Current white pine blister rust (Cronartium ribicola
J.C. Fisch) infection and mountain pine beetle (Dendroctonus ponderosae
Hopk.) impacts, combined with the effects of climate change and fire suppression, have placed whitebark pine (Pinus albicaulis
Engelm.) in an endangered status in Canada, and as a candidate species for listing under the United States Endangered Species Act [1
]. Whitebark pine is considered a keystone species in high-elevation forests in western Canada and the USA [3
]. The seeds are an important food source for many animals [4
] and whitebark pine delivers numerous ecosystem services including stabilizing the snowpack and runoff, and influencing treeline vegetation patterns [5
]. Canada is home to 56% of the global range for the species. While recovery planning is underway, restoration actions are in the early stages [8
]. Additional research on the effectiveness of early restoration efforts and strategies across the species’ range will accelerate science-based management and recovery of the species. The range-wide restoration strategy for whitebark pine [9
], recovery strategy for whitebark pine in Canada [8
] and recovery plan for Alberta [10
] highlight principles and possible actions to guide restoration. These include thinning competing tree species, collecting putatively blister rust-resistant seeds, screening seed stock for natural resistance, planting rust-resistant seedlings, and using prescribed fire to emulate the natural disturbance regimes that benefit whitebark pine regeneration [9
To support restoration action, spatially-explicit results and maps depicting whitebark pine stand conditions are required to identify restoration potential and set priorities for action at the scale of the Canadian Rocky and Columbia Mountains [9
]. We present monitoring data from whitebark pine stands surveyed across the Canadian Rocky and Columbia Mountains over three visits between 2003 and 2014. The study builds on previous work completed by Smith et al. [11
], who reported on the incidence of disease and mortality for these plots from 2003 to 2009. We examine how spatial, temporal, and climate variables influence patterns of white pine blister rust infection and mortality for trees (>1.3 m in height) and seedlings, with the goal of informing restoration planning for this region. Specifically, we sought to: (1) identify variables predicting live tree infection, seedling infection, canopy kill (i.e., severity of infection), mortality, and regeneration across this large region; and (2) present the results in spatially-explicit formats to assist land managers when setting priorities, targeting actions for areas to increase the likelihood of success, and coordinating restoration actions across boundaries.
We focused on blister rust because it is the most widespread threat across the whitebark pine range in Canada [13
]. Mountain pine beetle-infested trees and beetle-killed trees were included in the scope of our study. Although mountain pine beetle was only patchily encountered in our study area, and has been declining in most of the study area, it has been expanding into more northern, high elevation forests [15
]. Mountain pine beetle has documented interactions with blister rust, preferentially attacking rust infected trees [16
]. Canopy kill indicates both the severity of the blister rust infection and represents a loss in seed production when cone-bearing branches are killed by blister rust. We expect natural regeneration to decline as a result of canopy kill, and blister rust and beetle-caused mortality [18
]. We examined natural regeneration to better understand the relationships between regeneration and blister rust infection in the region, and to identify areas that may benefit from restoration treatment by thinning or prescribed fire, and/or seedling planting. The study is an important opportunity to examine stand conditions across an area with widely varying climatic and topographic conditions over a ten-year period, in order to understand how whitebark pine responds to mountain pine beetle and blister rust. For example, live tree infection, mortality, canopy kill, and regeneration may vary across different stages of an epidemic and may depend on other measurable variables (e.g., climatic conditions) [11
]. Understanding the trajectory of a disease over time and across a large region is key for restoration action.
2. Study Area and Methods
The total study area is approximately 92,000 km2
and spans over four degrees of latitude from the United States—Canada border in the south to the Willmore Wilderness Park in the north (Figure 1
). The study area encompasses national parks, numerous provincial parks, and a mix of private and public provincial lands along the continental divide in Alberta and British Columbia in the Canadian Rocky and Columbia Mountains. The landscape is highly variable with whitebark pine stands ranging in elevation from 1300 m to 2370 m. The lower-elevation eastern slopes are dominated by stream networks, generally more open canopies with lower precipitation in summer and winter. High elevation mountain areas include rock, glaciers, complex steep topography with varying aspects, slopes, and moisture regimes, particularly between areas east and west of the continental divide. Forest types at the study sites are primarily dominated by whitebark pine and include subalpine fir (Abies lasiocarpa
