Evaluating the Near and Mid Infrared Bi-Spectral Space for Assessing Fire Severity and Comparison with the Differenced Normalized Burn Ratio

: Fire severity, deﬁned as the degree of environmental change caused by a ﬁre, is a critical ﬁre regime attribute of interest to ﬁre emissions modelling and post-ﬁre rehabilitation planning. Remotely sensed ﬁre severity is traditionally assessed by the differenced normalized burn ratio (dNBR). This spectral index captures ﬁre-induced reﬂectance changes in the near infrared (NIR) and short-wave infrared (SWIR) spectral regions. This study evaluates a spectral index based on a band combination including the NIR and mid infrared (MIR) spectral regions, the differenced normalized difference vegetation index with mid infrared (dNDVI MID ), to assess ﬁre severity. This evaluation capitalized upon the unique opportunity stemming from the pre- and post-ﬁre airborne acquisitions over the Rim (2013) and King (2014) ﬁres in California with the MODIS/ASTER Airborne Simulator (MASTER) instrument. The ﬁeld data consist of 85 Geometrically structured Composite Burn Index (GeoCBI) plots. In addition, six different index combinations, respectively three with a NIR–SWIR combination and three with a NIR–MIR combination, were evaluated based on the optimality of ﬁre-induced spectral displacements. The optimality statistic ranges between zero and one, with values of one representing pixel displacements that are unaffected by noise. The results show that the dNBR demonstrated a stronger relationship with GeoCBI ﬁeld data when ﬁeld measurements over the two ﬁre scars were combined than the dNDVI MID approaches. The results yielded an R 2 of 0.68 based on a saturated growth model for the best performing dNBR index, whereas the performance of the dNDVI MID indices was lower with an R 2 = 0.61 for the best performing dNDVI MID index. The dNBR also outperformed the dNDVI MID in terms of spectral optimality across both ﬁres. The best performing dNBR index yielded median optimality statistics of 0.56 over the Rim and 0.60 over the King ﬁre. The best performing dNDVI MID index recorded optimality values of 0.49 over the Rim and 0.46 over the King ﬁre. We also found that the dNBR approach led to considerable differences in the form of the relationship with the GeoCBI between the two ﬁres, whereas the dNDVI MID approach yielded comparable relationships with the GeoCBI over the two ﬁres. This suggests that the dNDVI MID approach, despite its slightly lower performance than the dNBR, may be a more robust method for estimating and comparing ﬁre severity over large regions. This premise needs additional veriﬁcation when more air- or spaceborne imagery with NIR and MIR bands will become available with a spatial resolution that allows ground truthing of ﬁre severity. ﬁre severity. Field plots were sampled in order to represent the observed range of ﬁre severity in mixed conifer forests. The selection of the ﬁeld plots was constrained by road and trail accessibility, and the centre of the ﬁeld plot was always at least 100 m away from the nearest road or trail. Field locations were recorded using a Global Navigation Satellite System device.


Introduction
Wildfires in the Western United States have intensified during the recent decades with respect to their size, severity, and frequency [1][2][3][4][5]. Analysing the effects of fires on ecosystems by assessing how fires impact vegetation recovery and succession is critical to forest management [6,7]. Fire severity is a fire regime attribute that strongly influences postfire ecosystem trajectories [7]. Fire and burn severity are often used interchangeably [8].
burned and unburned areas significantly increases as we shift from the SWIR to the MIR spectral region.
This allows for better discrimination between burned and unburned pixels. MIR imagery is currently unavailable from space at spatial resolutions of less than 100 m. This moderate resolution is required for a direct comparison with field measurements of fire severity. Our study made use of the unique opportunity stemming from pre-and postfire airborne acquisitions at a spatial resolution allowing comparison with field data from the MODIS/ASTER (MASTER) sensor over two fires in California. By doing so, our study is the first to assess the NIR-MIR bi-spectral space for assessing fire severity. In addition, we compare the performance of bi-temporally differenced NDSI with NIR and MIR infrared, the differenced normalized difference vegetation index with mid infrared (dNDVIMID), with the performance of the most commonly used fire severity spectral index, the dNBR.

