1. Introduction
Burn severity is an important measurement of the effect that a wildfire has upon a landscape. Burn severity impacts vegetation mortality and soil nutrient composition, and causes increased runoff due to decreased infiltration resulting from soil hydrophobicity. The degree of burn severity can influence how long it takes for an ecosystem to recover and can change the composition of flora within an ecosystem. Because of these impacts, it is important for land managers to be able to assess the varying degrees of burn severity that result from fire events.
Burn severity can be measured differently depending on the interpretation of what it represents. Some studies have interpreted burn severity as a measurement of fire severity metrics and ecosystem responses [
1]. Other researchers interpret burn severity solely as the loss of organic matter in the soil or on its surface. The latter approach is used for Burned Area Emergency Response (BAER) assessments, which commonly use the delta Normalized Burn Ratio (dNBR) to derive a burn severity map, designated as the Burned Area Reflectance Classification (BARC). BARC maps generally provide adequate assessments of post-fire vegetation conditions and allow for rapid assessment of the immediate impacts of a fire event [
2]. BAER assessments commonly use the Composite Burn Index (CBI) for validation as it is heavily weighted towards the effects a fire has had on vegetation [
3].
Although commonly used, the CBI possesses a major limitation because it is based on ocular measurements as opposed to more quantitative field methods [
4]. This is because of how difficult it can be to take a significant number of accurate quantitative field measurements for each wildfire to calibrate spectral indices. However, this can lead to different assessments of CBI depending on the individual performing the assessment. Other measurements of burn severity provide a quantitative assessment of the level of burn severity, such as the amount of downed coarse wood, the number of live trees per unit area and ash depth.
The robustness of the dNBR Index has come Into question, with several studies suggesting that the index does not always provide accurate estimates and needs improvement [
5,
6,
7]. Miller and Thode [
7] found that dNBR performs poorly for pixels containing sparse vegetation because of dNBR detecting absolute change. dNBR detects change through the use of the whole image, and so a large change relative to the land cover within a given pixel may not be considered a large change in the context of the image as a whole. Different vegetation compositions affected by the same fire and possessing the same degree of burning can be assigned dissimilar dNBR values. To address this issue, RdNBR was proposed.
RdNBR is designed to assess the relative change instead of absolute change. This is accomplished with an additional step to the dNBR procedure in which the square root of the absolute value of the pre-fire NBR is used to calculate the quotient of dNBR. Miller and Thode [
7] found that RdNBR more accurately identified high-severity burns in areas of heterogenous vegetation composition. However, the proposed equation possessed its own issues, namely, that the square root used to calculate RdNBR produces large, difficult-to-interpret numbers.
An alternative burn severity index was proposed by Parks et al. [
8] and named the Relativized Burn Ratio (RBR). This index replaces the square root and absolute functions with the addition of 1.001 to ensure that all NBR values are greater than zero and altered in a way that preserves the level of NBR assigned to pixels. The RBR provides an index that estimates relative change without altering the output to the degree that the square-root in RdNBR does.
Although most studies using remote sensing data for assessing burn severity use Burn Ratios based on NIR and shortwave infrared (SWIR) [
9], a proposed alternative to the Burn Ratios is to include land surface emissivity (LSE). The inclusion of LSE adds a surface characteristic that is separate from incoming solar radiation for the assessment of burn severity [
10]. Quintano et al. [
11] found that LSE-enhanced vegetation indices resulted in better burn severity estimates when compared to standard spectral indices, with an increase of about 16% when used to map burn severity in Sierra del Teleno, Spain. However, LSE-enhanced vegetation indices can be difficult to generate as they require the LSE and temperature to be differentiated from surface radiance and atmospheric conditions.
Spectral mixture analysis (SMA) has also been proposed as an alternative to Burn Ratios. SMA is a technique that uses the spectral reflectance of the ‘pure’ spectral response of land cover, referred to as endmembers, to determine the proportion of a mixed pixel belonging to different cover types. This is accomplished by using the endmembers to analyze a pixel and determine the degree to which the radiance from a mixed pixel agrees with each endmember [
12]. Currently, SMA is not commonly used as a burn severity estimation technique. Studies that have compared spectral indices and SMA for estimating burn severity have shown the two approaches to be analogous [
2,
13]. However, SMA has not been shown to consistently outperform dNBR, as seen in Veraverbeke and Hook [
13], which compared SMA to several spectral indices (NBR, dNBR, RdNBR) for burn severity estimates. They found that dNBR outperformed SMA but also noted that both approaches performed adequately and that SMA has the benefit of providing transferable quantitative data and does not need field data for calibration.
