By controlling structural and compositional attributes, fire is the main disturbance shaping the North American boreal forest [1
]. Fires affect the forest’s structure by creating a mosaic of stands of different ages and sizes [3
], thus constantly rejuvenating stands and landscapes. Fire cycles, defined as the time required to burn an area equivalent to that of the study area [5
], determine the age structure of forest stands [7
] across the landscape. Fires also influence stands’ composition by controlling succession patterns, for instance, by favoring fire-adapted species such as jack pine (Pinus banksiana
) or black spruce (Picea mariana
]. Fire regimes are highly variable in space as a result of various environmental factors acting on different scales [12
]. Climate acts as a top-down factor from regional to continental scales. In Canada, for instance, the increasing gradient of fire activity observed from east to west is caused by the spatial variability in the frequency of drought events [15
]. However, topography [14
], surficial deposits and drainage [18
], or fuel type and availability [19
] are bottom-up factors which act from stand to regional scales. Fire regimes also vary in time; for example, the end of the Little Ice Age that occurred around 1850 represents a well-known transition to lower fire cycles in eastern Canada [8
]. Temporal variations in fire activity are mainly driven by climatic factors such as shifting air masses responsible for dry conditions [20
The coniferous boreal forest of Quebec, eastern Canada, experiences a gradient of dense, continuous forests to the south that transition to discontinuous, less productive forests [23
], and finally to the forest tundra in the north [2
]. The northern open forests are mainly constituted of lichen woodlands resulting from numerous factors such as limited post-fire regeneration due to low seed production [25
], unfavorable climate, short intervals between successive fires [26
], and high severity of large fires [30
Transition ecosystems are known to be extremely vulnerable to climate change [31
], and particularly so for the boreal forest where fire activity is expected to increase [32
]. Because the opening of these forests is closely related to fire activity, studying their fire regime is crucial. In Quebec, there is evidence that current fire activity is higher in northern discontinuous forests than in the commercial boreal forest further south [24
]. However, it is not clear whether climate is responsible for this latitudinal gradient, or if the underlying climate factors are constant.
Moreover, from a forest management perspective, it is important to understand how fire regimes vary depending on the latitude. In Quebec, northern discontinuous forests are protected from commercial forest harvesting by the legal limit of the commercial forest. It is thought that forest management could worsen the problem of regeneration failures at high northern latitudes under climatic influence. However, the spatial and temporal variability of fire regimes along the latitudinal gradient is still poorly known. The zonation of fire activity is also of interest, as zones with high annual area burned can jeopardize forests’ post-fire recovery [29
], although recent studies have shown that the boreal forest could express a certain resistance toward high burn rates [33
]. Learning more about the spatial consistency of high fire risk zones is particularly important in the context of the northern limit of the commercial forest because their expansion to the south could lead to a reduction in the area available for forest management.
The objective of this research is to assess the latitudinal variability of fire regime at the transition between continuous and discontinuous coniferous boreal forests in Quebec over the last 150–300 years, and its relation to climatic conditions. Even if a zonation of fire activity had been developed in previous studies based on current fire regimes [24
], the spatiotemporal consistency of the fire zones over a longer temporal scale has not been explored. The first step of this study was to reconstruct the fire history along four north-south roads almost equally distributed over the black spruce forest of Quebec, using fire archives and dendroecological surveys. We used survival models to assess whether climate was influencing fire risk—defined here as the relative hazard of burning compared with the road average—along each road. The latitudinal distribution of fire risk was then used to delimit homogeneous fire risk zones for each road. Then, the fire cycle of each fire risk zone was calculated allowing for an assessment of fire risk variability along the longitudinal gradient. Finally, fire cycles were compared to previous estimates based on the recent fire history in order to highlight the temporal variability of the fire regime in the study area.
2. Materials and Methods
2.1. Study Area
The study area is located in the boreal zone of Quebec and lies between latitudes 49.5° N and 53° N. The region is mainly coniferous and dominated by black spruce. It covers a gradient from closed, dense forests in the spruce-moss domain to the south, to more open and fragmented forests in the spruce-lichen domain to the north (Figure 1
b). The limit of the commercial forest crosses the study area separating managed forests in the south from unmanaged ones in the north.
