Assessing Wildfire Burn Severity and Its Relationship with Environmental Factors: A Case Study in Interior Alaska Boreal Forest

In recent years, there have been rapid improvements in both remote sensing methods and satellite image availability that have the potential to massively improve burn severity assessments of the Alaskan boreal forest. In this study, we utilized recent pre- and post-fire Sentinel-2 satellite imagery of the 2019 Nugget Creek and Shovel Creek burn scars located in Interior Alaska to both assess burn severity across the burn scars and test the effectiveness of several remote sensing methods for generating accurate map products: Normalized Difference Vegetation Index (NDVI), Normalized Burn Ratio (NBR), and Random Forest (RF) and Support Vector Machine (SVM) supervised classification. We used 52 Composite Burn Index (CBI) plots from the Shovel Creek burn scar and 28 from the Nugget Creek burn scar for training classifiers and product validation. For the Shovel Creek burn scar, the RF and SVM machine learning (ML) classification methods outperformed the traditional spectral indices that use linear regression to separate burn severity classes (RF and SVM accuracy, 83.33%, versus NBR accuracy, 73.08%). However, for the Nugget Creek burn scar, the NDVI product (accuracy: 96%) outperformed the other indices and ML classifiers. In this study, we demonstrated that when sufficient ground truth data is available, the ML classifiers can be very effective for reliable mapping of burn severity in the Alaskan boreal forest. Since the performance of ML classifiers are dependent on the quantity of ground truth data, when sufficient ground truth data is available, the ML classification methods would be better at assessing burn severity, whereas with limited ground truth data the traditional spectral indices would be better suited. We also looked at the relationship between burn severity, fuel type, and topography (aspect and slope) and found that the relationship is site-dependent.


Introduction
The circumpolar north has experienced a temperature change 1.5 times higher than the global average [1,2]. The intensification of global climate change is leading to an increase in frequency, size, severity, and duration of boreal forest fires due to the warming summer temperatures (hot, dry, and windy conditions), earlier snowmelt that lengthens the fire season, and variable precipitation [2][3][4][5][6]. In the past 20 years (2001-2020), wildfires in Alaska burned 30.9 million acres, which is 2.1 times more acres burned than during the previous two decades (1981-2000: 14.6 million acres burned). In 2019, 721 fires burned Fairbanks and are accessible by road ( Figure 1). These burn scars offer a unique opportunity to test remote sensing methods for mapping boreal fire burn severity and understanding burn severity's relationship to environmental factors. Application of advanced machine learning algorithms, such as Random Forest (RF) and Support Vector Machine (SVM), for mapping and assessing burn severity in the boreal forest has been limited [29]. The RF classifier is a supervised classification algorithm that uses decision trees to classify pixels into classes [30]. It has the ability to effectively process large volume datasets while avoiding overfitting [31], making RF an ideal classifier if a large amount of training data is available. Though only a few studies have used RF to classify burn severity [29], there is a volume of literature demonstrating its efficiency in mapping land cover, making it a good candidate for testing burn severity assessments [30,32,33]. Support Vector Machine (SVM) is another machine learning classifier that learns to label objects into classes using previously provided examples [34]. For remote sensing image classifications, SVM has been demonstrated to be one of the most successful classifiers [35]. SVM is an ideal classifier when training data is limited, runs on datasets with high dimensionality, and has been used previously to generate land cover maps [35][36][37][38]. It uses training sites to find the optimal separating hyperplane between classes [39]. Although RF and SVM have not been heavily used to study burn severity, they both have great potential for this application [29].
The primary goal of this study is to determine which of the aforementioned remote sensing techniques produce the most accurate products for assessing wildfire burn severity in the boreal forest. The secondary goal is to investigate how burn severity relates to fuel type and topography. To address these goals, we collected ground-based data within the Nugget Creek and Shovel Creek burn scars, trained the classifiers using this data, and then validated the performance of the classifiers with accuracy assessments. By performing this investigation of burn severity within the 2019 Nugget Creek and Shovel Creek fires, we hope to identify which remote sensing methods are best for assessing burn severity and to better understand how fuel type and topography influence burn severity.