Recently, the Sentinel-2 sensor system was launched by the European Space Agency (ESA). The system contains additional red-edge bands that facilitate the calculation of more indices that may be useful for burn severity estimates. Fernández-Manso et al. [
14] used Sentinel-2 imagery to calculate several red-edge indices, as well as several more ‘traditional’ spectral indices, for estimating burn severity. They found that two of the red-edge indices outperformed the other indices that were tested, showing the potential for red-edge indices to aid in the assessment of burn severity. However, the capabilities of red-edge bands for assessing burn severity have not been fully explored and further research is needed.
Although most burn severity studies that use remote sensing to assess severity rely on field-measured CBI [
5,
6,
7], few attempts have been published to determine which indices can be used for assessing more quantitative measurements of fire effects such as tree mortality by basal area and number of trees, char height and surface char. CBI is useful for the rapid ocular assessment of burn severity but is limited and may vary depending on the subjective judgement of the induvial assessor in the field. Saberi [
15] found that CBI estimates corresponded best to field measurements of tree canopy attributes but did not correspond as well to other field measurements like the deep char index. The authors suggest that spectral indices can be used to map CBI, which, in turn, can be used to map various fire effects (particularly those related to tree canopy attributes) using regression analysis. Hudak et al. [
16] attempted to relate several Landsat 5 TM-derived burn indices to fire effects, finding that none of the indices were highly correlated with the fire effects.
The objective of this study is to test the ability of several indices to estimate field-measured fire effects using Sentinel-2 imagery. At present, the most commonly used burn indices, such as dNBR, RdNBR and RBR, are limited to broad near- and shortwave infrared band intervals. Limited research has been published that examines comparisons of red-edge bands to traditional data to calculate burn indices. In Fernández-Manso et al. [
14], only post-fire indices were calculated, so this research aims to determine the effectiveness of using the delta index from pre- and post-fire imagery, as well as the post-fire indices. Additionally, alterations to the commonly used burn indices are made in which the narrow NIR band is replaced with a red-edge band to generate the indices and assess whether this substitution results in a more robust index.
By testing a broad range of indices, this paper seeks to determine the appropriate indices for estimating field-measured fire effects for two fires in the Greater Yellowstone Ecosystem. Given the limited availability of red-edge bands in free, publicly available data sets, this study contributes to an enhanced understanding of their utility for fire effects and associated estimates. Additionally, the findings from this study build on a limited body of knowledge regarding the specific effectiveness of Sentinel-2 red-edge bands for assessing post-fire effects. These bands are not present on similar publicly available sensors, such as Landsat 8 and 9.
3. Results
3.1. Descriptive Statistics for Spectral Indices
The spectral indices were generated for a 6.5 km buffered area around the Berry Fire and for the 2 km buffered area of the Maple Fire.
Table 3 shows the descriptive statistics for each spectral index calculated for the Berry Fire grouped into five categories: post-fire normalized red-edge indices, difference normalized red-edge indices, difference normalized Burn Ratios, other Burn Ratios and other indices (
n = 1,478,229 pixels). The non-Burn Ratio indices are defined in
Table 2. In
Table 3 the PF following the index acronym signifies post-fire, while the d before the index acronym signifies the result of the delta between pre- and post-fire imagery. For the Burn Ratios, the B# following the index signifies which red-edge band was used for NIR in the NBR equation.
Table 4 shows the same breakdown of indices for the Maple Fire (
n = 970,055 pixels). Note that the CLre, MSRren and PSRI indices are also red-edge indices, which are not normalized, and so are grouped under other indices, whereas the GNDVI indices are normalized but are not red-edge indices and so are also under other indices.
3.2. Correlation Results
To determine whether the indices and field measurements were related, Pearson correlations between the field measurement and spectral indices were assessed (
Table 5;
n = 27). Coarse wood percent cover and coarse wood mass were correlated with the most indices, with each field measurement possessing strong correlations with twenty-two spectral indices. Other field measurements that possessed strong correlations with spectral indices were post-fire dead PICO stumps, ash depth, coarse wood volume and percent cover of ghost logs.
Coarse wood mass possessed the strongest positive relationship with the spectral index, with NDre2_PF having a correlation of 0.886. The NDre2_PF index also possessed a strong positive correlation with coarse wood percent cover and coarse wood volume. Coarse wood mass also possessed the strongest negative correlation of −0.811 with both dNDre2 and GNDVI_PF.
dNDre2_PF was significantly related to the most field measurements, with six field-measured fire effects being strongly correlated with this index. The RdNBR_B8a index was significantly correlated with only one field measurement, coarse wood percent cover. Several indices were found not to possess strong correlations with any field measurements, including dPSRI, dNDVIre2n, dNDVIre3n, NDVIre2n_PF, NDVIre3n_PF and RdNBR_B5.