In order to cover the latitudinal gradient and because access is difficult in the north of the study area, four north-south roads that are almost evenly distributed from west to east were chosen as a means in which to reach forest stands and served as a basis for our sampling design. Those roads were divided into consecutive 2500 ha-cells (5 km long by 5 km wide) and will hereafter be referred to as transects (Figure 1
and Figure 2
The transects are located in the James Bay region (A), the Chibougamau area (B), the North Shore (C), and along the Romaine River (D). They are 81, 94, 83, and 34-cell-long, respectively. Transect D comprises less cells because of a shorter road. However, 28 cells were added to the original dataset using plots previously sampled for the Northern Forest Inventory Program of the Ministère des Forêts, de la Faune et des Parcs (MFFP) (NFIP; 2005 to 2009). These additional cells are located at a maximum distance of 45 km on either side of the transect.
2.2. Time-Since-Fire Data
As the study area is remote and not easily accessible, we developed a data collection strategy which attempts to maximize the use of fire archive data available for the area between 1924 and 2014 and to complement it with field sampling.
2.2.1. Fire Archive Data (1924–2014)
First, fire archives for the 1924–2014 period obtained from the MFFP were used to reconstruct the most recent fire history (Figure A1
). However, this database may gradually become less complete or lose dating precision with time, particularly towards the northern areas, as it becomes more likely that some fires were either overlapped by most recent fires, or not detected, reported, and archived. The database precision is excellent after 1972 [24
], very good after 1940, but some small fires can be missing or not perfectly delimited between 1924 and 1940. In some cases—mainly in the north and for the oldest fires recorded—the fire date is noted to the nearest five- to ten year interval. For those fires, the middle year of the range was used in the analyses.
Using ArcGIS 10.0, fires at least partially overlapping a cell were identified for each transect. The corresponding time-since-fire (TSF) was then assigned to the cells regardless of the relative importance of the area burned in the cell. When more than one fire partly overlapped a cell, only the most recent TSF was kept. This last situation concerned mostly areas where fires are recurrent (Figure 2
2.2.2. Field Sampling Design
The field campaign took place in 2013 for transects A, B, and C, and in 2014 for transect D. In the north of transect A, from the top of the transect to Broadback River, the TSF data from Héon et al. and Erni et al. [33
] collected along a 200 km long transect was used. The data was adapted to our study design and sampling effort by rescaling their original 2 km by 1 km sampling cells.
In each cell where no TSF was assigned from the fire archives, a point corresponding to a sampling plot was randomly generated from 100 m to 750 m on either side of the road. We assumed that those random points are representative of their corresponding cells in terms of TSF because in our study area, the mean fire size—calculated over the 1924–2014 period—is 5200 ha while our cells are 2500 ha. For this reason, a random point is very likely to capture the most representative fire that occurred in a cell.
In our effort to compare our results with those of Gauthier et al. [24
] who studied the fire regime between 1972 and 2009, all cells that burned between 1972 and 2014 needed to be assigned with a pre-1972 TSF. Therefore, an additional point was sampled outside the polygon of the post-1972 fire in those cells if no pre-1972 fire date was recorded in the fire archives.
For each plot, 10-dominant trees were sampled (section or core taken as close to the ground as possible), with priority given first to jack pine (Pinus blanksiana), and then to black spruce (Picea mariana), paper birch (Betula papyfera), trembling aspen (Populus tremuloides) and, lastly, to balsam fir (Abies balsamea). This order of priority was determined based on the rapidity of these species to regenerate after a fire in order to better approximate the real TSF. The trees’ age was determined by counting their annual growth rings. In a given cell, if all trees were the same age plus or minus 20 years, the TSF of the cell was considered equal to the age of the oldest tree. Otherwise, the age of the oldest tree was considered a minimum TSF, and therefore right-censored. Because post-fire trees are eventually replaced by new ones as succession proceeds, assessing the exact TSF becomes more difficult as a stand ages.
