From Background to Benchmark: A Framework for Preserving Spatial Structure in Wildfire Occurrence Modeling

Wildfires represent a growing global challenge, with increasing impacts on ecosystems, air quality, infrastructure, and human safety. In Canada, wildfire activity has intensified in both frequency and severity over recent decades, underscoring the need for robust spatial analyses to better understand the conditions under which fires escape initial suppression. Numerous studies have leveraged advances in artificial intelligence and machine learning to model wildfire occurrence and behavior. Many of these approaches rely on event–non-event (case–control) study designs, where fire locations are contrasted with non-fire locations to identify controlling environmental and anthropogenic factors. While fire event locations are generally well defined in historical records, selecting non-event (non-fire) locations remains a critical and often under-addressed challenge. Existing studies have employed a range of strategies to define non-fire points, including random sampling, uniform grids, distance-based buffers, environmental stratification, and background sampling. Poorly defined non-event locations can introduce substantial spatial bias, distort background conditions, and ultimately undermine model inference and interpretation. In wildfire applications, non-fire locations must satisfy multiple constraints: they should be accessible to fire occurrence, respect land–water and administrative boundaries, and reproduce the spatial structure of observed fire patterns without clustering too close to fire events or dispersing into ecologically irrelevant regions. To address this issue, we propose a two-stage methodological framework specifically designed for wildfire case–control studies, demonstrated using escaped wildfires in Quebec, Canada.
In the first stage, six background-pool (BP) generation strategies were developed to create large sets of geographically plausible non-fire candidates. These strategies progressively incorporate wildfire-relevant constraints, including minimum distance buffers around escaped fires, land–lake masking, grid-based spatial stratification, density weighting, and explicit enforcement of the Quebec boundary. The final background-pool version integrates all constraints and introduces a hybrid distance-based acceptance scheme that combines a strict exclusion zone near fires with a smooth distance-decay function beyond this threshold.
In the second stage, five control-set (CS) selection methods were evaluated to construct 1:1-matched fire–non-fire datasets across multiple fire-size thresholds. The final method balances regional representation and spatial clustering by using an adaptive grid and a composite distance metric that accounts for both proximity to individual fires and distance to local fire centroids. This approach explicitly matches the spatial “clumpiness” of escaped wildfires rather than simply maximizing separation between events and controls.
Model performance was assessed using distance-based diagnostics, spatial variance metrics, and point-pattern validation based on Ripley’s K-function. The proposed framework consistently produced non-fire patterns that are statistically indistinguishable from observed escaped wildfire patterns. Overall, this study provides a transparent, wildfire-specific template for selecting non-event locations, thereby supporting more reliable spatial inference in wildfire risk assessment and fire behavior modeling.