The Duff Moisture Code and the limits of sustainable combustion: examining the evidence for a widely used threshold

Lightning ignition

We assembled datasets of lightning strikes, daily fire weather station observations and lightning fire occurrence for the provinces of British Columbia, Alberta, Saskatchewan, Manitoba and Ontario for the years 2000–2019. These locations were chosen as they represented a good span of the Canadian forest with very significant lightning fire occurrence and were areas with datasets we had available because of ongoing applied research on lightning fire occurrence modelling.

Lightning flash data were from a Canadian Forest Service (CFS) research dataset from the Canadian Lightning Detector Network (CLDN: Kochtubajda and Burrows 2020). These lightning flash data included attributes such as each strike’s location, time of occurrence, multiplicity and average discharge strength. We used multiplicity (which represents the number of individual discharges (strokes) associated with a lightning flash) to summarise the data as lightning stroke counts.

Daily fire weather data were assembled from various agency archives and from Environment Canada records. These daily fire weather data observations included calculated FWI System outputs at each station throughout each fire season (Van Wagner 1987).

Fire occurrence records were assembled from individual fire agencies. These records contained information on the estimated start location of each reported fire, its general cause and the day it was reported to the fire agency. We used only lightning fires. An estimated start date for a fire is sometimes provided in these records. However, Wotton and Martell (2005) found that this start date estimate was unreliable; on ~50% of the estimated start days, there was no lightning active within a 25 km radius of the ignition location. We used the lightning flash data to estimate the likely start date at each fire location. For a 10 km radius around each fire ignition location, lighting stroke counts were summarised for each day, moving backwards in time from the fire report date for 21 days. The first day on or prior to the report date for which there was lightning activity within a 10 km radius of the fire location was assumed to be the start date. In previous lightning fire occurrence modelling work (Wotton and Martell 2005), this was found to provide a better start date estimate than looking for the closest strike. If no lightning was found in the 21 days prior to the report date, the fire was excluded from our analysis.

In provinces where the area under fire management spans major ecological zones, we summarised data and modelled ignition probability separately within these regions (Ecological Stratification Working Group (ESWG) 1996). For provinces with a ‘modified response zone’ (where fires that occur are often intentionally left unsuppressed), we split provinces by these general boundaries as well; in a ‘modified response zone’, detection effort may be lower (Fig. 1). For this analysis, we excluded fires from these modified response zones.

Fig. 1.

Regions used in the lightning fire ignition analysis. Each province is divided into its major ecological zones (ECWG 1995) and further divided to exclude the portions of modified response zones in the north of some provinces. The delineations of the fire management boundaries shown here are only simple approximations from fire management strategies circa 2000. We use them here to avoid large areas where detection efforts may be significantly different. In practice, such boundaries are not hard lines on a map – response decisions can vary on a fire-by-fire basis; therefore, this map should not be used as a reference for fire management zone boundaries in Canada.

The base daily unit of analysis was a 20 by 20 km cell to account for potential inaccuracies in lightning and fire location (Wotton and Martell 2005). DMC values from the weather stations were interpolated using thin-plate cubic splines to the centre of each grid cell for each day. To capture the core of the fire season in Canada, we only included data from May through August. To demonstrate the effect of season, we also broke this full fire season into a spring period (May only) and a summer period (June through August). Because fire management policies and procedures can vary from province to province, we analysed data from each province separately. After first exploring data visually, we used shape-constrained generalised additive modelling (Wood 2017) to examine the probability of ignition from a single lightning stroke as a function of DMC. Shape-constrained generalised additive models can ensure that relationships between the predictor and response variable follow specific patterns. In this case, it was used to ensure the relationship between lightning ignition and fuel moisture was physically logical; the probability of ignition can only increase as the fuel moisture level decreases. Applying the shape constraints allowed us to reduce model overfitting from the variability in observations. All modelling was performed using the ‘scam’ package (Pya and Wood 2015) in R statistical software (R Core Team 2024).

Forest floor consumption

The CFS, in partnership with different fire management agencies, has continued to carry out large-plot fire behaviour experimentation since the work Van Wagner (1972) summarised. These experimental burning observations have formed the core of the fire behaviour models in the Canadian Forest Fire Behaviour Prediction (FBP) System (Forestry Canada Fire Danger Group (FCFDG) 1992). For the present analysis, we selected datasets from well-documented Canadian burning experiments included in the CFS Fire Behaviour dataset. Only experimental burning projects conducted in fuels with a pine overstorey were included; our goal was to be consistent with the general fuel type in Van Wagner’s original study. The experimental burning projects were: the Darwin Lake Project carried out in mature pine in northern Alberta (Quintilio et al. 1977); the Kenshoe Lake Project carried out in mature pine near White River, Ontario (Stocks 1989); the Sharpsand Creek Project in immature pine north of Thessalon, Ontario (Stocks 1987), and the International Crown Fire Modelling Experiment, carried out in mature pine north of Fort Providence, Northwest Territories (Stocks et al. 2004). Forest floor consumption observations and DMC values from the burn day were analysed similarly to the data presented in Van Wagner (1972, fig. 2 in that report).