2.2.3. Relative Importance of Recent Fires
Road layout depends on different physical attributes of the landscape as they are usually built on specific surficial deposits. They are also more used by humans than the rest of the landscape. Both these particularities of our transect roads, which have all been built after 1924, can bias the TSF distribution toward more recent fires, thereby affecting estimations of burn rate. Moreover, in some areas, fires can be rare and yet very large. When roads cross such large fires, it can lead to an overestimation of the number of burned cells compared with the average burn rate of the corresponding region. For all those reasons, we computed for each transect the proportion of recent fires (after 1924) in the cells as well as in a 45 km-wide buffer around the transect. Because recent fires were slightly overestimated in all transects comparatively to the 45 km-wide buffers, the cells that burned after 1924 were down-weighted in the analyses. Weights were calculated in each transect in order to match the relative frequency of recent fires with the relative area recently burned in the surrounding landscape. Weights of cells burned between 1924 and 2014 of transects A, B, C, and D are 0.57, 0.41, 0.36, and 0.46, respectively; while all other cells were assigned a weight of 1.
2.3. Climate Data
The Fire Weather Index (FWI) System consists of six indices derived from meteorological observations—namely temperature, relative humidity, wind speed, and 24-h rainfall—which provide numeric ratings of relative potential for wildland fire. The Fine Fuel Moisture Code (FFMC), the Duff Moisture Code (DMC), and the Drought Code (DC) constitute the fuel moisture codes, and the Initial Spread Index (ISI), the Buildup Index (BUI), and the Fire Weather Index (FWI) constitute the fire behaviour indices. We extracted the value of these indices for each cell within each transect using the BioSIM 9 software [36
]. BioSIM allows the user to compensate for the scarcity of weather stations in a study area by interpolating climate data from nearby weather stations, with adjustments for elevation, latitude, and longitude [36
]. We extracted the mean value of each index in the FWI system over the period of 1971 to 2000 at the cell’s centroids. Means of each index in each cell were calculated for spring months (April to June), summer months (July to September), and fire season months (May to August).
2.4. Statistical Analyses
Survival analyses are often used in fire studies because of their ability to examine the time required for an event to occur, which in our case refers to the TSF, and its relationship with one or more covariates. They produce a survival distribution function corresponding to the probability of having gone without fire at each time t
, from which a fire cycle can be calculated. Not only are survival analyses adapted to time to event data, but they also allow for censored observations in the modelling process. This is a major advantage compared to regular fire cycle analyses, which strictly assume that TSF is given by the stand age, with no distinction in stands that have been attributed a minimum age. This often leads to an underestimation of the fire cycle that can be attenuated with survival analyses [37
Among the different types of survival analyses available, we selected a semi-parametric model known as the Cox proportional hazard regression [38
]. Although this model is one of the most commonly used for survival analyses in other fields—mainly in medicine—still very few fire studies have explored its potential (e.g., [39
The Cox proportional hazard model is made of two distinct parts. The first part corresponds to the baseline hazard function (cumulative hazard function when all covariates values are set to zero), i.e., the non-parametric portion of the model, which is initially left unspecified. This has great advantage over parametric models (such as the Weibull distribution, which has been widely used in fire cycle studies) because it avoids making arbitrary and possibly incorrect assumptions about the form of the baseline hazard function. Instead, it is derived from the empirical TSF distribution. In terms of fire history, it means that since the model does not assume a constant risk of burning through time, it allows for variations in the fire regime that could have happened in the past, resulting, for example, from human activities or climate change. Those variations are therefore taken into account while calculating the fire cycle, giving a more precise fire history estimate [37
]. The second part of the Cox model is parametric and is estimated using the method of partial likelihood [38
]. It is used to evaluate the relationship between the tested covariates and survival. Our survival models were built using the coxph
function of the survival
R package [41
Cox proportional hazard models were used all along the analysis process. They were built either with covariates in order to test the effect of climate on fire risk and delimitate fire risk zones along the transects, or as null models in order to calculate the fire cycle of each fire risk zone previously determined. Analyses were performed for each transect separately, as four independent entities, each representative of their surrounding region. Indeed, the four transects are under very different climatic regimes, and merging them into the same analysis process would make impossible the estimation of the climate effect at the scale of one particular transect. Moreover, we wanted to identify variables affecting the fire risk independently for each transect. Although analyses are realized per transect and allow for the latitudinal assessment of the fire risk variability, calculating fire cycles provides a means of assessing the longitudinal variability by comparing fire activity among transects.