Study Area
The Shovel Creek burn scar is located approximately 20 miles northwest of Fairbanks and was ignited by a lightning strike. The fire started June 21, 2019 and burned 91 km 2 [40]. The Nugget Creek burn scar is located 35 miles east of Fairbanks and was also started by a lightning strike on June 21, 2019 and consumed 77 km 2 [41]. The suppression cost of the Nugget Creek fire was around USD 56,000 while the suppression cost for the Shovel Creek fire was over USD 25 million [42,43]. The Shovel Creek fire was suppressed more heavily as it threatened life and property in the Murphy Dome community.

Field Work
During the summer of 2020, we surveyed burn severity within the Nugget Creek and Shovel Creek burn scars following the Composite Burn Index (CBI) protocol modified for Alaska [28,44] (Supplementary Figure S1). CBI is a standard survey protocol used to assess ecological changes following a wildfire [20]. The modified CBI is broken into five different strata, including (a) substrates; (b) herbs, low shrubs, and trees less than 1 m; (c) tall shrubs > 1 m and trees 1 to 2 m; (d) intermediate trees 2-8 m; and (e) big trees > 8 m. Each stratum contains 4-5 variables that assess the amount of change by looking at various factors, such as species mortality, char height, changes to soil and duff layers, changes in species composition, and several other factors that are rated by the surveyor [38,44]. CBI takes the average score of all the factors rated to assign a burn severity value that can range from 0 to 3.0. At Nugget Creek, the CBI values for the fire-disturbed plots ranged from 0.34 to 2.46, and at Shovel Creek the values ranged from 0.09 to 2.38. We used a Trimble R10 RTK GPS to get the precise locations of the field plots (30 m in diameter). The location of the plots surveyed was based on site accessibility and visual interpretation of burn severity from pre-and post-fire images in order to capture the range of severity. At Shovel Creek, we surveyed 52 CBI plots while at the Nugget Creek burn scar, we surveyed 28 CBI plots ( Figure 2). We grouped the CBI values into four classes, taking into account the impact of fire on the five strata that were rated as part of the CBI assessment and the spread of the CBI values within the two burn scars: unburned (CBI = 0), low (CBI = 0.01-1.49), moderate (CBI = 1.50-1.99), and high (CBI = 2.0-3.0) ( Table 1).

Burn Severity Mapping from Sentinel-2 Images
Optical satellite images with bands that cover the visible, near-infrared, and shortwave infrared wavelengths are ideal for mapping and assessing burn severity as vegetation and burned areas can be accurately differentiated using these bands [45]. We used Sentinel-2 Level 1-C reflectance satellite imagery, publicly available on the USGS EarthExplorer portal (source: https://earthexplorer.usgs.gov/). The Sentinel-2 bands used in this study had a spatial resolution of 10-20 m. Bands 2, 3, 4, and 8 have a 10-m pixel size while bands 5, 6, 7, 8A, 11, and 12 have a pixel size of 20 m. For the Shovel Creek fire, we used a pre-fire image collected on May 16, 2019 and a post-fire image acquired on May 20, 2020. For the Nugget Creek fire, we used a pre-fire image collected July 24, 2018 and a post-fire image collected May 17, 2020. The pre-and post-fire images were adequate for assessing burn severity because the images were collected in the summer season and the vegetation was in a similar growth stage.
For both burn scars, we generated pre-and post-fire NDVI and NBR maps, as well as differenced NDVI (dNDVI = preNDVI-postNDVI) and differenced NBR (dNBR = preNBR − postNBR) maps (Table 2 and Figure 3). We used linear models to link spectral indices with CBI values. The estimated CBI values were then broken into three burn severity classes: low-severity (CBI: 0.01-1.49), moderate-severity (1.5-1.99), and high-severity (2.0-3.0). We used all the CBI plots to assess the user, producer, and overall accuracy of the spectral indices-based burn severity maps (Nugget Creek: 28 plots; Shovel Creek: 52 plots). The Sentinel-2 sensor is designed to register image products to a global reference image framework that usually has a residual geolocation error. The absolute geolocation performance specification for Sentinel-2 is 12.5 m [46]. We used a Trimble R10 RTK GPS (with mm accuracy) to get the precise locations of field CBI plots (30 m in diameter). In order to account for the location offset between the surveyed CBI plots and image pixel, we used a 20-m buffer around each plot for accuracy assessment [46,47].

