Predictors of fuel hazard
Based on our analysis we recommend using a combination of climate, fine scale topographic attributes and plant response to climate change for fuel projections. We found CO2 fertilisation to contribute up to 12% of P4_5 in Victoria (Figure 4). Empirical models that do not include rainfall seasonality are unlikely to capture the actual change of fuel hazard score under future climate in Victoria where rainfall seasonality is a key driver of plant productivity (Figure 5) despite the optimal LAI ranked only in the middle in the importance list (Figure 2). Although this finding does not directly apply to other regions, the approach is generalisable and could help extract knowledge from field-based observations across the world. In contrast, land surface models running on coarse resolution (>5 km) generally cannot resolve terrain-driven variation in plant growth (e.g., Clark et al., 2016; Wu et al. 2022; but see Fiddes et al., 2022) and are unable to capture fine-scale variations in fuel hazard scores (Figure 6). Although spatial resolution might not be critical in capturing global trends in fire risk, fine scale predictions are crucial for operational fire management on regional scales (Bale et al., 1998; Nyman et al., 2015; Inbar et al., 2018).
Despite the good overall predictive performance, this machine learning approach has four major limitations: 1. It cannot provide information about the transient response of vegetation structural change due to gradual climate change; 2. It does not mechanistically model the impacts of plant composition and range shifts on fuel structure; 3. The ML models do not explicitly consider climate driven changes in fire regimes and the associated feedbacks with vegetation, which potentially affects vegetation distribution and fuel accumulation (e.g., Murphy and Bowman, 2012). 4. Human activities (e.g., land use change and fire hazard reduction efforts) are not included in the models but could result in substantial changes in future fire characteristics (e.g., Wu et al., 2022). Our approach aimed to quantify the potential for changes in fuel hazard as set by environmental and biological constrains. The shortcomings of this study could be addressed by process-based models which require significant developments in the computational capacity and the understanding of climate-vegetation-fire interactions.