2.4.1. Climate Influence on Fire Risk
For each transect, survival models were built in order to examine the influence of the different FWI indices—hereafter referred to as climate variables—on TSF. A supervised forward model selection was conducted in order to select the climate variables that best explained the fire risk. This multi-step process was conducted using the Akaike Information Criterion (AIC). Figure 3
summarizes the different steps of the model selection process. First, univariate models were built in order to test for each climate variable individually. For each model, the AIC and ΔAIC
(i.e., the difference from the model having the lowest AIC) were calculated. Models with a ΔAIC
higher than 6 from the best univariate model were discarded [42
]. The second step consisted in adding a second variable to each model selected. Only variables that were not collinear with the first one (threshold: correlation coefficient of Pearson < 0.7) were tested as second variables. A second variable was kept only if the model with two variables showed a lower AIC value by at least 2 than the AIC of the corresponding univariate model, in which case the univariate model was discarded. The same process of adding variables was repeated until the model could not be improved by any additional variable. The AIC of all selected models were compared and those having a ΔAIC
value higher than 2 from the best model were discarded. Among all models having a ΔAIC
lower than 2, only the most parsimonious ones were retained, and the one with the lowest AIC value was kept as the final model. The AIC of this model was finally compared with the AIC of a null model to ensure the overall improvement. Bootstrap was then applied by randomly sampling with replacement (1000 iterations) the original dataset containing TSF and climate variables to extract a 95% confidence interval on the variables’ estimates using the lower and upper percentiles.
2.4.2. Relative Fire Risk and Latitudinal Risk Zonation
For each transect, the predicted fire risk of each cell was extracted from the final best-fitted model containing the selected climate variables (Figure 3
). A 95% confidence interval on the fire risk was calculated using the same bootstrapping process detailed in the previous section. Because the Cox model is a relative risk model, the predicted risk is relative to the sample used in the model, so it can only be interpreted within a transect. The mean risk of a transect is set to one, and is associated with the mean value of the variables used in the model. The value of the risk can then take any positive value and show how many times the risk equals the mean risk of the transect. We chose to graphically represent the results using the log-transformed values of the predicted risk. This scale indeed implies the same range of risk values on both sides of the mean risk value, which on this scale equals zero. For each transect, the log-transformed predicted fire risk variations along the latitudinal gradient allowed us to identify fire risk zones where the fire risk was diverging significantly from the mean risk of the transect. Each transect was thereby separated into different zones, where each was attributed either a low, moderate, or high fire risk relative to the mean risk of the entire transect.
2.4.3. Fire Cycle
Calculating the fire cycle of each transect zone allows for the comparison of fire activity between transects, as we are no longer dealing with relative estimates within transects. Fire cycles can therefore be used to assess the fire activity variability along both latitudinal and longitudinal gradients. Moreover, in order to compare our results with those of Gauthier et al. [24
] who regionalized the entire coniferous boreal vegetation domain based on fire cycles over the period 1972–2009, fire cycles were calculated for two different periods (i.e., previous to 2014 and to 1972). Calculating fire cycles with and without the 1973–2014 years also allows for the ability to highlight the impact of recent years (post-1972) on past fire regimes, and therefore to assess the temporal variability of fire activity over these two periods.
