We describe the development and implementation of an operational human-caused wildland fire occurrence prediction (FOP) system in the province of Ontario, Canada. A suite of supervised statistical learning models was developed using more than 50 years of high-resolution data over a 73.8 million ha study area, partitioned into Ontario’s Northwest and Northeast Fire Management Regions. A stratified modelling approach accounts for different seasonal baselines regionally and for a set of communities in the Far North. Response-dependent sampling and modelling techniques using logistic generalized additive models are used to develop a fine-scale, spatiotemporal FOP system with models that include nonlinear relationships with key predictors. These predictors include inter- and intra-annual temporal trends, spatial trends, ecological variables, fuel moisture measures, human land-use characteristics, and a novel measure of human activity. The system produces fine-scale, spatially explicit maps of daily probabilistic human-caused FOP based on locally observed conditions along with point and interval predictions for the expected number of fires in each region. A simulation-based approach for generating the prediction intervals is described. Daily predictions were made available to fire management practitioners through a custom dashboard and integrated into daily regional planning to support detection and fire suppression preparedness needs.

Wildland fire occurrence prediction (FOP) modelling supports fire management decisions, such as suppression resource pre-positioning and the routeing of detection patrols. Common empirical modelling methods for FOP include both model-based (statistical modelling) and algorithmic-based (machine learning) approaches. However, it was recently shown that many machine learning models in FOP literature are not suitable for fire management operations because of overprediction if not properly calibrated to output true probabilities. We present methods for properly calibrating statistical and machine learning models for fine-scale, spatially explicit daily FOP followed by a case-study comparison of human-caused FOP modelling in the Lac La Biche region of Alberta, Canada, using data from 1996 to 2016. Calibrated bagged classification trees, random forests, neural networks, logistic regression models and logistic generalised additive models (GAMs) are compared in order to assess the pros and cons of these approaches when properly calibrated. Results suggest that logistic GAMs can have similar performance to machine learning models for FOP. Hence, we advocate that the pros and cons of different modelling approaches should be discussed with fire management practitioners when determining which models to use operationally because statistical methods are commonly viewed as more interpretable than machine learning methods.

This paper presents the results from employing survival analysis methods to model the probability distribution of the control time of forest fires. The Kaplan–Meier estimator, log–location–scale models, accelerated failure time models, and Cox proportional hazards (PH) models are described. Historical lightning and people-caused forest fire data from the Province of Ontario, Canada from 1989 through 2004 are employed to illustrate the use of the Cox PH model. We demonstrate how this methodology can be used to examine the association between the control time of a suppressed forest fire and local factors such as weather, vegetation and fuel moisture, as well as fire management variables including the response time between when a fire is reported and the initiation of suppression action. Significant covariates common to both the lightning and people-caused models were the size of the fire at the onset of initial attack, the Fine Fuel Moisture Code and the Initial Spread Index. The response time was also a significant predictor for the control time of lightning-caused fires, whereas the Drought Code and time of day of initial attack were significant for people-caused fires. Larger values of the covariates in these models were associated with larger survival probabilities.

Daily, fine-scale spatially explicit wildland fire occurrence prediction (FOP) models can inform fire management decisions. Many different data-driven modelling methods have been used for FOP. Several studies use multiple modelling methods to develop a set of candidate models for the same region, which are then compared against one another to choose a final model. We demonstrate that the methodologies often used for evaluating and comparing FOP models may lead to selecting a model that is ineffective for operational use. With an emphasis on spatially and temporally explicit FOP modelling for daily fire management operations, we outline and discuss several guidelines for evaluating and comparing data-driven FOP models, including choosing a testing dataset, choosing metrics for model evaluation, using temporal and spatial visualisations to assess model performance, recognising the variability in performance metrics, and collaborating with end users to ensure models meet their operational needs. A case study for human-caused FOP in a provincial fire control zone in the Lac La Biche region of Alberta, Canada, using data from 1996 to 2016 demonstrates the importance of following the suggested guidelines. Our findings indicate that many machine learning FOP models in the historical literature are not well suited for fire management operations.

Spring fire activity has increased in parts of Canada, particularly in the west, prompting fire managers to seek indicators of potential activity before the fire season starts. The overwintering adjustment of the Canadian Fire Weather Index System’s Drought Code (DC) is a method to adjust and carry-over the previous season’s drought conditions into the spring and potentially point to what lies ahead. The occurrence of spring fires is most strongly influenced by moisture in fine fuels. We used a zero-inflated Poisson regression model to examine the impact of the previous end of season Drought Code (DC_f) and overwinter precipitation (P_ow) while accounting for the day-to-day variation in fine fuel moisture that drives ignition potential. Impacts of DC_f and P_ow on area burned and fire suppression effectiveness were also explored using linear and logistic regression frameworks. Eight fire management regions across the boreal forests were analyzed using data from 1979 to 2018. For the majority of regions, drier fall conditions resulted in more human-caused spring fires, but not in greater area burned or reduced suppression effectiveness. The influence of P_ow was much more variable pointing to the conclusion that P_ow alone is not a good indicator of spring drought conditions.

Forest organic layers are important soil carbon pools that can, in the absence of disturbance, accumulate to great depths, especially in lowland areas. Across the Canadian boreal forest, fire is the primary disturbance agent, often limiting organic layer accumulation through the direct consumption of these fuels. Organic layer thickness (OLT) and fuel load (OLFL) are common physical attributes used to characterize these layers, especially for wildland fire science. Understanding the drivers and spatial distribution of these attributes is important to improve predictions of fire behaviour, emissions and effects models. We developed maps of OLT and OLFL using machine learning approaches (weighted K-nearest neighbour and random forests) for the forested region of the province of Alberta, Canada (538, 058 km²). The random forests approach was found to be the best approach to model the spatial distribution of these forest floor attributes. A databased of 3, 237 OLT and 594 OLFL plots were used to train the models. The error in our final model, particularly for OLT (5 cm), was relatively close to the variability we would expect to find naturally (3 cm). The dominant tree species was the most important covariate in the models. Age, solar radiation, spatial location, climate variables and surficial geology were also important drivers, although their level of importance varied between tree species and depended on the modelling method that was used.