Index
Formula Pre-fire NBR-Post-fire NBR For running supervised classifications, we first removed the unburned pixels from the image. In order to mask out unburned pixels, we tested different indices and threshold values by visually inspecting how well they masked out the unburned pixels. The NDVI index with a threshold of 0.4 was most effective for the Nugget Creek burn scar (all 10 unburned CBI plots are correctly identified; Supplementary Table S7) while at the Shovel Creek burn scar the dNBR index with a threshold of 0.24 was the most effective (all 14 unburned CBI plots are correctly identified; Supplementary Table S4). We removed unburned areas to avoid potential misclassification of the low severity class as unburned. Sentinel-2 bands 1, 9, and 10 are mainly used for atmospheric corrections and cirrus-cloud screening and are not useful for land cover classification; therefore, we removed them from the image classification [48]. We used only post-fire images to run the SVM and RF (500 trees, 30 rules per tree) supervised classifications [49]. We used half of the burned CBI plots surveyed at each site (9 at Nugget Creek and 20 at Shovel Creek) to train the classifiers. Each CBI plot that was used for training contained 2-4 pixels. While the remaining half of the CBI plots were used for accuracy assessment. We classified the images into low, moderate, and high burn severity classes based on the CBI values (see values in Table 1). We used the remaining CBI plots (9 at Nugget Creek and 18 at Shovel Creek) with a 20-m buffer (to account for image positional offset) to assess the classification accuracy. For RF classification, we also tested if including elevation, aspect, and slope as dependent variables with the Sentinel-2 bands would improve the overall mapping accuracy. Since it did not, we did not use this mapping technique for downstream analysis, and only reported the results from Sentinel-2 bands products. We reported accuracy as the percent of correctly mapped CBI plots. We used ESRI's ArcMap 10.7 software for all image processing and analyses.

Burn Severity Relationship to Environmental Factors
Using the 2019 Alaska Vegetation and Wetland Composite (AVWC) map we calculated the percent area occupied by each level II vegetation class within the Nugget Creek and Shovel Creek burn scars. The AVWC maps vegetation at a scale analogous to level IV of the Alaska Vegetation Classification [50,51]. We analyzed the relationship between vegetation type, topography (aspect and slope), and burn severity. We used a 2 m Arctic Digital Elevation Model (DEM) as a source for generating the slope and aspect maps of the burn scars (source: Polar Geospatial Center, University of Minnesota). We used ESRI's ArcMap 10.7 [49] software for image processing, and RStudio 1.2.5019 [52] for all statistical analyses. We generated plots visualizing these relationships between environmental variables (slope grade, aspect, and fuel type) and burn severity class using ggplot2 version 3.3.2 [53]. Regression analysis, namely ordinal regression, has been used in previous remote sensing studies to generate predictive models with fire severity using an ordinal scale [54,55]. We used Ordinal Logistic Regressions (OLR) to generate models with burn severity classes as the response, and aspect, slope, burn scar site (Nugget Creek or Shovel Creek), and fuel type as the explanatory variables, using the plor function in the MASS package (version 7.3-51.5). We used a two-way ANOVA test to quantify the effect of each explanatory variable on severity class with a type III sums of squares, which identifies the variables that are significant independently of one another. Lastly, we performed model selection using the Akaike Information Criterion (AIC), which is a common model selection tool used to test the quality of ecological models relative to other models generated using a single dataset [56]. We used AIC to test the statistical significance of the full model (severity class ~ aspect + slope + site + fuel type), and all combinations of the third-, second-, and first-order models. This model selection allowed us to determine which combination of explanatory environmental variables are best at explaining burn severity in our dataset.