3.3. Regression Results
Within the statistical analysis software JMP 14, all-possible-models (i.e., best subsets regression) was used to construct models for predicting field measurements using the spectral indices. All-possible-models regression tests all possible subsets of the predictor variables, and returns models that contain one variable, two variables, etc., along with their summary statistics. This allows the researcher to quickly assess the performance of every combination of predictor variables to determine which models performed best. With a sample size of twenty-seven field plots, a maximum of three independent variables were allowed for model construction. Each model was assessed based on the significance of its independent variables and on the variable multicollinearity. The
p-value of each model covariate had to be less than 0.05 for the model to be accepted. Multicollinearity was assessed using the Variance Inflation Factor (VIF), where all input variables had to possess VIF values of < 10. The models for one, two and three input variables that met these criteria and possessed the highest R
2 for a given field measurement are reported in
Table 6.
Of the field measurements, coarse post-fire dead PICO stumps, coarse wood percent cover, coarse wood volume, coarse wood mass, ash depth and percent cover of ghost logs possessed models with R2s above 0.6. Of these, coarse wood mass achieved the highest R2 (0. 847), followed by Coarse wood volume (R2 = 0.833). Ash depth had the lowest R2 (0.636), with post-fire dead PICO stumps possessing the second lowest (R2 = 0.663) of the variables with R2s greater than 0.6.
Of the single-variable models, Coarse wood volume possessed the highest R2 (0.784) with NDre2_PF as the input variable. The single-variable model for post-fire dead PICO density performed the worst, with an R2 of 0.174 when dNBR_B6 was used as the input variable. For the two-variable models, coarse wood mass performed the best, with an R2 of 0. 847, and initial regeneration of post-fire aspen density performed the worst, with an R2 of 0.249. Of the three-variable models, coarse wood mass possessed the highest R2 (0. 847), whereas initial regeneration of post-fire aspen density had the lowest R2 (0.448). Several field measurements did not have any models that met the p-value and/or VIF criteria, and so their variables are reported as none and their statistics as N/A.
The PRESS statistic was used to determine which of the models generated for each field measurement possessed the best predictive power. This statistic determines model performance by leaving one sample out at a time to determine how well the data predict the left-out sample [
20]. The model with an R
2 > 0.60 that yielded the highest PRESS R
2 for each field measurement was determined to be the best model for predicting the field-measured fire effect. Standard least squares was then used to plot these models and determine their corresponding prediction equation (
Figure 4).
Of the variables used to construct these models, dNDVIre3n, NDre1_PF and dNBR_B5 were the only indices used in more than one model. A total of three Burn Ratio indices were used across all the models, whereas ten red-edge indices were used. Of the Burn Ratio indices, all the selected input variables for the models used red-edge bands in place of narrow NIR.
4. Discussion
4.1. Correlations between Spectral Indices and Field Measurements
When examining the correlation between the various spectral indices and the fire effects measurements, the strongest correlations were found to involve indices generated using band 5 for all field measurements. Of the red-edge indices, the only indices to possess strong correlations with field measurements not generated using band 5 were CLre_PF and dCLre. Of the red-edge NDVI indices, both post-fire and difference NDVIre1n outperformed their NDVIre2n and NDVIre3n counterparts. Previous research has reported similar findings, with red-edge indices generated using the band closest to red, band 5, outperforming the other red-edge indices, as well as more traditional spectral indices, for burn severity detection [
14,
21].
4.2. Spectral Indices’ Ability to Estimate Field Measurements
Although many studies have estimated burn severity using spectral indices [
5,
6,
7], only a few have attempted to estimate field-measured fire effects using these indices [
16,
22,
23]. Although field-measured fire effects are not as commonly assessed because of the time-intensive nature of these measurements, they provide valuable ecological information that can be used in fire recovery efforts. However, previous research has shown little relationship between Landsat-derived burn indices and field-measured fire effects [
16,
22,
23]. This can be attributed to a number of variables, including the spatial resolution of Landsat images, the radiometric resolution of the sensors used during the time of these studies being inadequate to capture the slight variations in radiance, and the lack of spectral bands in the red-edge region. Previous research has shown slight improvement in the performance of Sentinel-2 NBR-based indices when compared to Landsat 8 NBR-based indices [
24,
25]. However, this research was limited to indices that could be calculated by both sensor systems, which eliminates the use of red-edge indices. The results of this research suggest that several field-measured fire effects can be estimated using the Sentinel-2 sensor constellation, and the use of red-edge indices improved Sentinel-2’s performance of this task.