Lightning ignition

Table 1 shows an overall summary of lightning, fire and the occurrence of days above and below the existing DMC threshold of 20 for the portion of each ecozone within each province for the 2000–2019 study period. Overall, 35,343 lightning fires were included across the entire dataset. The percentage of lightning fires occurring at DMC values below 20 (Table 1), while never a majority of fires, reveals that a large number of fires occur at levels below the existing DMC = 20 threshold.

Fig. 2 shows the probability of ignition of a sustainable fire by an individual lightning stroke as a function of DMC from the generalised additive modelling. The term ‘sustainable fire’ is used here to emphasise that our dataset, by definition, contains only fires detected and reported to the agency. Lightning ignitions that self-extinguish before they are ever detected are not included here, nor can their numbers currently be estimated in any reliable way. The data points shown in Fig. 2 are for visualisation and represent empirical estimates of the probability of ignition created by summarising fires and lightning strokes from the observed data into 40 bins, each representing one 2.5 percentile segment of the data (ordered by DMC). The probability of ignition from a lightning stroke can be seen to rise systematically, though non-linearly, with increasing DMC value; there is little indication of a change in the rate of change of the relationship around the DMC value of 20 across the various zones.

Fig. 2.

Empirical probability estimates and general additive models (GAMs) for each study region: (a) British Columbia, (b) Alberta, (c) Saskatchewan, (d) Manitoba, (e) Ontario. The shaded regions represent the 95% confidence interval from the fitted GAM. The points represent the empirical proportion of observed lightning-caused fires binned into 2.5% intervals of DMC (i.e. partitioning DMC into 40 equal categories). The grey dashed line indicates the currently used DMC of 20 threshold.

The positive and increasing relationship between DMC and the probability of lightning fire ignition is similar through the intensively managed portions of most of the boreal forest and taiga plains regions studied here (Fig. 2), except for the Northwest region of Ontario (Fig. 2e), where the rate of rise in probability is larger. The Montane Cordillera and Pacific Maritime ecozones in BC provide an interesting contrast to the boreal/taiga zones; the rate of increase in ignition probability with DMC is higher in these regions than seen through most of the boreal forest. The probability data for the Montane Cordillera in BC (in Fig. 2a) suggest that, in this region that regularly experiences high DMC values and a lot of lightning fire activity, the increase in the probability of ignition as DMC reaches very high values continues approximately linearly to the extremely dry end of the range of duff moisture. This may provide helpful guidance for modellers and fire managers extrapolating these empirical probability relationships in boreal forest regions beyond the historical range of DMC as the climate changes.

Although the majority of lightning fires in Canada occur in the summer, we also explored the ignition probability–DMC relationship for the spring and summer. Fig. 3 shows both spring and summer models for several of the regions we studied. The full-season model (the black line in Fig. 3) is very similar to the summertime model because the overwhelming majority of data (both fire and lightning) come from the summer period (e.g. the spring period has 7% the number of fires and 3.5% the lightning of the summer period). This much smaller spring period sample size leads to less consistent (and we suggest much less reliable) models of the ignition process, reflected somewhat in the wider confidence bands in the Fig. 3 spring models. There is some inconsistency across the country. For the same DMC value in the western provinces, ignition probability tends to be higher in the spring than summertime values; in Ontario, the opposite is observed. This seasonal exploration (Fig. 3), however, reveals no additional indication of a threshold around a DMC value of 20.

Fig. 3.

The probability of ignition per lightning stroke from generalised additive modelling for spring and summer periods and for selected regions in our study area: (a) Alberta, (b) Manitoba, (c) northwestern Ontario, and (d) northeastern Ontario. The blue line and associated shaded 95% confidence bands are from the spring data; the red line and associated shaded 95% confidence bands are for the summer period. The solid black line represents the full season (May through August) model from Fig. 2. The grey dashed line indicates the currently used DMC of 20 threshold.

Forest floor consumption

Fig. 4 shows forest floor consumption and DMC values summarised for a series of experimental burning projects carried out in Canada over the last several decades. The observations from Van Wagner’s experimental burns through the 1960s (shown in his original work as fig. 2 in Van Wagner 1972) indeed suggest a threshold for forest floor consumption at a DMC of just over 20 (Fig. 4a); however, none of the data collected on subsequent, similar experimental burning projects show a similar relationship (Fig. 4b–e). A trend of increasing forest floor consumption with increasing DMC value is typically discernable; however, no meaningful or consistent intercept value where forest floor consumption is effectively negligible can be seen. The plotted results from Van Wagner (1972), which have been adopted in recent decades as ‘evidence’ of this threshold for forest floor consumption, seem to be more of an exception than a consistent observation of a meaningful threshold.

Fig. 4.

Forest floor consumption (kg/m²) observations from experimental fires associated with DMC values: (a) a replotting of data from all fires originally studied in Van Wagner (1972); (b) Darwin Lake Project; (c) Kenshoe Lake Project; (d) Sharpsand Creek Project; (e) International Crown Fire Modelling Experiment (ICFME).