In order to calculate the observed fire cycle of each transect zone, a stratified null Cox model was built for each transect. No variables were added to the models in order to capture the observed fire cycle per zone, as opposed to a predicted fire cycle, based on the prevailing climate conditions. A special strata
term specifying which cell belongs to which transect zone was added to the models in order to take into account how transects were split into different zones. The estimated cumulative hazard of burning (baseline hazard function) could then be extracted for each transect zone [41
], representing the accumulated hazard of burning through time. The time it takes for the cumulative hazard to reach 1 is equivalent to the fire cycle [37
]. To estimate the fire cycle, the time at which the cumulative hazard reached or exceeded 1 was then divided by its associated cumulative hazard. In case the cumulative hazard never reached 1, the fire cycle was calculated as the time at which the cumulative hazard reached its maximum value, divided by this maximum cumulative hazard value. A 95% bootstrap confidence interval on the fire cycle was calculated using 1000 randomizations with replacement of the original TSF dataset. The confidence interval was computed using the lower and upper percentiles.
The frequency distributions of TSF (Figure 4
) show that whereas most recent fires are dated to the year, they are mostly dated with a minimum TSF beyond 90–100 years. Transects A, B, C, and D show 26%, 32%, 43%, and 44% of minimum TSF data, respectively, thus underlining the importance of considering censored data in survival analyses. Transects C and D are located in the North Shore region of Quebec where the proportion of balsam fir, a fire-sensitive species, is much greater than in the other transects, suggesting that these stands did not establish themselves immediately after a fire event. In these old stands, it is usually difficult to date the TSF precisely, which explains the higher percentage of censored data in these two transects. The minimum dates we recorded for transects A, B, C, and D are 1719, 1703, 1731, and 1663, respectively.
Peaks of TSF can sometimes correspond to single fires. For example, the most recent peak in transect A results at 80% from a very large 2013 fire, although immense fires are common in this region [33
]. In transect D, a large fire occurred in the 1940s in its southernmost part that covered close to 26% of the entire transect. Unlike in transect A, this fire appears as an exceptional event when compared to the surrounding landscape. It is not only the largest, it also covers more than 31% of the area burned since 1924 within a 200 km-wide area centered on the transect and delimited in the north by the breakpoint in the latitudinal zonation section (see below), and in the south, by the bottom of the transect. Moreover, the fire is about 16 times larger than the mean size of all fires that have occurred in this area since 1924. For this reason, the cells associated to this particular fire were either removed or re-associated with a previous TSF (obtained from field data or fire archives). We will refer to this configuration as transect D2. This approach also allows for a clearer demonstration of how this fire is influencing our results, as it could either lead to an overestimation of the fire risk or to a misinterpretation of the climate’s influence on the fire risk.
3.1. Climate Influence on Fire Risk
In all transects, the selected models have a lower AIC than the null models with ΔAIC
values higher than 7 (Table 1
), meaning that null models can be discounted [43
]. Moreover, in all transects except B, the ΔAIC
values with null models are higher than 10, a threshold indicating with high certainty that the selected models are highly plausible [44
are all above 0.35, except for transect B.
All variables in AIC selected models have significant effects on the fire risk (Table 2
). In transect A, the Drought Code (DC) during the fire season increases the fire risk. In transects B and C, models with two variables were selected. The first variables (lowest p
-value) with the most important positive effect on the fire risk, are maximum DC and DC during spring, respectively. The second variables selected show a slight negative effect on the fire risk for both transects, suggesting an adjustment to the positive effects of the first variables. In transects D and D2, the main climatic factor selected is the Fine Fuel Moisture Code (FFMC) during fire season. In contrast to the other transects, FFMC decreases the fire risk even though it is an indicator of sustained flaming ignition and fire spread [45
In the Cox proportional hazard model, the relevant estimates are the exponentiated coefficients, which represent the multiplicative effect on the risk of burning. Thus, if we take the example of transect C (Table 2
), when holding the DC fire season constant, an increase of 1 in the DC spring value increases the risk of burning by an average factor of 2.12. Likewise, an increase of 1 in the DC fire season value decreases the risk of burning by a factor of 0.79 on average.
3.2. Relative Fire Risk and Latitudinal Risk Zonation
For each transect, we defined homogeneous fire risk zones based on whether or not the predicted risk diverged from the mean fire risk of the transect (Figure 2
and Figure 5
). Because the predicted fire risks extracted from the models are relative to each transect, relative risk values cannot be compared from one transect to another.
