Predictive models of tree mortality and survival are vital for management planning and
understanding fire effects in forests and woodlands, yet the underlying mechanisms of firecaused
tree mortality remain poorly understood. This shortcoming limits the ability to accurately
predict mortality and develop robust modelling applications, especially under novel future
climates. Our project reviewed the current understanding of the mechanisms of fire-induced tree
mortality, recommended standardized terminology, and described model applications and
limitations. We evaluated accuracy of the fire-induced tree mortality models from the First Order
Fire Effects Model (FOFEM; https://www.firelab.org/project/fofem) software system using a
national dataset we developed. Lastly, we explored if climate data can improve prediction of fire induced
tree mortality and conclude with key knowledge gaps and future directions for research.

We used a post-fire tree mortality dataset built from 40 contributed datasets from across the USA
to formally evaluate the accuracy of fire-induced tree mortality models from the FOFEM
software system. The Fire-Induced Tree Mortality (FITM; https://www.firelab.org/project/fire induced-
tree-mortality) database includes 173,120 tree-level observations of fire injury and
survival or mortality. The database includes 160 tree species from 435 prescribed fires and
wildfires occurring from 1981 to 2016.

Using the FITM database, we evaluated mortality models available in FOFEM, including the
general and species-specific formulations for 45 tree species. These models are also included in
FFE-FVS (https://www.firelab.org/project/ffe-fvs) and BehavePlus
(https://www.frames.gov/behaveplus/home). Of the 69 models evaluated, ~75% of models tested
had excellent or good predictive ability, while 17 had poor performance. The FOFEM5 model
consistently over-predicted angiosperms mortality. For conifers, FOFEM5 over-predicted
mortality for thick-barked species, but under-predicted mortality at low levels of crown scorch
levels with moderate bark thickness. The species-specific models had higher AUCs than
FOFEM5 models for 15 of the 22 models. Poorly performing models were primarily
angiosperms or thin-barked conifers. This suggests that other approaches, such as different
model forms, better bark thickness estimates, and additional predictors, may be warranted for
these taxa.

The project also investigated the addition of climate data to improve model accuracy in
predicting tree death from fire. We evaluated the effect of climatic water deficit (CMD),
summarized over three temporal windows (3-years pre-fire, fire year, and 3-years post-fire) as a
predictor to 11 of the FOFEM models. These models were selected because they had excellent
data quality and model performance. In all cases, CMD significantly improved model
performance, but this did not always translate in a significant improvement in classification
accuracy, based on statistical comparisons of the ACUs.

We suggest a two-pronged approach to future research: (1) continued improvements and
evaluations of empirical models to quantify uncertainty and incorporate new regions and species
and (2) acceleration of basic physiological research on the proximate and ultimate causes of fireinduced
tree mortality to incorporate processes of tree death into models. Advances in both
empirical and process fire-induced tree modelling will allow creation of hybrid models that could
advance understanding of how fire injures and kills trees, while improving prediction accuracy of
fire-driven feedbacks on ecosystems and landscapes, particularly under novel future conditions.