Nugget Creek Burn Severity Mapping
At Nugget Creek, all of the indices (NDVI, NBR, dNDVI, and dNBR) had a high correlation to CBI (|r| > 0.75), with dNDVI having the highest (|r| = 0.80) and NDVI the lowest (|r| = 0.75) ( Figure 4). Linear models produced for each index indicated a significant relationship between the respective indices and CBI (p < 0.05) that were then used to produce burn severity maps. The best performing index using only a post-fire image was NDVI, with a 96.43% accuracy while the NBR had an accuracy of 75% (Table 3). The dNBR (89.29% accuracy) was the best performing differenced index while the dNDVI had a 71.43% accuracy. Both supervised classification products had a 66.67% accuracy (Table 3 and Figure 5). Overall, the NDVI-based burn severity map was the best product at Nugget Creek (Table 3).

Shovel Creek Burn Severity Mapping
At Shovel Creek, the indices had low to moderate correlations (r: 0.33-0.66) to CBI ( Figure 6). The NDVI had the highest correlation (|r| = 0.66) and dNDVI the lowest correlation (|r| = 0.34) (Figure 6). Linear models at Shovel Creek produced for each index had a significant relationship between the respective indices and CBI values (p < 0.05). Out of all indices, the post-fire NBR-based severity map had the highest accuracy (73.08%) and the dNDVI had the lowest accuracy (65.38%). The RF and SVM supervised classification map products performed similarly with both of the products, having an 83.33% accuracy (Table 3 and Figure 7). Overall, the RF and SVM supervised map products outperformed all other products at assessing burn severity within the Shovel Creek fire (Table 3).

Burn Severity Relationship to Environmental Factors
According to the AVWC, 80.1% of the pre-fire vegetation cover in the Shovel Creek burn scar was conifer forest and 11.8% was mixed forest (conifer and deciduous) ( Table  4). In the Nugget Creek burn scar, 72.8% of the area was conifer forest while 25.5% mixed forest (Table 4). In both fires, conifer forest occupied the majority of the burn scars, confirming that both fires mostly burned conifer vegetation. We conducted several statistical tests to investigate the relationship of fuel type, slope, aspect, and burn scar site with burn severity classes (unburned, low, moderate, and high). First, we used type III sums of squares with all explanatory variables. We found that aspect and burn site (Shovel Creek and Nugget Creek) were the two statistically significant explanatory variables for burn severity classes (Table 5). We used AIC to determine which regression model best explained burn severity; of all possible models, the model with aspect and burn scar site as explanatory variables had the lowest AIC value and best fit. When we visualize the relationships between burn severity and aspect at both burn scar sites, we can see that aspect and burn severity classes have a stronger relationship at the Nugget Creek burn scar (R = 0.44, p = 0.02) than at the Shovel Creek burn scar (R = 0.095, p = 0.5 Figure 8). The slope had a statistically significant relationship with burn severity classes (R = 0.38, p = 0.049) only at Nugget Creek. Fuel type did not have a statistically significant relationship with burn severity classes at either site. These observations suggest that models for predicting burn severity using environmental variables are sitedependent, and likely unsuited for modeling risk of high burn severity at a larger scale.

