Fire refugia, sometimes referred to as fire islands, shadows, skips, residuals, or fire remnants, are an important element of the burn mosaic, but we lack a quantitative framework that links observations of fire refugia from different environmental contexts. Here, we develop and test a conceptual model for how predictability of fire refugia varies according to topographic complexity and fire weather conditions. Refugia were quantified as areas unburned or burned at comparatively low severity based on remotely sensed burn severity data. We assessed the relationship between refugia and a suite of terrain‐related explanatory metrics by fitting a collection of boosted regression tree models. The models were developed for seven study fires that burned in conifer‐dominated forested landscapes of the Western Cordillera of Canada between 2001 and 2014. We fit nine models, each for distinct levels of fire weather and terrain ruggedness. Our framework revealed that the predictability and abundance of fire refugia varied among these environmental settings. We observed highest predictability under moderate fire weather conditions and moderate terrain ruggedness (ROC‐AUC = 0.77), and lowest predictability in flatter landscapes and under high fire weather conditions (ROC‐AUC = 0.63–0.68). Catchment slope, local aspect, relative position, topographic wetness, topographic convergence, and local slope all contributed to discriminating where refugia occur but the relative importance of these topographic controls differed among environments. Our framework allows us to characterize the predictability of contemporary fire refugia across multiple environmental settings and provides important insights for ecosystem resilience, wildfire management, conservation planning, and climate change adaptation.

Altered fire regimes can drive major and enduring compositional shifts or losses of forest ecosystems. In western North America, ponderosa pine and dry mixed‐conifer forest types appear increasingly vulnerable to uncharacteristically extensive, high‐severity wildfire. However, unburned or only lightly impacted forest stands that persist within burn mosaics—termed fire refugia—may serve as tree seed sources and promote landscape recovery. We sampled tree regeneration along gradients of fire refugia proximity and density at 686 sites within the perimeters of 12 large wildfires that occurred between 2000 and 2005 in the interior western United States. We used generalized linear mixed‐effects models to elucidate statistical relationships between tree regeneration and refugia pattern, including a new metric that incorporates patch proximity and proportional abundance. These relationships were then used to develop a spatially explicit landscape simulation model. We found that regeneration by ponderosa pine and obligate‐seeding mixed‐conifer tree species assemblages was strongly and positively predicted by refugia proximity and density. Simulation models revealed that for any given proportion of the landscape occupied by refugia, small patches produced greater landscape recovery than large patches. These results highlight the disproportionate importance of small, isolated islands of surviving trees, which may not be detectable with coarse‐scale satellite imagery. Findings also illustrate the interplay between patch‐scale resistance and landscape‐scale resilience: Disturbance‐resistant settings (fire refugia) can entrain resilience (forest regeneration) across the burn matrix. Implications and applications for land managers and conservation practitioners include strategies for the promotion and maintenance of fire refugia as components of resilient forest landscapes.