Results
We first compared the predictions of the three random forest models, one
for each fuel stratum, against the evaluation dataset. The models agreed
well with the field-based observations in the evaluation data sets in
terms of accuracy, no information rate and Kappa coefficient (Table S3).
For all three fuel strata, the models achieved an accuracy over 0.65,
which is much higher than the ‘no information rate’ (0.3-0.4). The
models also had Kappa coefficients over 0.5, indicating good model fits
for all five score values despite the imbalanced sample size. Overall,
the evaluation suggested good and consistent model performance for all
three fuel strata and all five fuel hazard scores.
The random forest models also helped identify the key predictors of fuel
hazard score among the 15 layers of gridded input. The importance of
each predictor, which is measured in terms of the contribution to the
overall model accuracy, is shown in Figure 2. Climate predictors in
general ranked higher than other predictors in all models. However,
solar radiation for January consistently ranked high in all strata,
suggesting the importance of slope gradient and aspect for long-term
climate controls on vegetation productivity and fuel hazard. Rainfall
seasonality is among the most important predictors for the hazard score
of the elevated fuel stratum, indicating the temporal distribution of
rainfall may be more important than the absolute total. Minimum RH is a
climate factor that is important for all three models, but its
importance is much smaller for predicting the hazard score of the
elevated fuel stratum than for the other two fuel hazard scores. The
importance of optimal LAI was intermediate. Plan and profile curvature
as well as clay fraction in the topsoil and available volumetric water
capacity are consistently ranked low in importance for predicting
hazards cores of all strata. For hazard scores of surface and near
surface fuel strata, Tmax and PPT also ranked high in
importance, indicating importance of short-term topoclimatic drivers in
these two strata. It is notable that the importance based on prediction
accuracy and Gini Index are different for all strata. For example, for
the hazard score of the elevated fuel stratum, MAP is ranked high in
importance according to the Gini Index but not when importance is
measured by model accuracy.
After model evaluation, we used the fitted models to predict changes in
fuel hazard scores for each fuel stratum under projected climate change
and increasing Ca. The mean predicted change in the
P4_5, averaged across the nine climate models, showed a
distinct spatial distribution (Figure 3). For mid 21stcentury (2045-2060), the model predicted an increased
P4_5 in the north and southeast of Victoria under both
the RCP4.5 scenario (Figure 3a) and the RCP8.5 scenario (Figure 3 c, d)
with more severe climate change resulting in larger changes. This
pattern remained consistent later in the 21st century
(Figure 2b). The same projection for hazard scores of surface and
near-surface strata are shown in Figures S5 and S6. The changes in the
P4_5 (-0.2, 0.2) in surface and near-surface strata
were much smaller than that of elevated stratum.
We explored the model projection of P4_5 of elevated
fuel stratum along a climate aridity gradient (Figure 4). For current
conditions, the results show a clear decrease in the median probability
of high fuel hazard score in elevated fuel stratum as aridity increases,
reflecting the pattern seen in the input data (Figure 4a) during
2000-2015. For future conditions, the model predicts an increased
P4_5 of elevated stratum in dry region (AI
>3.5) but not in more mesic regions under RCP8.5 by
2085-2100 (Figure 4b).
The strong spatial pattern in predicted change of the probability of
high hazard scores for the elevated fuel stratum encouraged a further
investigation of the predictor of that spatial variation. We first
assessed the impact of changes in rainfall seasonality. Comparing the
‘manipulated run’ (i.e., projections with rainfall seasonality from
2000-2015) to the ‘projection’ showed that the spatial variation in
predictions is driven by rainfall seasonality (Figure 5). With the
manipulated run, the model predicted no change in P4_5of the elevated fuel stratum, with predictions forming a narrow
distribution with a single maximum (blue bars in Figure 5a). With the
actual projections, the model predicted a bimodal distribution with
around half of the pixels having little change in probability while the
other half with a clear increase by up to 0.4 in the
P4_5 of the elevated fuel stratum (red bars in Figure
5; Figure 3d compared with Figure 3a).
We also assessed the impact of changes in CO2, via the
change in optimal LAI. The ‘fertilisation effect’, varied from -9% to
12% in the region under RCP8.5 by 2085-2100 (Figure 5b). Notably, the
‘fertilisation effect’ is larger in regions with current low LAI (Figure
S4g).
We chose a mountainous region with complex topography (Dargo, Australia)
to explore the potential of the model to describe fine scale (90 m)
variation in fuel hazard (Figure 6). Dargo has large small-scale
variations in topography (Figure S4). A model driven solely by coarse
resolution predictors, such as gridded climate, would predict no
variation in fuel hazard score within this region of ca. 4 km x 5 km.
The random forest model with key terrain attributes as predictors,
however, predicted strong variations in the probability of high hazard
scores in the elevated fuel stratum. The P4_5 of the
elevated fuel stratum in the valleys was relatively low compared to
ridges (Figure 6 a,b). Notably, the predicted response of
P4_5 of the elevated fuel stratum to climate change is
not uniform throughout the landscape (Figure 6 c,d). Wet valleys (Dark
green in Figure 6d) saw higher increases in probability of high fuel
hazard scores in elevated fuel stratum during 2085-2100 under RCP8.5
(Figure 6c) compared to the rest of the landscape.