In transect A, the fire risk gradually increases from south to north, thus allowing two zones (north and south) to emerge. On average, the northern zone shows a fire risk 15.61 times higher than the mean risk of the transect, while the southern zone shows a risk 3.33 times lower. 95% confidence intervals (CI95) for the average relative fire risk of each zone can be found in Table A1
In transect B, the predicted fire risk increases from moderate (not significantly different from the mean fire risk of the transect) in the south to high in the north, except within the zone between latitudes 51.707° N and 52.105° N (coinciding with the plateau located between the Otish Mountains to the east and the Tichegami Mountains to the west, Figure 1
a), where the risk abruptly drops (dashed lines in Figure 5
b). The fire regimes of high elevation areas are often idiosyncratic because hilltops and upperslopes can be subject to lower fire frequency [40
] due to shorter fire seasons and lower temperatures. Because those mountains are not representative of the regions they cross, this section has been removed from the rest of the analyses. Transect B was divided into a southern zone with an average fire risk similar to the mean risk of the transect (1.05 times higher), and a northern zone with a fire risk 3.70 times higher than the mean risk of the transect (excluding the plateau near the Otish Mountains).
In transect C, the predicted fire risk increases stepwise from south to north. This transect was therefore split into a high risk zone in the north, with a risk 10.48 times higher than the mean risk of the transect, a moderate risk zone in the center, with a risk 1.92 times higher than the mean risk, and a low risk zone in the south, showing a risk 2.63 times lower than the mean risk.
In transect D, the fire risk increases from south to north, although the southernmost portion was highly influenced by the 1940s fire (Figure 5
d,e). The transect was thus split into three zones when this fire is included in the analysis, with the southernmost zone—almost exclusively inside the 1940s fire area—showing a moderate to high fire risk 3.75 times higher than the mean risk of the transect. The rest of the transect includes a high fire risk zone in the north and a low fire risk zone in the center, showing fire risks 48.14 times higher and 1.75 times lower than the mean risk of the transect, respectively. When removing the 1940s fire from the analysis, transect D2 could be split into two zones (north and south) with risks 21.30 times higher and 1.45 times lower than the mean risk of the transect, respectively. The breakpoint separating the northern and southern zones in D2 was located 4.4 km south of the northern breakpoint in D. This slight difference results from the increased mean fire risk of the transect due to the 1940s fire in transect D.
3.3. Fire Cycles
In all transects, fire cycles (Table 3
) calculated over the whole period are seen to lengthen from north to south. Globally, there is also a lengthening from west to east. Fire cycles were calculated over two different periods, prior to 1972 and prior to 2014, in order to check for recent changes in fire regimes. Except for zones A north and B north, the fire cycles calculated for the period before 1972 are shorter in all zones than those calculated over the whole period (prior to 2014). The temporal variability in fire activity in each transect zone (Figure A3
) also shows that over the last 300 years most fire activity recorded in each transect zone occurred before 1972, except for A north and B north.
Considered as a whole, the latitudinal breakpoints separating our fire risk zones are largely consistent with the recent evaluation of the northern limit of the commercial forest. This limit has been drawn across the coniferous boreal forest of Quebec to delineate the more opened forests to the north from the tall and dense forests that are suitable for forest management to the south [23
]. As there is a direct link between high fire activity and the opening of forests [26
], as we have shown that fire risk appears to be higher in the northern zones over the last 150–300 years, and as these zones have been relatively stable through time, it seems reasonable to conclude that the limit between open and closed forests has also been somewhat stable. If the expected climate change leads to a fire activity level that remains in the same range of variability as the last 150–300 years, which Girardin et al. [63
] consider a plausible scenario, this limit may also remain stable in the future. Indeed, boreal forests south of the northern limit of the commercial forest seem to be well-adapted to large changes in fire activity [63
]. However, if fire activity increases beyond its range of variability in the south, dense forests could start opening [2
] and thus eventually change the location of the limit.