(Hooker) Nuttall), Engelmann spruce (Picea engelmannii
Engelmann), and lodgepole pine (Pinus contorta
Douglas ex Loudon), with limber pine (Pinus flexilis
E. James) co-occurring with whitebark pine in the Waterton Lakes National Park area.
Methods for establishing plots and assessing whitebark pine health were those recommended by Tomback et al. [19
] and used in Smith et al. [11
]. A 10 m wide belt transect was extended to a variable distance that attempted to sample a minimum of 50 whitebark pine trees >1.3 m, with a minimum of 10 trees that were living or recently dead. In 2009, most plots were shortened to 50 m to align with refinements to Tomback et al. [19
]. Field crews were trained to recognize blister rust symptoms in whitebark pine, worked in teams of two to three people, and surveys occurred between the time of snow melt and mid-August, before aecia fade. Within each plot, all whitebark pine trees (>1.3 m in height) were marked with numbered aluminum tags. Plots were surveyed over repeat visits between 2003 and 2014, forming three distinct sampling periods for future analysis—2003–2007; 2008–2012; and 2014. During each visit, we recorded diameter at breast height for all trees to the nearest 0.1 cm. Living trees were visually assessed using binoculars for presence-absence of active or inactive branch and stem cankers caused by white pine blister rust (WPBR) and for mountain pine beetle infestation (beetle entry holes with pitch plugs or J-shaped galleries). Active cankers were identified by diagnostic orange-yellow aecial blisters containing aeciospores, or empty white spore sacs later in the season. Inactive WPBR cankers were identified by their spindle shape, broken bark and, because rodents feed on active blister rust cankers, the presence of gnawing or bark stripping [20
]. Surveyors recorded the presence of the following symptoms of infection for each tree: roughened bark, branch flagging, topkill, swelling, oozing sap, and rodent chewing. Trees with uncertain stem or branch cankers were not included in the analysis for infection. The cause of death for all dead whitebark pine was recorded where it could be determined. The percentage of the total canopy killed (dead branches) was estimated (as a proportion of the entire canopy volume) to the nearest 10% for each tree and also summarized as an average per plot. All dead branches were considered, whether caused by blister rust, bark stripping, or mechanical damage, which are difficult to distinguish. We used canopy kill as a proxy for the severity of the infection. All live whitebark pine ≤1.3 m in the plot were considered seedlings and were placed in two size-classes (short, ≤50 cm; and tall, >50 cm), and assessed for the presence or absence of active or inactive cankers.
Spatial and temporal patterns in whitebark pine stand and tree health were analyzed using the following response variables: (1) presence and absence of white pine blister rust in live trees and seedlings; (2) tree mortality (tree status: live/dead); (3) average percent canopy kill in a plot; and (4) regeneration (count of seedlings in a plot). We selected predictor variables based on findings from previous research [11
] and hypotheses based on our own observations in our ecosystem (Table 1
). We obtained mean (1981–2010) annual and seasonal climate data for each plot location using ClimateWNA (an extension of ClimateBC) [25
]. Using a geographic information system (GIS; ArcGIS 10, ESRI, Redlands, CA, USA), we calculated landscape predictor variables that we could not measure in the field. Surveyors measured diameter at breast height (cm) for all trees, except those with a krummholz form. Aspect was transformed from a circular measure to a continuous variable better suited for modelling. Cooler and wetter north–northeast orientations were assigned values close to zero, whereas hotter and dryer south–southwest orientations were closer to 1.0. Distance of the study stand to the continental divide was measured using Euclidian distance and recorded as increasingly negative and increasingly positive west and east of the divide, respectively.
All statistical analyses were performed with the software R (Version 3.1.1; [27
]). We applied data exploration following Zuur et al. [28
] to the final data to assess homogeneity, potential outliers, and zero-inflation, and all predictor variables were standardized prior to analysis. Spatial interdependency was assumed based on the study design given the clustering of trees within plots and repeated measures of plots over time. We used Variance Inflation Factor (VIF) analysis and Spearman’s rank correlations to test for collinearity. If VIF was higher than 3 and pair-wise correlation was p
> 0.5 (or p
< −0.5), we eliminated the variable considered less relevant biologically.
We used a mixed effects logistic regression (package: lme4) to model the relationships between explanatory variables and the following response variables: probability of WPBR in (1) live trees (n
= 16,716) and (2) seedlings (n
= 11,709), and (3) the probability of mortality (n
= 22,052), all measured at the tree level. Because the response variable for canopy kill was a proportion, we used a beta regression mixed model (package: glmmADMB) to model the average proportion of tree canopy kill per plot (n
= 463). We used a negative binomial mixed model (package glmmTMB) to model whitebark regeneration because counts of seedlings per plot were over-dispersed (n
= 463); we included an offset to adjust for unequal plot sizes. When visual inspection suggested a non-linear relationship between the response and predictor variables, the quadratic term was included in model selection. We applied a stepwise backward regression method to select the best models using the Akaike information criterion [29
], choosing the model with the fewest parameters when models were considered equivalent (ΔAIC
< 2). We used data from plots measured at least two times for the disease, mortality, and canopy kill models; for regeneration, we used only data from plots measured for three visits to protect against finding year effects on regeneration due to new plots added rather than a change in regeneration.
