fuel moisture percentage per species per sampling event; and an average fuelbed scenario in which we used the mean fuel moisture of all species combined per sampling event.

Fuel moisture and soil moisture. To test for an association between fuel curing and soil moisture across study locations, we fit log10-transformed, observed fuel moisture against soil moisture and location in a linear mixed-effect model with site and date as nested random effects to account for spatial variation and repeated measures using the lmer function, estimated 95% confidence intervals, and ran post hoc pairwise comparisons on species as above.

Results

Fuel Curing Rates

From the end of the growing season through mid-winter, fuel moisture varied among the Highmoor and Ukulinga study locations (χ 2 = 8.51, P = 0.004, 95% CI = 0.11 to 0.51; Figure 1). Although fuel moisture was statistically different among species at both Highmoor (χ 2 = 67.82, P < 0.001) and Ukulinga (χ 2 = 7.14, P = 0.03), the effect was limited to Highmoor (95% CI = 0.11 to 0.23) and did not vary from zero at Ukulinga (95% CI = −0.03 to 0.05) (Figure 2). At Highmoor, post hoc pairwise comparisons indicated that the C3 Festuca had higher fuel moisture than all three C4 grasses—Alloteropsis (P < 0.001), Aristida (P = 0.02), and Themeda (P = 0.02)—while among C4 grasses, Themeda and Aristida had significantly higher fuel moisture than Alloteropsis (P < 0.001 for both). At Ukulinga, post hoc pairwise comparisons provided marginal statistical evidence that Aristida had lower fuel moisture than Alloteropsis (P = 0.05) and Themeda (P = 0.06).

Figure 1
figure 1

Second-order polynomial linear regression lines generalize the relationship between estimated fuel moisture (log10 scale) and time at two grassland locations in South Africa. Analysis of deviance indicated that location was a significant term in a linear mixed-effect regression model (χ 2 = 8.51, P = 0.004, 95% CI = 0.11 to 0.51). Sampling period spans end of the growing season into late winter.

Figure 2
figure 2

Second-order polynomial linear regression lines generalize the relationship between estimated fuel moisture data (log10 scale) and time for four species at two grassland locations in South Africa. Analysis of deviance indicated that species was a significant term in linear mixed-effect regression models for both locations (P < 0.05), but estimated 95% confidence intervals suggest that the effect is limited to Highmoor (0.11 to 0.23) and not Ukulinga (−0.03 to 0.05). Festuca costata occurred at Highmoor only. Sampling period spans end of the growing season into late winter.

Fire Behaviour

Overall, predicted fire intensity increased for all species at both locations as the growing season gave way to the dormant winter season (Figure 3). At the end of the study period, when fuels were considered fully cured (<30% moisture), Ukulinga had consistently greater predicted fire intensity (all species >5000 kW m−1) than Highmoor (all species <5000 kW m−1). At Highmoor, Festuca had the lowest predicted fire intensity after C4 grasses cured (Figure 3).

Figure 3
figure 3

Trollope’s predicted fire intensity over time for three species at two grassland locations in South Africa, plus Festuca costata at Highmoor. Within each species for each sample date, fire intensity was predicted at two levels of fine dead fuel moisture—low and high—calculated as the second and fourth quantiles of observed moisture, respectively. Grey bar in background represents the predicted fire intensity based on the average fine dead fuel moisture of all species combined. Sampling period spans end of the growing season into late winter.

At Ukulinga, the increase in predicted fire intensity as fuels cured was linear and consistent across species, with the exception of a sharp divergence on the 10 May sampling event, which was likely driven by differential responses of fuel moisture to a rain event the previous day. At Highmoor, predicted fire intensity increased later and more slowly than at Ukulinga. Predicted fire intensity varied considerably within and among species at Highmoor due to differences in high and low moisture content, but within-species variation in predicted fire intensity was relatively low at Ukulinga (Figure 3). The average scenario, which predicted fire intensity using mean fuel moisture across all species, had a relatively linear increase over time at Ukulinga, but was later and slower to increase at Highmoor (Figure 3).

