Fig 1.
Overview of the study area encompassing 23 wildfires that burned throughout ponderosa pine-dominated forests in the Colorado Front Range from 1996–2013.
We used systematic aerial image interpretation of 2015 NAIP imagery to identify post-fire presence (Conifer Refugia) or absence (Conifer Absence) of at least one mature conifer in each 30-m grid cell within each fire perimeter. Each fire is mapped using the binary variable Conifer Refugia or Conifer Absence.
Fig 2.
Demonstration of methods used to create binary Conifer Refugia and Conifer Absence variable.
Conifer Refugia is defined by the presence of post-fire mature conifers, and Conifer Absence contained no live mature trees following the fire. Panels A-D span the same extent within the 2012 High Park fire. Pre-fire black and white aerial images were used to enhance the classifications of 2001 NLCD Pre-fire Forest Cover (A). Post-fire four band aerial images in 2015 (B) were used to identify presence or absence of a mature post-fire conifer (C) in each 30-m grid cell that was recorded by Monitoring Trends in Burn Severity (MTBS) Thematic Burn Severity Classes (D).
Table 1.
List of 25 landscape predictor variables generated for the random forest model meant to classify Conifer Refugia and Conifer Absence.
An expected positive relationship between the variable and Conifer Refugia is denoted with (+) and an expected negative relationship with (-).
Table 2.
Spatial metrics of post-fire patches of Conifer Absence (A) and Conifer Refugia (R) within 23 fires that burned ponderosa pine-dominated forests along the Colorado Front Range from 1996–2013.
Using image interpretation of 1 m aerial imagery to identify presence or absence of a mature conifer in each 30-m grid cell, Conifer Refugia patches are defined by the contiguous cells with the presence of post-fire mature conifers, and Conifer Absence patches are defined by contiguous cells with the absence of post-fire mature conifers. Median gives the median attribute value across all individual fires and Total gives the attribute value for all fires combined.
Fig 3.
Spatial characteristics of Conifer Refugia and Conifer Absence within 23 fires that burned ponderosa pine-dominated forests along Colorado’s Front Range 1996–2013.
A) Percent Conifer Refugia for each fire plotted across time (no significant trend, p≥0.05), B) significant linear relationship between the log of Maximum Patch Size of Conifer Absence and the log of Fire Size (R-squared = 0.85, p≤0.0001), and C) Distribution of Distance from Conifer Absence to Conifer Refugia seed source. Thirty-eight percent of the area within the 23 fire perimeters is greater than 50m from a seed source. Using 1-m aerial image interpretation to identify presence or absence of a mature conifer in a 30-m grid cell, Conifer Refugia is defined by the post-fire presence of at least one mature conifer, and Conifer Absence as the opposite.
Fig 4.
Relationship between Conifer Absence and Conifer Refugia and MTBS burn severity metrics within 21 fires which burned ponderosa pine-dominated forests along Colorado’s Front Range 1996–2013.
(A) Thematic Burn Severity Classes (TBSC) as classified by Monitoring Trends in Burn Severity (MTBS) and (B) Relative differenced Normalized Burn Ratio (RdNBR). Using 2015 1m aerial images, we classified Conifer Refugia as the presence of a mature post-fire conifer, and Conifer Absence as the absence of a post-fire mature conifer in 30m grid cells. In (A), Percentage of Conifer Refugia (green) and Conifer Absence (orange) in TBSC is shown within each bar. Horizontal dashed line in (B) denotes the results of a classification and regression tree predicting Conifer Refugia and Conifer Absence using MTBS TBSC and RdNBR, in which a single split at RdNBR <544 best predicted Conifer Refugia and had an overall accuracy of 78%.
Fig 5.
Relationships between forest cover and MTBS thematic burn severity and Conifer Refugia and Conifer Absence.
Percent area of (A) Monitoring Trends in Burn Severity (MTBS) thematic burn severity classes (TBSC) and (B) post-fire Conifer Refugia and Conifer Absence with Pre-Fire Forest Cover classes burned by 23 fires in ponderosa pine-dominated forests along Colorado’s Front Range 1996–2013. Pre-Fire Forest Cover was classified by the 2001 NLCD Tree Cover and augmented with 1994 and 1999 aerial imagery. Values above bars in (B) show the percentage point between Conifer Absence and Conifer Refugia (negative numbers denote Conifer Absence is greater than Conifer Refugia).
Fig 6.
Relationships between weather variables and Conifer Refugia and Conifer Absence.
Percent area of Conifer Absence and Conifer Refugia within 23 fires that burned ponderosa pine-dominated forests along Colorado’s Front Range 1996–2013 within climate variable classes for A) Fire Danger Rating (Low, Moderate, High, Very High, Extreme), B) Maximum Wind Speed, C) Minimum Relative Humidity, and D) Maximum Temperature. Values above bars show the percentage point between Conifer Absence and Conifer Refugia (negative numbers denote Conifer Absence is greater than Conifer Refugia).
Fig 7.
Relative variable importance in predicting Conifer Refugia.
Results of Random Forest model shown in Relative Variable Importance (relativized Mean Decrease in Accuracy) for 20 variables for predicting location of Conifer Refugia within 23 wild fires that burned throughout ponderosa pine-dominated forests in the Colorado Front Range from 1996–2013. Presented variables (20 out of 25) were selected for model inclusion in at least six of the 11 model iterations. Mean decrease in accuracy is the normalized difference of the accuracy of the classification when the data for that variable are included versus when they have been randomly permutated. Values were relativized to sum to one for each model run. Higher values indicate greater importance and a higher error value when this variable is removed from the model. Dots represent the Relative Variable Importance for the primary model with the lowest Out of Bag (OOB: 23.2%) error rate, giving it an accuracy of 76.8%, and fewest selected variables (15 out of 25). Horizontal segments show the range (minimum to maximum) of Relative Variable Importance across all model iterations. Numbers in parentheses give the median ordered rank of the Relative Variable Importance across 11 model iterations.
Fig 8.
Partial-dependence plots for variables predicting Conifer Refugia.
The random forest partial-dependence plots for continuous variables for the primary model with the lowest OOB and fewest selected variables (n = 15). Legend provides rank of variable importance. Partial dependence plots show the dependence of the probability of Conifer Refugia on each individual predictor variable after averaging out the effects of the other predictor variables. The y axis (Refugia Prob.) is defined as the logit probability of Conifer Refugia/2. Values for Refugia Prob. above zero indicate a higher likelihood of the presence of Conifer Refugia and value below zero indicate a higher likelihood of the absence of Conifer Refugia.
Fig 9.
Partial-dependence plots for variables predicting Conifer Refugia.
The random forest partial-dependence plots for continuous and categorical variables for the primary model with the lowest OOB and fewest selected variables (n = 15). Legend provides rank of variable importance. Partial dependence plots show the dependence of the probability of Conifer Refugia on each individual predictor variable after averaging out the effects of the other predictor variables. The y-axis (Refugia Prob.) is defined as the logit probability of Conifer Refugia/2. Values for Refugia Prob. above zero indicate a higher likelihood of the presence of Conifer Refugia and value below zero indicate a higher likelihood of the absence of Conifer Refugia.