Wildfires are a major cause of disturbance in North American boreal forests, burning an annual average of approximately 10,000 km2
in the 1960’s and more than 31,000 km2
in the 1990’s [1
]. There is high interannual variability in the area burned [2
], and boreal fires can exert a considerable influence on the species composition, diversity, and stability of many forest ecosystems [3
]. Fires can cause large fluxes of carbon within an ecosystem [4
] and release a tremendous quantity of carbon to the atmosphere [5
]. This can lead to a boreal forest becoming a significant net carbon source, particularly when combined with severe insect disturbances [2
]. The near-complete mortality of forest vegetation after boreal fires in North America has modified wildlife habitats, including the elimination of lichen that caribou forage upon [6
]. In colder permafrost environments, the active layer depth also often shows a long-term increase after a fire, due to the removal of the insulating surface organic layer [7
]. Of particular concern is the projected increase in the area, frequency, and severity of fire under a changing climate, and its potential to exacerbate changes in vegetation succession and ecosystem function [8
Burn severity is a term used to describe the magnitude of ecological change caused by fire [11
]. Knowledge of burn severity is important for predicting the post-fire vegetation community [12
], assessing the availability of residual seed sources to promote regrowth [13
], predicting long-term forest successional patterns [13
], and determining changes in bird abundance [14
]. Boreal burn severity has been assessed in North America using a variety of methods that span spatial scales from plot measurements [15
], to high resolution air photos [18
], airborne hyperspectral imagery [19
], airborne Lidar [20
], and satellite imagery for mapping at landscape to regional scales [21
A commonly applied field method for assessing burn severity in forests is the Composite Burn Index (CBI) [22
]. The CBI provides a numerical index of vegetation damage from fire by integrating visual estimates of damage, stratified by five vertical layers (substrate, herbs and low shrubs < 1 m, tall shrubs and saplings 1–5 m, sub-canopy trees, and canopy trees), using a scale of 0 (unchanged) to 3 (high severity). The CBI provides a rapid means of assessing burn severity that integrates a variety of ecologically significant burn severity indicators. Some limitations and challenges in applying the CBI include: (1) the qualitative nature of its assessments, which could vary among observers [23
]; (2) the difficulty and cost involved in accessing remote burn sites on the ground; (3) the potentially large number of plots required to represent the range of conditions in a large, heterogeneous burned area; (4) a need to infer pre-fire fuel conditions for assessing consumption and mortality [23
]; and (5) a limited ability to quantify organic layer consumption [24
], which is an important measure of severity in northern boreal forests. Despite these limitations, the CBI is a widely used field method in burn severity assessment [23
A wide range of satellite-based approaches have also been developed to map burn severity over larger forested areas [11
]. A common method is to compute the difference between the pre-fire and post-fire Normalized Burn Ratio (NBR), which is based on the near-infrared and shortwave infrared bands from 30 m resolution Landsat imagery [21
]. The difference in the NBR (dNBR) has been effective for mapping burn severity in Canadian forests [16
], but less so for Alaskan spruce forests [29
], where the severity strongly depends on the depth of burning in the surface layer [29
]. Since dNBR only measures a spectral response change, it is the calibration with the plot-based CBI that provides the ability to transform larger-area dNBR values into maps of varying burn severity [22
]—an approach that has been applied across different regions of boreal forest [21
]. For example, a non-linear relationship between CBI and dNBR was derived for several boreal fires within Western Canada, by Hall et al. [16
]. Some considerations for applying this CBI-Landsat calibration approach include: (1) the requirement to identify CBI plots that should be relatively homogeneous at a 90 m, 3-by-3 Landsat pixel scale; (2) the need for cloud-free, post-burn Landsat imagery that closely matches the CBI sampling date; and (3) an understanding of the complex physical relationship between CBI’s multiple, integrated measures of severity and spectral changes, measured using satellite-based indices [11
Driven by a need to acquire highly detailed information over larger areas, relative to field plots, unmanned aerial vehicles (UAVs) have been growing in popularity as a means of conducting inexpensive, on-demand remote sensing that can bridge measurement scales between field plots and satellite imagery [32
]. While there have been a number of applications of thermal sensors on UAVs to map actively burning fires [26
], there has been a limited use of UAVs for assessing vegetation burn severity. One application involved the use of a multispectral imager mounted on a large, military-grade UAV system, to generate NBR maps in the Western US [33
]. More recently, a post-fire digital terrain model (DTM) derived from a UAV survey was differenced with a pre-fire DTM from airborne LiDAR, to estimate the depth of surface burning in a tropical peatland [34
]. The purpose of our study was to assess the potential for using a small, multicopter UAV to map indicators of boreal forest burn severity at landscape scales, and to up-scale this information to calibrate Landsat spectral indices, thus representing biophysically meaningful measures of severity over larger areas. As a result, we also conducted a limited comparison of plot-based CBI values with both UAV and Landsat indices of burn severity.