Burn Severity Mapping
Several previous studies used traditional spectral index-based remote sensing methods to map burn severity in the boreal forest [57,58]. Epting et al. (2005) identified NBR as a better index for assessing burn severity than NDVI in the boreal forest [57]. Furthermore, they found the differenced indices (dNDVI and dNBR) outperformed post-fire indices (NDVI and NBR). The differenced indices are better than post-fire indices because they measure the amount of change over time [59]. Using 347 sites across 6 fires, Murphy et al. (2008) found that dNBR had a relatively weak correlation with CBI (R 2 adjusted, 0.11-0.64) [58]. In our study, we found that the post-fire NDVI and NBR indices outperformed the differenced indices at Shovel Creek, and the post-fire NDVI index outperformed the differenced NDVI (dNDVI) at Nugget Creek (Table 3). In fact, the NDVI-based burn severity map for the Nugget Creek burn scar had the highest accuracy (96%) between the two sites. Burn severity factors such as vegetation type and location impact the performance of the spectral indices and are not transferable between sites [29,58].
At the Shovel Creek burn scar, the RF and SVM (83%) outperformed all index products. Parks et al. (2019) used 19 fires to predict burn severity in Alaska using RF and had a R 2 value of 0.62 between the observed vs. predicted CBI, which is similar to the R 2 values of the dNBR index and CBI reported by Murphy et al. (2008) [29,58]. The supervised classification (RF and SVM) products outperformed the traditional index products at Shovel Creek as more CBI training plots (20) were available for this site compared to the Nugget Creek site (9 plots). At Nugget Creek, the traditional index-based burn severity map products (NDVI: 96%; dNBR: 89%) outperformed the RF and SVM products. We had only 9 CBI plots to train the RF and SVM classifiers for the Nugget Creek site, which may have caused the lower accuracy (RF: 67%; SVM: 67%) for the classified map products compared to the index-based map products. Field surveys of burn severity at remote sites are not always feasible due to logistical and funding challenges. So, when adequate field data to train a supervised classifier are not available, the index-based burn severity mapping would be ideal. Field surveys of burn scars to train supervised classifiers are time consuming and expensive but can result in accurate map products. The differences in the accuracy between the burn severity map products may also be the results of the spatial spread of the CBI plots collected. Our findings suggest that index-based methods can result in reliable burn severity map products, and offer insights on fire damage. However, if adequate field data are available to train supervised machine learning classifiers, then the classifiers can outperform index-based methods.
At the Nugget Creek fire, the SVM and RF mapped more of the moderate burn severity class and less of the low and high burn severity classes compared to the traditional indices (Table 6). This may be the result of the classifiers having limited training plots (2) to use to train the low burn severity class. At the Shovel Creek fire, the SVM and RF mapped more of the high burn severity class and less of the low burn severity class compared to the indices (Table 7). A common misclassification was CBI plots with values close to the edge of the class boundary (i.e., a CBI 1.49 being misclassified in the moderate class) (Tables S1-S12). In most products, a high burn severity was the most accurate class. This is likely a result of the high burn severity class being the most spectrally distinct from the other burn severity classes. In some applications, more than three burn severity classes may be needed. Overall, we found that both spectral indices and machine learning classifiers can adequately assess burn severity.

Burn Severity Relationship to Environmental Factors
The relationship between fire and environmental factors is complex and varies across a landscape [60]. Looking at the vegetation classes that were in the burn scar prior to the fire shows that the classes that contain conifer species (conifer and mixed forest) appear to be more likely to burn than other vegetation types and is consistent with a previous study [61]. Needles and branches of coniferous species are more flammable due to the presence of the resinous substance and are also distributed continuously from ground to treetop, allowing the fire to easily spread to the tree crowns [62]. In the Alaskan boreal forest, north-facing slopes are more conducive for black spruce growth because of colder soil conditions, lower solar insolation, and being less productive, whereas the dry southfacing slopes are more conducive for hardwood growth [63]. Hardwoods are less flammable fuel types (more moisture) while black spruce are incredibly flammable due to their lower moisture content, resin production, and low branches that promote crown fires [62]. Presence of diseased trees and dead trees can alter the relationship between burn severity and environmental factors. A case study on the basin complex fire in California suggested that a disease-caused increase in wildfire fuels can result in increased burn severity [64]. At Shovel Creek, there was not a significant relationship between burn severity and topography (aspect and slope). While at Nugget Creek, aspect had a strong correlation with burn severity, showing that the burn severity and topography relationship is site-dependent. The Nugget Creek fire was located far away from human infrastructure and little effort was put into containing the fire compared to the Shovel Creek fire that was suppressed aggressively as it was close to human infrastructure. For the Shovel Creek fire, the total suppression cost was over USD 25 million while at Nugget Creek the suppression cost was USD 56,000 [42,43]. The efforts to contain the fire may have caused the burn severity relationship with environmental factors to appear insignificant for the Shovel Creek fire. Therefore, the inclusion of topography to predict burn severity or create a burn severity probability map over a large landscape or region can be problematic.
Understanding the relationship between burn severity and environmental variables will help researchers and fire managers predict areas most at risk of high-severity fires, allowing for proper mitigation and restoration strategies that reduce the negative effects of fire. Topography has a complex relationship with fire that varies across landscapes and with daily climatic conditions, making it hard to model fire and severity risks [60,65]. Mapping burn severity helps managers see the effects that fires have on the landscape and allow managers to better understand changes to ecosystem functions and services caused by the fire. By creating accurate burn severity maps in the boreal forest that consider the environmental factors and their relationship to burn severity, we gain insight into what factors favor more severe impacts from fires. Areas with a high burn severity are of particular interest to land managers as they indicate areas with the largest amount of change to vegetation ecosystem properties due to fire. These areas are more susceptible to the colonization of invasive species, permafrost thaw, erosion, and potential vegetation type conversions [59]. Land managers are also interested in the changes to wildlife habitat and recreational activities caused by fire [59]. The relationship between burn severity, environmental factors, and vegetation recovery should continue to be investigated to advance the understanding of ecosystem responses to fire and ability to predict burn severity.