We used standard residual plots to validate the canopy kill and seedling regeneration models and k-fold cross validation to evaluate the predictive performance of all the final regression models. Cross-validation involved randomly splitting data nested within plots into five groups of equal size, fitting each top model to data from four of the folds and using the model to predict to data from the fifth fold. We report the predictive performance of the cross-validated model using the Area Under the Curve (AUC) of the receiver-operating characteristic (ROC) plot for binary responses and the root mean square error (RMSE) for beta and negative binomial response variables. The AUC describes the overall ability of the model to correctly discriminate between two observations (e.g., disease presence and absence) and results above 0.7 and 0.9 indicate moderate and high model performance, respectively [30
]. The RMSE measures the difference between predictions and observations of the test-fold datasets and models with good predictive power have a small RMSE relative to the range of the response variable.
To visualize the results for blister rust infection, mortality, canopy kill, and regeneration, we used ARCGIS 10.1 Spatial Analyst to create spatial predictions for the best fitting models. We applied the mean coefficients to each predictor of GIS layers to develop the maps and we clipped the output by elevation to eliminate predictions outside the natural elevation range for whitebark pine in our dataset. We classified the predictions in the resulting maps in 10 quantile bins to represent areas of increasing relative probabilities, and for regeneration, to represent increasing seedling counts.
3.1. Incidence of Infection and Mortality
We sampled whitebark pine trees >1.3 m during three visit periods: (1) 2003–2007 (8385 trees), (2) 2008–2012 (7440 trees), and (3) 2014 (8000 trees). The trees were clustered in 181 permanently-marked plots; 135 plots were surveyed three times, while 24 plots had two visits and 22 plots were sampled only once. Over 84% of the trees included in our analyses were measured three times in 2003/04, 2009, and 2014. During visit 1, 84% (n = 7009) of trees were alive, while 76% (n = 5690) of trees were alive at visit 2, and 74% (n = 5886) of trees were alive at visit 3. Of these trees examined, we removed live trees with heavy lichen loads obscuring the bark from the dataset, because they could not be properly assessed for WPBR (63, 242, and 363 trees were removed for visits 1, 2, and 3, respectively).
For the 135 plots surveyed three times, 38% of living trees were infected at visit 1, 45% at visit 2, and 53% were infected at visit 3 with WPBR (had active or inactive cankers) (Figure 2
). Trees that died between sampling intervals were removed from the incidence of infection calculations and were included in calculations of rates of mortality. Infection levels increased 1.5% year−1
between the first visit and the third visit (~2003 to 2014). Of the infected live trees, 20% had stem cankers in visit 1, increasing to 27% in visit 2, and 29% in visit 3. While large trees may live for many years with stem cankers, the cankers often cause canopy kill and significantly reduce cone and seed production in the tree. The largest live tree sampled exhibited an 88 cm diameter at breast height (DBH), while the mean tree DBH across the study area was 10 cm. Mortality levels increased 0.8% year−1
over the ten years between the first and third visits, from 17% at visit 1, to 22% and 25% at visits 2 and 3, respectively (Figure 2
For dead trees in plots surveyed three times, cause of death could be determined for 25% (348 of 1376) of surveyed trees in visit 1, 32% (566 of 1750) in visit 2, and 33% (700 of 2114) in visit 3. Of the trees for which we could attribute a cause of death in visit 1, 64% (223) had definite signs of WPBR (active or inactive cankers), increasing to 75% (423) in visit 2 and 78% (544) in visit 3. Evidence of mountain pine beetle (MPB) infestation (J-shaped galleries) informed the cause of death for 35% (125), 25% (143), and 22% (152) trees in visit 1, 2, and 3, respectively.
To assess seedling blister rust infection and regeneration, we counted 1986, 1726, and 2576 short seedlings and 1521, 2048, and 1882 tall seedlings in plots over the three visits, respectively. Of the tall seedlings, 25% (388), 22% (443), and 28% (535) were infected in visit 1, visit 2, and visit 3, respectively, while approximately 7% (152, 124, 190) of short seedlings were infected during all three visits.
