Providing a benchmark for process-based models
Although many process-based models have a fire component, the evaluation of those models focused on the carbon and water components (Luo et al., 2012; Hantson et al., 2020). Existing evaluations of the fire modules relied on satellite derived burned area (e.g., Arora and Boer 2005; Hantson et al., 2020). Our predicted fuel hazard represented the potential to use machine learning approaches to upscale field-based observations to the resolution of land surface models and thus could be used as a reference for benchmarking process-based models alternative to satellite-derived burned areas.
Current process-based models run on coarse horizontal grids (Eyring et al., 2016; Friedlingstein et al., 2022) which cannot capture the fine scale variation of fuel shown in field observations. Although fine spatial resolution simulations accounting for fuel variation are possible at regional scales (Fiddes et al., 2022), the computational demand prevents such implementation at the regional and global scales. Our machine learning approach could be used in hybrid modelling framework to improve the model behaviour at fine spatial resolution (Reichstein et al., 2019).