Of the fire effects estimated by the spectral indices, those related to tree canopy characteristics resulted in the best estimates. This agrees with the findings of Saberi [
15], who found that CBI and the three primary burn severity indices (dNBR, RdNBR, and RBR) were more highly correlated with tree canopy fire effects than they were with other effects. Additionally, we found that the red-edge spectral indices explained approximately 64% of field-measured variation in ash depth and 72% of the variation in percent cover of ghost logs.
Of the best-performing models, coarse wood mass and coarse wood volume possessed the highest R2 values, at 0.837 and 0.833, respectively. Both models were generated using the same red-edge indices, with neither of the models requiring a Burn Ratio index. These high accuracies, combined with the use of solely red-edge indices, suggest that further research into the utility of using red-edge indices for estimating and mapping various fire effects should be explored.
4.3. Performance of Red-Edge Bands and Indices
With the launch of the Sentinel-2 sensor constellation, red-edge bands for index generation have become freely available. Fernández-Manso et al. [
14] show that red-edge indices can accurately discriminate between levels of burn severity and found that indices generated using Sentinel-2 band five were most suited to this task. Our results show that indices that were generated using band five were included in all six of the best-performing models, suggesting agreement with other research [
14,
21,
26].
Interestingly, of the best-performing models, only one contained a non-red-edge index (percent dead PICO stumps). The two best-performing models (coarse wood volume and mass) both only used red-edge indices, and both achieved R
2 > 0.8. The indices used in these models relied on bands 5, 7 and 8a. This suggests that red-edge indices, which have shown promising results in estimating burn severity [
14,
21,
27], may also be useful for estimating fire effects.
4.4. Sources of Uncertainty
Although these results are promising, there are a few sources of uncertainty. The largest sources of uncertainty are the sample size of the field data (n = 27) as well as the limitation of the data to a single ecosystem. As a result, the data do not provide a comprehensive explanation for the tested dependent variables, and our results should be considered preliminary.
The field measurements were collected for circular subplots measuring 30 m in diameter, but the spatial resolution of the Sentinel-2 data was 20 m. We used the average of the pixels that fell within a 30 m buffer to address this issue, but some of these pixels lay partially outside the buffer and other pixels were excluded because too small a proportion of these pixels fell within the buffer. This may lead to the spectral reflectance of the pixels corresponding to these measurements only partially representing the measured conditions and/or including reflectance from outside the buffer in the average.
Additionally, a geolocation error between images can create uncertainty in index calculation and value-to-points extraction. For Sentinel-2, this error is less than 1 pixel in most cases, with errors exceeding this threshold primarily because of coarse corrections. No coarse corrections were documented for any of the images used in this analysis; however, a single pixel error could potentially impact the results.
The use of samples from two separate fires is also a source of uncertainty. These fires started and ended around the same time (summer 2016 to fall 2016) and were both located in the Greater Yellowstone Ecosystem, possessing similar vegetation and landcover. However, different image acquisition dates and a more limited number of samples for the Maple Fire could create uncertainty in the results. However, because of the lack of snow cover in the November imagery and the evergreen forests that make up the majority of the in-scene vegetation for both fires, we do not expect the difference in acquisition dates to considerably influence our results. After examining the residuals for the primary regression models, it was noted that overall, the two fires’ residuals were similarly distributed, except in the case of post-fire dead PICO stumps, where the upper end of the predicted values for the Maple Fire possessed larger negative and positive residuals than any residual for the Berry Fire. This can be explained by the limited sampling for both fires, which ideally would have at least thirty sample plots per fire. Because of this limited sample size, these results should be considered preliminary and further research should be conducted to determine their validity.
5. Conclusions
This study assessed the ability of spectral indices, both traditional and red-edge-based, to estimate various field-measured fire effects. Several fire effects were accurately estimated using a combination of red-edge and Burn Ratio indices and multivariate regression. These fire effects included post-fire dead PICO stumps, coarse wood percent cover, coarse wood volume, coarse wood mass, ash depth and percent cover of ghost logs. Of the indices generated, the most useful for estimating these fire effects were red-edge indices, especially those generated using Sentinel-2 band 5 (0.6955–0.7134 μm).
Despite the field data being of limited sample size and from a single forested ecosystem, this research shows that red-edge indices have potential for mapping various fire effects when used in combination. Further, although the methodology used to calculate the indices and evaluate the comparisons is not novel, this study contributes to a growing body of literature, emphasizing the improved performance of red-edge-based indices over traditional near- and shortwave infrared-based indices. Future research should incorporate a larger validation data set and extend to other ecosystems, as these results are preliminary.