Airborne Imagery and Processing
There are currently no spaceborne sensors with a MIR band at a spatial resolution of approximately 30 m. This study capitalized upon the unique opportunity stemming from the pre-and post-fire airborne acquisitions over the Rim and King fires with the MASTER instrument. The MASTER scanner acquires data over the spectral range from 0.4 µm to 13 µm in 50 spectral bands. The spatial resolution depends on the flying height and was approximately 35 m over the Rim and King fires [33]. We used Level 2 surface reflectance data in this study. The MASTER bands in the visible to short-wave infrared spectral regions were atmospherically corrected using a radiative transfer model, while surface reflectance in the mid infrared regions was retrieved following Kaufman and Remer [34]. All surface reflectance bands were topographically corrected using the modified ccorrection approach [35]. The pre-fire image acquisition covered approximately 97% of the Rim fire perimeter and 92% of the King perimeter with a two km buffer in the two months preceding the fires and the entire perimeter, including buffers, after the fire. For the King fire, however, the MASTER data covered the full fire perimeter with a two km buffer, in both the pre-and post-fire imageries.
The pre-and post-fire image acquisition dates were close to anniversary date acquisitions for both fires and phenological differences between images were thus minimal. In this study, MASTER band 9, centred at 0.87 µm, was used as the NIR band in the spectral index calculations (Step 2 Figure 2). Three suitable MASTER bands covered both the SWIR and MIR spectral ranges. This allowed calculation of three different NBR and NDVIMID indices (Step 2 Figure 2). The MASTER bands 21 (centred on 2.16 µm), 22 (centred on 2.21 µm) and 23 (centred on 2.26 µm) cover the SWIR spectral region. The MASTER bands 28 (centred on 3.61 µm), 29 (centred on 3.76 µm) and 30 (centred on 3.91 µm) cover the MIR spectral region. We thus calculated three different NBR indices. NBR1 was calculated with MASTER band 21 as SWIR band, NBR2 with band 22 and NBR3 with band 23 (Table 1).

Airborne Imagery and Processing
There are currently no spaceborne sensors with a MIR band at a spatial resolution of approximately 30 m. This study capitalized upon the unique opportunity stemming from the pre-and post-fire airborne acquisitions over the Rim and King fires with the MASTER instrument. The MASTER scanner acquires data over the spectral range from 0.4 µm to 13 µm in 50 spectral bands. The spatial resolution depends on the flying height and was approximately 35 m over the Rim and King fires [33]. We used Level 2 surface reflectance data in this study. The MASTER bands in the visible to short-wave infrared spectral regions were atmospherically corrected using a radiative transfer model, while surface reflectance in the mid infrared regions was retrieved following Kaufman and Remer [34]. All surface reflectance bands were topographically corrected using the modified c-correction approach [35]. The pre-fire image acquisition covered approximately 97% of the Rim fire perimeter and 92% of the King perimeter with a two km buffer in the two months preceding the fires and the entire perimeter, including buffers, after the fire. For the King fire, however, the MASTER data covered the full fire perimeter with a two km buffer, in both the pre-and post-fire imageries.
The pre-and post-fire image acquisition dates were close to anniversary date acquisitions for both fires and phenological differences between images were thus minimal. In this study, MASTER band 9, centred at 0.87 µm, was used as the NIR band in the spectral index calculations (Step 2 Figure 2). Three suitable MASTER bands covered both the SWIR and MIR spectral ranges. This allowed calculation of three different NBR and NDVI MID indices (Step 2 Figure 2). The MASTER bands 21 (centred on 2.16 µm), 22 (centred on 2.21 µm) and 23 (centred on 2.26 µm) cover the SWIR spectral region. The MASTER bands 28 (centred on 3.61 µm), 29 (centred on 3.76 µm) and 30 (centred on 3.91 µm) cover the MIR spectral region. We thus calculated three different NBR indices. NBR 1 was calculated with MASTER band 21 as SWIR band, NBR 2 with band 22 and NBR 3 with band 23 (Table 1). Table 1. Spectral indices used in this study. The MASTER band 9, centred on 0.87 µm, was used as the NIR band in the spectral index calculations The MASTER bands 21 (centred on 2.16 µm), 22 (centred on 2.21 µm) and 23 (centred on 2.26 µm) cover the SWIR spectral region. The MASTER bands 28 (centred on 3.61 µm), 29 (centred on 3.76 µm) and 30 (centred on 3.91 µm) cover the MIR spectral region.