Conclusions
In this study, we were able to show that both supervised classification methods and spectral indices can accurately map burn severity in the boreal forest. At the Nugget Creek fire, with 28 CBI plots, the spectral indices NDVI (96%) and dNBR (89%) produced higher accuracy products than the supervised classification RF product (78%). While at the Shovel Creek fire, where we sampled 52 CBI plots, the RF and SVM products (83%) outperformed the spectral index-based burn severity maps. This shows that when we have more ground truth field data, the better the supervised classifier performs in assigning pixels to a burn severity class. In most cases, using pre-and post-fire images to create a differenced index did not improve the accuracy of assessing burn severity. Overall, both supervised classification and spectral indices create highly accurate burn severity maps in the boreal forest. Based on our study, coniferous forests appear more likely to burn than other vegetation types. We also found that the influence of environmental variables on burn severity is site-dependent and further research needs to be done to understand this relationship and explore additional factors in order to predict burn severity on the landscape. The products from this study will help fire managers identify areas of concern based on burn severity information that can then be integrated into management practices that will help mitigate and reduce the negative effects of fire.
Supplementary Materials: The following are available online at www.mdpi.com/2072-4292/13/10/1966/s1, Figure S1: CBI data sheet used to survey wildfire burn severity at the Shovel Creek and Nugget Creek burn scars. The data sheet was originally created by Key and Benson (2006) and modified by Barnes et al., (2020) for boreal forest in Alaska, Table S1: Shovel Creek Fire NDVI burn severity error matrix, Table S2: Shovel Creek Fire NBR burn severity error matrix, Table S3: Shovel Creek Fire dNDVI burn severity error matrix, Table S4: Shovel Creek Fire dNBR burn severity error matrix, Table S5: Shovel Creek Fire RF burn severity error matrix, Table S6: Shovel Creek Fire SVM burn severity error matrix, Table S7: Nugget Creek Fire NDVI burn severity error matrix, Table S8: Nugget Creek Fire NBR burn severity error matrix, Table S9: Nugget Creek Fire dNDVI burn severity error matrix, Table S10: Nugget Creek Fire dNBR burn severity error matrix, Table  S11: Nugget Creek Fire RF burn severity error matrix, Table S12: Nugget Creek Fire SVM burn severity error matrix.  Data Availability Statement: In this section, please provide details regarding where data supporting reported results can be found, including links to publicly archived datasets analyzed or generated during the study. Please refer to suggested Data Availability Statements in section "MDPI Research Data Policies" at https://www.mdpi.com/ethics. You might choose to exclude this statement if the study did not report any data.
Acknowledgments: Thanks to Jennifer Barnes, Lisa Saperstein, Alison York, Heidi Strader and Ga-Briella Branson for feedback on our research. Thanks to Colleen Haan for assisting in fieldwork. A special thanks to the Alaska Center for Conservation Science for creating and allowing access to the Alaska Vegetation and Wetland Composite map. The Alaska Vegetation and Wetland Composite map can be found at https://accscatalog.uaa.alaska.edu/dataset/alaska-vegetation-and-wetlandcomposite. Thanks to the Polar Geospatial Center for creating and allowing access to the Arctic DEM. The Arctic DEM can be found at https://www.pgc.umn.edu/data/arcticdem/. A big thank you to the anonymous reviewers and our editor Aguero Gui.

Conflicts of Interest:
The authors declare no conflict of interest. The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, or in the decision to publish the results.