The separability statistic (S
) derived for each UAV index indicated that the Excess Greenness (ExG) index (Figure 3
b) provided the best discrimination of the burned and green vegetation samples (S
= 2.2), followed by NormG (S
= 1.9) and NormG-R (S
= 1.0). We therefore applied the binary classifier to ExG and assessed the resulting map accuracy (Figure 3
c) using the reference samples. The natural breaks (Jenks) classification produced a very high accuracy, with only 68 of 5268 (1.3%) burned samples and 5 of 3794 (0.13%) unburned samples being classified incorrectly. This high accuracy is the result of the two classes’ ExG mean values being separated by more than two standard deviations (S
= 2.2), and the fact that reference samples were selected with a relatively high confidence and did not include borderline cases, such as partially scorched crowns.
The two-sample Kolmogorov-Smirnov test indicated that the probability distributions of UAV and Landsat index values from the model calibration (70%) and validation (30%) datasets were significantly similar (p-values 0.32–0.64). A non-linear, exponential function in the form Y = abx was found to be suitable for modeling the relationships between the two Landsat severity indices (post-NBR and dNBR) and three UAV severity indicators (percent green fraction, percent green fraction above 5 m, and percent char fraction).
shows the scatterplots, best-fit exponential equations and lines, coefficients of determination (R2
), and root mean squared errors (RMSE) for these six modeled relationships. The post-NBR showed stronger relationships with both the green tree fraction (Figure 4
b; Adjusted R2
= 0.81) and char fraction (Figure 4
c; Adjusted R2
= 0.79), while the dNBR was more closely related to the total green fraction (Figure 4
d; Adjusted R2
= 0.69). Initial visualizations of the scatterplots in Figure 4
may suggest that there is a possible pattern of increasing residual error with changing post-NBR or dNBR values, but standardized residual plots proved otherwise. Example standardized residual plots for Figure 4
e,f, where this pattern appeared most obvious, do not depict a pattern of increasing variance across the range of dNBR (Figure 5
a,b, respectively). The Wilcoxin signed-rank test also showed no statistical differences (p
> 0.05) in the predicted versus actual green and char fractions from the 30% validation sample (n
= 167), when these were predicted using the models developed from the 70% training sample (n
= 388) (Table 3
). However, green tree fraction predictions were statistically different (p
< 0.05) from either Landsat predictor (Table 3
Non-linear models for predicting the overall CBI using the Landsat indices and UAV severity indicators are summarized in Table 4
. The post-NBR and dNBR Landsat indices could explain similar amounts of CBI variance (Adjusted R2
= 0.53–0.59), with relationships that were somewhat weaker (Adjusted R2
= 0.82–0.85) than those reported in [16
], based on dNBR. The results also indicated the UAV green fraction indices were comparable to the Landsat indices in their ability to predict the CBI.
Previous studies of boreal forest burn severity have shown that the Landsat dNBR index generally demonstrates the strongest relationship to the field-based CBI designed to measure overall burn severity [21
]. The NBR exhibits a large drop after burning, owing to the removal of leafy vegetation, which decreases NIR scattering, SWIR water absorption and canopy shadowing [22
]. In this study, we similarly found that dNBR was effective for predicting CBI (Table 4
), but also demonstrated that Landsat indices were even more closely related to specific indicators of burn severity derived from UAV mapping (Figure 4
). The post-NBR and dNBR indices could each explain a similar amount of the variation in the UAV-derived green vegetation fraction, while the post-NBR was a better predictor of the UAV char fraction. These Landsat-UAV relationships are consistent with [38
], where Landsat severity indices were most highly correlated to the green crown fraction and total charred ground surface from a set of 32 fire effect field measures from Alaska.
