Introduction

Recent catastrophic fires around the world have drawn attention to the need for improved fire risk assessments (Bowman et al., 2020; Duane et al., 2021). The likelihood of fire in terrestrial ecosystems is a function of: (i) the fuel hazard (i.e., the amount, density and three-dimensional distribution of plant biomass, both dead and alive), (ii) fuel dryness, (iii) the weather conditions, and (iv) the availability of ignition sources (Bradstock, 2010; Boer et al., 2017). Existing fire behaviour models can capture the impacts of weather on fire spread rate and intensity but require spatially explicit information about a range of fuel attributes as input (Tolhurst et al., 2008). Although current fuel hazard can be mapped using a combination of ground-based and remote-sensed observations (Pierce et al., 2012), quantification of future changes in these patterns in response to climate change requires predictive models.
Fuel hazard is closely related to variation in the composition and structure of the vegetation, which in turn are shaped by plant responses to long-term environmental conditions and disturbance regimes (Kelley, et al., 2019). Consequently, predictions of future fuel hazard need to incorporate the potential impacts of climate change. There are two major ways that climate change affects fuel hazard. First, the rising atmospheric CO2 concentration (Ca) fertilises plants via an enhancement of photosynthesis (Ainsworth and Rogers, 2007), potentially resulting in an increase in plant biomass (Ainsworth and Long, 2005; Norby et al., 2005; Zhu et al., 2016; Walker et al., 2019). Any increase in plant biomass is likely to result in higher fuel loads, but the magnitude of change and how it will interact with other environmental factors remains uncertain (Bradstock, 2010). Second, rising air temperatures and altered rainfall patterns have distinct effects on plant productivity and species composition, both of which could lead to altered fuel hazard (Archibald et al., 2013). It is thus critical to account for plant responses to climate change when projecting future fuel hazard.
Changes in plant biomass under future climate can be predicted with a range of modelling approaches, which have been used to estimate fuel loads. For example, Clarke et al. (2016) projected future fuel loads using the net primary productivity (NPP) predictions from a land surface model, assuming a linear relationship between NPP and fuel load. However, existing evidence suggests that fire regimes (i.e., fire frequency, intensity, season, type, and extent) could varied within a biome with similar NPP - a single biome could have more than one fire regime while the same fire regime can be observed in different biomes (Archibald et al., 2013). Consequently, predictions of the spatial variation of fuel hazard under climate change need to be constrained by and evaluated against the fuel hazard observations under different environmental controls (i.e., climate, soil and topography).
Ground-based fuel observations have been routinely collected by fire management agencies in Victoria, Australia since 1995 (e.g., Hines et al. 2010), and have provided valuable insight on the spatial variation of fuel hazard at landscape to regional scales (e.g., Jenkins et al., 2020; McColl‐Gausden et al., 2020). At each survey site, an ordinal score is assigned to each fuel stratum based on visual estimates of fuel hazard. The potential of these fine-resolution data to help inform process-based model predictions of future fuel hazard remains under-utilised. Assuming the spatial variation of fuel hazard along climatic gradients is indicative of how fuel hazard may change with climate over time (i.e., space-for-time substitution; Picket, 1989), these ground-based fuel surveys contain possibly the best information about the potential for changes in fuel hazard across Victoria in response to projected climate conditions.
Random forest models have been used to synthesise field-based fire observations and environmental drivers with demonstrated success (Pierce et al., 2012; Jenkins et al., 2017; McColl‐Gausden et al., 2020). However, previous machine learning approaches have generally ignored plant responses to climate change (e.g., McColl‐Gausden et al., 2020), due to the this ‘space-for-time’ approach needing additional process-based information on vegetation responses to novel climate.
The past developments of empirical approaches thus exposed the limitation of pure statistical analysis and advocate novel ways to combine strengths of process-based plant biomass predictions with data-driven approaches (e.g., Jenkins et al., 2020). Yang et al. (2018) modelled the change of leaf area index (LAI; an indicator of plant foliage biomass) under changing climate and rising Ca. Incorporating this LAI model and random forest models could help address the lack of plant responses to future climate in current machine learning frameworks. This combined framework could unite the strength of fine-resolution empirical observations and the process-based plant responses to climate changes, addressing the weakness of previous regression and process-based models.
Here, we used random forest models to predict spatial variation in fuel hazard at fine spatial resolution across the state of Victoria as a function of climate, soil and topographic attributes as well as modelled plant responses to climate change. The goal was to assess the potential of change in fuel hazard in response to projected future climate conditions and Ca. Although the training and evaluation of the models focused on a specific region, the methods and conclusions built a quantitative understanding of anthropogenic impacts on future fuel hazard, which is applicable to similar regions across the world.