Wildland fires have a multitude of ecological effects in forests, woodlands, and savannas across the globe. Predicting and understand tree mortality from fire has been a major focus of past research, as trees provide myriad biological and economic services. We assembled a database of individual-tree records from prescribed fires and wildfires in the United States. The Fire and Tree Mortality (FTM) database includes records from 159,660 individual trees with records of fire injury (crown scorch, bole char, etc.), tree diameter, insect attack (when available), and mortality up to ten years post-fire. Data span 146 species, from 435 fires occurring from 1981-2016. First, we evaluated the performance of species-level empirical post-fire tree mortality models used in the First Order Fire Effects Model (FOFEM) software system. Second, we examined how taxa and species’ traits explain patterns of performance of existing models, and may provide a guide to determining which variables best predict mortality. Third, we look at spatial variation in the accuracy of existing models, and test if new models that include spatial covariates—first/second fires, method of assessing canopy damage, subspecies, drought stress, and long-term climate—improve mortality predictions. We use logistic regression and machine-learning methods to build predictive mortality models.
Results/Conclusions
Our FOFEM model assessment covered 45 tree species, using data from 96,278 trees, 366 fires, and 34 datasets. Approximately 75% of models tested had either excellent or good predictive ability; of the 69 models evaluated, 42 had AUC values indicating excellent performance. The basic model in FOFEM, the Ryan and Amman (“R-A”) often over-predicted mortality for angiosperms. For conifers, R-A over-predicted mortality for thin-barked species and for small diameter trees. These results, and ongoing development of new machine-learning models, indicated additional predictors, selected using species’ traits, can improve model accuracy and yield ecological insights. Model error of the FOFEM models varied between fires, partially because of lower accuracy when crown volume scorch was calculated from other variables, as opposed to measured in the field, and because of lower accuracy in second-entry fires. We examined spatial variables for two widely-distributed conifer species: Pinus ponderosa (Ponderosa pine, n=43,029), and Pseudotsuga menziesii (Douglas-fir, n=14,681), but found that co-variance between multiple spatial predictors make interpreting the influence of each spatial predictor challenging. In-progress modeling for those two species, using well-balanced subsets of data with strong controls for methodological differences may yield additional insights into spatially variable predictors, including subspecies, drought stress, and site productivity.