It is important to note that the strongest modeled relationship was for the UAV green tree fraction predicted using the post-NBR (Figure 4
b). Despite the strength of this relationship, the high degree of flattening at low post-NBR values makes it relatively less useful as a means to calibrate the NBR, compared to the total green fraction or char fraction relationships. The major reason for this is that the large proportion of crown fire within our study sites led to a majority (57%) of 30 m Landsat pixels containing less than five percent green tree crown (Figure 6
, FP15_21). As a result, this UAV indicator had a relatively minor influence on 30 m post-NBR values over most of the post-NBR range.
Superimposing a raster of the regression model residuals over the UAV orthomosaic can provide insight into the specific reasons for scattering in the model results shown in Figure 4
. An examination of the residuals for the first plot (post-NBR vs. green fraction) shows that values are often spatially clumped, rather than random. While there don’t appear to be consistent causes for either model over- or under-prediction of the UAV green fraction, some patterns are apparent. For example, some areas where the actual UAV-derived green fraction is higher than the Landsat-predicted fraction (i.e., underestimation) are associated with relatively open and unburned canopies that contain fewer shadows, which may otherwise obscure unburned ground vegetation (Figure 6
, FP15_20). This can result in a higher UAV-mapped green fraction compared to areas where shadows obscure more of the ground surface. By contrast, some areas where the actual green fraction is lower than the amount predicted (i.e., overestimation) correspond to standing scorched tree canopies with brown/orange needles (Figure 6
, FP15_19), or with a ground layer of dropped needles (Figure 6
, FP_22). These canopies encountered a less severe fire at the surface and lower canopy strata, but it was severe enough to scorch the needles through proximal heating, without direct flame contact. The Jenks classifier was used to create a binary, burned versus green vegetation classification, whereby scorched crowns were not treated as a separate, less severe vegetation response class.
Most Landsat-based studies of boreal burn severity have applied the dNBR, which is a bi-temporal index representing spectral changes from pre- to post-fire conditions [16
]. However, investigations in Alaska, where black spruce is the predominant tree species, have shown that the post-NBR can be more highly correlated than the dNBR to field-measured fire effects [29
]. The choice of a uni- or multi-temporal index to characterise severity should be based primarily on the accuracy for measuring fire effects, but may also reflect other considerations. For example, an index that measures spectral changes such as the dNBR, should be theoretically preferable for studies of burn severity, since, by definition, this term describes the degree of ecosystem change resulting from fire. The CBI, which was developed as a field measure of severity for calibrating dNBR, is similarly based on ecological changes, although pre-fire conditions must often be inferred [22
]. A bi-temporal index also permits a more accurate mapping of the extent of burning (a pre-requisite for determining severity), by avoiding commission errors over unvegetated areas that can result from only using a post-fire index. In our study, the post-NBR and dNBR were comparable in their ability to predict UAV-based green vegetation fractions, while the post-NBR was a better predictor of the UAV char fraction. The better performance of the post-NBR may be due, in part, to the char indicator having little or no dependence on the pre-burn conditions that influence the dNBR values. The inclusion of a pre-burn NBR to generate the dNBR has the potential to add spectral variability unrelated to fire, due to image differences in solar illumination, atmosphere, phenology, and spatial registration [31
], or differences in the pre-fire forest condition [48
]. Based on these results, a potential strategy for upscaling the UAV indicators using Landsat imagery may be to use the dNBR for mapping the extent of fires and green vegetation fractions, followed by the post-NBR to estimate char fractions.