3.2. Blister Rust Infection in Live Trees
Blister rust infection in live trees has increased significantly since 2004 (Table 2
, Figure 3
a) and infection is more common in larger diameter trees (Table 2
, Figure 3
b). Tree diameter ranged from approximately 1 cm to 88 cm, and in 2014, the median tree diameter was 6 cm for healthy trees, and 11 cm for blister rust infected trees. Blister rust infection for live trees was variable across the study area. Figure 3
a compares the 2014 probability of infection for a 10 cm diameter tree at a mid-latitude location (50% probability of infection) and a southern location (>80% probability of infection), at mean values for growing degree days and distance to divide. The probability of infection was higher for trees in stands located west of, or near, the continental divide, and at sites with a longer growing season (Table 2
). The probability of blister rust infection map (Figure 4
a) indicates that the risk of blister rust infection for a 10 cm diameter tree ranges from a high in south-east BC of 96% to a low of 2% along the northern edge of the eastern slopes. The blister rust infection model had a moderate discrimination ability within the range of the predictor variables used in the study area [31
] with an average cross-validated area under the curve of 0.78 (SD = 0.02), indicating that the model can discriminate between blister rust infected and uninfected trees 78% of the time.
3.3. Whitebark Pine Mortality
We found that latitude, growing degree days, distance to the continental divide, and year were significant predictors of tree mortality (Table 2
). Figure 5
a illustrates that the probability of tree mortality has increased by approximately 10% between visit 1 and visit 3 (~10 years). Inclusion of the quadratic term for latitude indicates that mortality was highest in the southern range, declined to the mid-latitude stands in Banff National Park, and increased slightly near the northern extent of the study area in Jasper National Park (Figure 5
b). The probability of mortality map demonstrates that the spatial pattern in tree mortality ranged from 91% probability of mortality in south to 1% along the east slopes (Figure 4
b). The average cross-validated area under the curve of the model predicting whitebark pine mortality was 0.80 (SD = 0.06), indicating that the model can discriminate between dead and live whitebark trees 80% of the time.
3.4. Whitebark Pine Canopy Kill
Percent average canopy kill increased over time from approximately 20% during visits 1 and 2 to 40% during visit 3 (Table 2
, Figure 6
a). The greatest increase in canopy kill occurred in the southern portion of the study area (Figure 6
b). Canopy kill was higher west of, and near, the continental divide, and at stands with longer growing degree days (Table 2
, Figure 7
). We included the interaction between latitude and year based on visual inspection of the variables and standard diagnostic plots. The five-fold cross-validation indicated a good predictive performance: the root mean squared error was 0.02 (SD = 0.001), which is small relative to the range of the canopy kill response variable.
3.5. Whitebark Pine Seedling Infection
Blister rust infection in seedlings declined from visit 1 to visit 2, but increased from visit 1 levels in visit 3 (Table 2
, Figure 8
a). The 2014 probability of infection for a tall seedling (based on mean values for predictor variables) is almost 20%. As with whitebark trees, seedling infection is spatially variable (Table 2
, Figure 9
a), and is higher in tall seedlings than short seedlings (Table 2
, Figure 8
a). Seedlings are much less likely to be infected at middle and northern latitudes (Figure 8
b), and are more likely to be infected at sites with higher spring radiation (Table 2
). The average cross-validated area under the curve of this model for live seedling infection was 0.78 (SD = 0.02).
3.6. Whitebark Pine Natural Regeneration
Natural regeneration in plots ranged from 0 to 2254 (mean = 463) seedlings per hectare, estimated from plots that ranged in size from 140 to 2830 m2
; however, the majority of plots (80%) were 500 m2
. Eleven plots (n
= 181 plots) had no regeneration. We estimated live basal area using DBH measurements from all whitebark. Average basal area was 9.2 m2
per hectare (SD = 7.3) in 2014. We tested basal area and canopy kill, which reduces the cone producing potential in a stand, for inclusion in the model, but only canopy kill was retained. Regeneration was highest in the southern and northern regions of the study area and on south-west aspects (Table 2
, Figure 10
a). Seedling counts peaked where there were moderate growing degree days, declined as average canopy kill increased at a plot (Figure 10
b), and increased slightly in 2014 compared to the previous years (Table 2
). The natural regeneration map illustrates the spatial variation in regeneration and highlights pockets of low regeneration that occur in the southern east-slopes and southern continental divide areas (Figure 9
b). Validation plots and cross-validation indicated a good predictive performance: the RMSE is 20.6, which is small relative to the range (0–138) of seedling counts. All predictions (plots and maps) used the plot size of 500 m2
for the offset to account for variable plot size in seedling counts.