3 Results
3.1 Global and biome distributions of ecoregion flammability
thresholds
Ecoregion flammability thresholds (EFTs) were identified for 772 of the
867 ecoregions classified in the Terrestrial Ecosystems of the World
dataset (Olson et al., 2001), representing a total area of 128.5 M
km2 – about 87% – of the Earth’s terrestrial
surface (Figures 2a and 3). The global mean EFT identified was 12.2%,
with a median of 11.5%, and an interquartile range of 6.92%. Only 17
ecoregions representing 0.8 M km2 reported thresholds
over the expected maximum of 30%; most of these, however, are likely
due to the limitation of fuel moisture playing a smaller role in the
importance of fire for that ecoregion, or artifacts associated with
insufficient fire data. The inflection point slope estimates (IPSEs)
extracted from the P(BA < DFFMC) models shows the
strongest relationship between DFFMC and burnt area appears to be
primarily in different types of desert, tropical and subtropical
savannas, higher-latitude forests, and some select tundra environments
(Figure 2b).
Across biomes, the lowest identified EFTs were associated with types of
shrublands, woodlands, and savanna biomes (Figure 3). The highest
identified EFTs, on the other hand, tended to represent higher-latitude
forests, as well as tundra. The average wetter temperate and tropical
forested biomes closely followed these high-latitude biomes, but also
tended to hold the widest range of EFTs, suggesting more complex
geospatial relationships between fire, fuel moisture, and anthropogenic
stressors. The Kruskal-Wallis rank sum test for differences in EFTs
between biome types was strongly significant (p <
0.001), while the pairwise multiple comparison statistics calculated
using Dunn’s test highlight like and unlike biome types. The statistical
grouping of biomes with similar EFTs shows a clear gradient across
climate, productivity, and vegetation types, albeit with a wide
intra-biome variability associated with the complexity of ecoregions
within biomes (Figure 3). See supporting information for pairwise
multiple comparison statistics, as well as EFT summary statistics across
both biome type and all existing hierarchical realm and biome
classifications.
3.2 Climatic controls on ecoregion flammability
thresholds
The two-dimensional nonmetric multidimensional scaling (NMDS) ordination
reveals the role of eco-climatological factors in effecting the
identified EFT across the different biome types (Figures 4a-b). The
first dimension is primarily driven by precipitation and associated
proxies: Higher annual precipitation, net primary productivity (NPP),
and more tree cover are interlinked, influencing fuel-fire relationships
in moisture-limited or -driven ecoregions. The second scaled dimension
is driven primarily and negatively by temperature and is strongly linked
to moisture-limited ecoregions with lower EFTs. Precipitation
seasonality has a negative effect across both the first and second
scaled dimensions, with higher values of precipitation seasonality
representing stronger seasonal dryness (e.g., deserts and savannas with
arid summer conditions). Similarly, higher herbaceous vegetation cover
is associated with cooler, drier ecoregions. Outside of higher-latitude
tundra and forest types, Earth’s biome types on average hold lower EFTs
driven by temperature and precipitation on the second NMDS axis (Figure
4b).
The average EFT is well below the expected 30% limit for carbon fibre
saturation across the full range of the productivity-fire activity index
gradient (Figure 5a), noting that the majority of the Earth’s ecoregions
by areas have a NPP below 10 t C ha-1year-1 (Figure 5b). Within the intermediate NPP range
(i.e., roughly 1.2 – 5.4 t C ha-1year-1 per year for the evaluated ecoregions) where
fire activity indices are maximal, EFTs are moderately higher than on
either side of this productivity range or in those ecoregions with less
fire activity. The lowest EFTs occur within the least productive
environments across the entire range of fire activity. Low EFTs also
occur to the right of the intermediate NPP range within those ecoregions
that are frequently burnt. Both these areas of extremely low EFTs also
tend to have the sharpest associated IPSEs (Figure 5c). There ultimately
exists strong relationships between DFFMC and burnt area across most of
the Earth’s ecoregions, apart from those highly productive, wet tropical
forests, (Figure5b).
Our application of generalized additive modelling of EFT against
environmental predictors showed a strong effect of the three bioclimatic
indicators, with less explanatory power for biogeographic and ecological
variables. This final model reported a correlation coefficient of
determination (R2 ) of 0.682 and a
root-mean-square error (RMSE ) of 0.038 for the complete
distribution of EFTs. See supporting information for comparable
statistics for training, testing, and cross-validation datasets. Our
analysis of variable importance shows that RMSE dropout loss was
sensible between predictor variables (supporting information). The three
nonlinear bioclimatic indicators representing mean annual precipitation,
mean annual temperature, and mean precipitation seasonality report the
greatest influence on the model’s accuracy, followed distantly by the
terrestrial realm and biome classifications as random effects. Median
percent herbaceous cover reported a consistent, negligible effect on the
model RMSE loss. The effects statistics confirm this, showing that
precipitation is the strongest predictor of the global EFT distribution.
In contrast to the variable importance analysis, the effect of
precipitation seasonality on EFTs is larger and more significant than
temperature (Table 1). See supporting information for model evaluation
steps including conditional effects and both the predicted and expected
posterior distributions.