Fuel Moisture and Soil Moisture

Analysis of deviance indicated the significance of the regression model that fit fuel moisture against soil moisture and location (χ 2 = 6.95, P = 0.008). Fuel moisture and soil moisture had a positive relationship (95% CI = 0.01 to 0.09), and Ukulinga generally had lower soil moisture than the wetter, more loamy Highmoor (95% CI = 0.39 to 2.60). The slope of the fuel moisture-soil moisture relationship varied substantially among sampling sites at Highmoor, with less spatial variability in slope at Ukulinga (Figure 4). Only at Highmoor was the regression model with species term significantly different from the intercept-only model (χ 2 = 11.76, P = 0.008, 95% CIsoil = 0.01 to 0.08, 95% CIspecies = 0.02 to 0.28), and only Alloteropsis differed from other species (Aristida, P = 0.02; Festuca, P = 0.002; Themeda, P = 0.06).

Figure 4
figure 4

Linear regression lines generalize the relationship between fuel moisture and soil moisture at two sampling locations over four sampling events in two South African grasslands. Analysis of deviance indicated that, over all sites and locations, soil moisture was a significant term in a linear mixed-effect regression model (χ 2 = 3.97, P = 0.05, 95% CI = 0.01 to 0.08). Sampling period spans end of the growing season into late winter. Each symbol represents mean of two samples per sample point, and differences in line and character shape represent spatially discrete sampling points within study locations.

Discussion

Our data on fuel and soil moisture in South African grasslands were characterised by variability from several sources. First, variability among sampling sites and sampling dates within locations highlights the importance of not only replication and sample sizes, but also analytical sensitivity to this variation—accounting for it, not attempting to average it away—through hierarchical sampling and analysis (Twidwell et al. 2009). Secondly, variability among locations and species indicates that broad biogeographic and compositional similarities among plant communities do not adequately describe fuelbed dynamics. Although limited to four sites across two locations in a single year, the temporal breadth and species-level specificity of our study constitute novel data that suggest areas where additional research is needed—or perhaps not needed—to gain a better understanding of fuelbed ecology in grasslands worldwide, beyond the context of extreme fire weather and behaviour.

We provide evidence that fuel moisture has a linear relationship with soil moisture, and found that the nature of the relationship (as characterised by the slope of the regression line) is spatially variable and might be soil-type dependent. While a positive association between soil moisture and live fuel moisture has been described for woody plants in a savannah-scrub type ecosystem (Qi et al. 2012), to our knowledge, ours is the first report on such an association in mesic grassland. The relationship is promising because sampling fuel moisture directly is labour-intensive, but soil moisture can be monitored continuously with electronic sensors and even remotely sensed (Qi et al. 2012). The spatial resolution and temporal replication of our data are too limited to inform accurate predictions, but they do suggest value in additional research on the relationship between fuel moisture and soil moisture, especially given the dependence of fire behaviour on fuel moisture (Govender et al. 2006, McGranahan et al. 2012).

An important ecological function of fire in mesic African rangelands is controlling the encroachment of woody plants, or “bush” (Trollope 1980). Effecting bush mortality—or at least substantial set-back of vigorous re-sprouters—requires relatively high fire intensity, which is a function of several factors including fuel load and moisture (Twidwell et al. 2009). We set fuel load as the average productivity for each grassland and determined that predicted fire intensity varies substantially with fuel moisture at each location. Conventional wisdom places the bush control threshold at approximately 2000 kW m−1 (Trollope and Trollope 2004). While all of our estimated values technically exceeded 2000 kW m−1, obtaining fire intensity substantially greater than the minimal threshold value required one and two and a half months curing time at Ukulinga and Highmoor, respectively.