The strong relationship between the dNBR/post-NBR and the UAV-derived severity indicators, permits the translation of the Landsat indices into measures of burn severity with a greater biophysical underpinning. Figure 7
shows the pre- and post-fire Landsat image and predicted residual green vegetation and char fractions based on the NBR-UAV non-linear regression functions, for a larger 60 km2
area within our study region. Considering the diversity of the fuel types sampled (Table 1
) and the results from the hold-out validation sample, these regression relationships should be applicable to other areas within in the post-burn Landsat image that burned under similar conditions. However, the relationships may not be valid for other images captured at different post-fire regeneration intervals, and should also be tested for their extendibility over a wider range of forest ecological conditions [16
The area of charred surface was derived from UAV mapping to indicate the extent of burning in the surface organic layer. A second potentially useful metric of surface burning not investigated here, is the presence of light coloured ash, which indicates that the fire has completely combusted organic material, such as large woody debris. The amount of ash cover has been related to surface fuel consumption in the boreal forest [19
], and could thus provide a valuable UAV-based measure of burn severity in the surface layer. The extent of light ash could potentially be derived using a colour index similar to our Char Index, but by identifying a bright, rather than dark, colourless surface. However, note that ash, unlike char, could be challenging to separate from other surfaces such as bright rock, based on colour alone. A very high resolution digital surface model derived from UAV photogrammetry and/or image segmentation techniques could potentially assist with this [39
]. Ash cover can also be rapidly removed by wind and rainfall, so UAV surveys may need to be conducted relatively soon after burning. The mapping of light ash would be a more important consideration for boreal burns containing larger proportions of ash than were encountered in our study sites [50
], and where it would be more likely to influence Landsat burn indices at a 30 m scale [42
In this study we made limited use of the UAV-based vegetation height model to separate the portion of residual green vegetation that lies above 5 m in the tree canopy. A vegetation height model could be further exploited to stratify UAV severity metrics according to the five CBI height classes, in an attempt to mimic specific factors that form the overall CBI rating. UAV-based colour and height information could also be investigated, to quantify more complex indicators of burn severity, including the area of scorched needle-leaf crowns, char height [51
], density of standing and downed boles with no foliage, and the presence of exposed mineral soil arising from the complete combustion of surface organic material. The calculation of these metrics may require higher resolution UAV photos with a larger overlap than those collected in this study [39
], to facilitate characterizing the detailed structure of residual dead and living trees. The goal of follow-on research should therefore focus on a more comprehensive UAV-based protocol for plot-scale assessments of burn severity, which could include select field-measured attributes, such as depth of surface burning [15
], that may be difficult to address using UAV photogrammetry alone. If successful, the major advantage of such a hybrid protocol would be the ability to conduct more quantitative, repeatable, and objective burn severity assessments with a closer physical linkage to satellite-based indices, to allow for robust up-scaling.
Some other potential applications of UAV-based, SfM modeling for studying boreal fires include: (1) providing reference data for training and validating burned area mapping algorithms that use Landsat’s or higher resolution satellite sensors; (2) mapping the charred standing bole density to help guide post-fire salvage logging efforts; and (3) quantifying long-term regeneration rates and structural changes in recovering burned forest, similar to previous studies that have used LiDAR [20
]. Because of the relatively limited extents that can be surveyed using UAVs, relative to manned aircraft or satellite-based sensors, these applications would have to examine sites that had been carefully chosen to represent larger-area conditions.
Finally, the development of UAV-based burn severity indices in this study was not without certain limitations and challenges. Here we highlight several:
Our surveys were conducted under clear, sunny conditions that caused strong shadows to cover much of the ground in areas with dense residual tree cover. Shadowing could be minimized by conducting surveys in light, overcast conditions.
Surface char was completely obscured in some areas, due to a dense layer of scorched conifer needles that dropped from the above canopy.
We did not validate the accuracy of the tree canopy height model and assumed that it was sufficiently accurate for broadly separating ground-level vegetation and tree crowns that lie above 5 m. For reference, Wallace et al. [52
] created UAV/SfM-based canopy height models that could estimate 4.7–16.2 m eucalyptus tree heights, with a root mean square error of 1.3 m.
The binary Jenks classifier, used for separating both green vegetation and shadows, was found to be simple and effective, yet may not be as accurate as more advanced classifiers or those that include a third, less severe vegetation response class for brown/orange scorched canopies.
A rigorous validation of very high resolution, UAV-derived indicators of burn severity is challenging, because of the requirement to collect reference data that is at least as detailed and spatially precise. The approach used in this study for mapping green vegetation and char was to relate photo-interpreted conditions, based on ground and UAV photos, to a georeferenced UAV orthomosaic. However, the validation of more advanced and structural measures of burn severity from UAV photogrammetry will require precisely georeferenced field-based measurements.
We used a consumer-grade RGB camera mounted on a UAV to capture images in JPEG format, that provided simple digital number intensity values. Note that such data have an unknown relationship to scene radiance [53
], and would be impacted by any changes in solar illumination during a survey. While we found that such imagery is suitable for separating broad and distinct classes, such as green vegetation and charred surface, more detailed vegetation characterization could benefit from using miniaturized multi- or hyper-spectral instruments that provide additional spectral information, and where radiance is normalized based on incident light sensors. Note, however, that such sensors currently provide coarser pixel resolutions (e.g., 1–2 megapixels) that may make the SfM modeling of a tree structure challenging.