Species-level variation in fuel moisture and predicted fire intensity at Highmoor might reflect a global syndrome of high-moisture, C3 grasses reducing fire intensity and spread, but more focused research is required. It is difficult to distinguish whether we can attribute higher moisture and lower fire intensity of the C3 Festuca costata in our data to phylogeny or physiology. Patterns in traits among South African grasses were recently attributed to phylogeny rather than different photosynthetic pathways, owing to unique responses among Aristida species compared to other C4 grasses (Ripley et al. 2015) consistent with our data on Aristida junciformis (Figures 2 and 3). Festuca costata’s pattern of higher fuel moisture and lower predicted fire intensity than C4 grasses is reminiscent of Festuca species in other rangelands: in Patagonia, Festuca pallescens (St.-Yves) Parodi has lower flammability than neighbouring grasses (Gonzalez et al. 2015); and in North American tallgrass prairie, the invasive Festuca arundinacea Schreb. (syn. Lolium arundinaceum [Schreb.] S.J. Darbyshire) reduces fire spread (McGranahan et al. 2013). However, the effect of F. arundinacea is specifically attributed to high live fuel moisture during the otherwise dormant period of native C4 grasses (McGranahan et al. 2012), and Trollope’s fire intensity equation does not distinguish between live and dead fuels. F. costata fuel moisture must be measured as live and dead components and modeled as such; the BehavePlus system does so, and although no fuel models have been specifically developed for African rangeland (Trollope and Trollope 2002), recent versions allow users to customise several fuelbed parameters (Andrews et al. 2008).

Ecologically meaningful differences in fire behaviour among species might create positive feedbacks in the plant community that create and increase variability among patches. Disturbance that follows a spatially discrete pattern—such as heterogeneous plant community composition—can reinforce the underlying pattern by creating increasingly contrasting patches internally similar in response and increasingly variable at the landscape level (Kotliar and Wiens 1990, Adler et al. 2001). In these grasslands, emergence of patchiness likely depends on the fuel moisture at which these grasslands are burned: early-season fires conducted when variability in fuel moisture is high are more prone to variable fire intensities, especially at Highmoor, than later-season burns in homogeneously cured stands (Figure 3). But some areas of South Africa seek to increase the variability of burn season with more prescribed ignitions outside of the dry season (e.g., van Wilgen et al. 2004). In the Drakensberg grasslands, F. costata often occurs as discrete patches (D.A. McGranahan, North Dakota State University, Fargo, USA, unpublished data) that could benefit from lower-intensity burns and increased competitive advantage following fire relative to neighbouring C4 species, a classic model of positive feedbacks in fire regime alteration (Brooks et al. 2004).

Variation in fire intensity among grass types and patches might have landscape-level implications that could be exacerbated by global environmental change, but further research into the interactions between plant biology and soil is required to determine the magnitude and direction of these changes. Drakensberg grasslands sit at an altitudinal transition between C4-dominated montane grassland and sub-montane grassland interspersed with Festuca costata (Killick 1963). If the competitive advantages conferred by C4 photosynthesis are considered, C3 species like F. costata might be expected to retreat upslope as local environments become hotter and drier (Edwards et al. 2010). But many Festucoid grasses are infected by fungal endophytes that can increase drought tolerance in the host (Hahn et al. 2008, Gibert and Hazard 2011). F. costata has only recently been identified as an endophyte host, and any ecological effects of the association remain to be determined (McGranahan et al. 2015). Understanding F. costata fire ecology will be enhanced by better knowledge of how F. costata responds to increased air temperature, altered moisture regimes, and atmospheric carbon dioxide fertilization, and how these relationships are affected by endophyte infection.

Our study suggests ways in which grassland fuelbeds might be considered both in current decision making and in plans to mitigate and manage global environmental change. We show that in compositionally similar grassland communities, both environmental factors and fuelbed dynamics can have substantial spatial variation: Ukulinga, a low-elevation location where grasses cured rapidly and consistently under hotter air temperature and a discrete rainy season, presented little variability for managers to consider; whereas the environmental context of Highmoor contributed to slower curing response, differences between species, and association between fuel and soil moisture. While the magnitude of differences in our data do not suggest radical alteration of fire practice and planning, we do hope that fire managers in South Africa and abroad consider heterogeneity within grassland fuelbeds and recognize seasonal changes to ensure that objectives are obtainable prior to burns, and to explain post-burn spatial variation in fire effects within a fuelbed that might have appeared homogeneous.