NDSIs
Pre-and Post-Fire Differences Similarly, we calculated three different NDVI MID indices. NDVI MID,1 was calculated with MASTER band 28, NDVI MID,2 with band 29 and NDVI MID,3 with band 30. After this, all indices were bi-temporally differenced by subtracting the post-fire index values from the pre-fire index values (Step 3 Figure 2).

Relationship with Fire Severity Field Data
Fire severity was assessed in the field using the Geometrically structured Composite Burn Index (GeoCBI) [36] in 33 plots in the Rim fire and 52 plots in the King fire. The GeoCBI is a modified version of the Composite Burn Index (CBI) [36,37]. The main modification of the GeoCBI in comparison with the CBI is that it accounts for the fractional cover of vegetation strata and as such plot-level fire severity is estimated as the weighted average of the fire severity estimates of the different vegetation strata. Both the CBI and GeoCBI have been widely used to validate remotely sensed fire severity assessments [17,20,36,38]. The GeoCBI divides ecosystems into strata and thereby differentiates between one substrate and four vegetation strata. The strata are: (i) substrates; (ii) herbs, low shrubs and small trees shorter than 1 m; (iii) tall shrubs and trees of 1 to 5 m; (iv) intermediate trees of 5 to 20 m; and (v) big trees taller than 20 m (Table 2). For each stratum, several aspects related to fire severity can be rated, when present. Some examples are soil and rock cover and its colour changes, the percentage change in leaf area index, char height and percentage of green vegetation [16,36]. Ratings are based on ocular assessment and scale between zero and three, in which zero represents unburned and three represents high severity. Based on the stratum averages, GeoCBI is calculated in proportion to the corresponding fraction of cover of each stratum, resulting in a weighted average between zero and three that expresses plot-level fire severity (Table 3) [16]. The field plots consisted of 30 m × 30 m squares and were taken in larger areas that were homogeneous with regards to pre-fire vegetation and fire severity. Field plots were sampled in order to represent the observed range of fire severity in mixed conifer forests. The selection of the field plots was constrained by road and trail accessibility, and the centre of the field plot was always at least 100 m away from the nearest road or trail. Field locations were recorded using a Global Navigation Satellite System device.  Table 3. Photo examples of high, moderate and low fire severity plots in the Rim fire.

Sample Plot Description
Remote Sens. 2021, 13, 695 7 of 20 Table 3. Photo examples of high, moderate and low fire severity plots in the Rim fire.

Sample Plot Description
High fire severity GeoCBI rating: 3.00 Large portions of downed fuels are consumed. Substantial soil exposure and soil colour change. Shrubs are absent and only few resprouts are present. The overstorey is mostly consumed, some brown needles have remained.

Sample Plot Description
High fire severity GeoCBI rating: 3.00 Large portions of downed fuels are consumed. Substantial soil exposure and soil colour change. Shrubs are absent and only few resprouts are present. The overstorey is mostly consumed, some brown needles have remained.

Spectral Index Optimality
Spectral index optimality gauges an index's sensitivity to the change of interest, in this case, fire-induced ecosystem changes and thus fire severity [7,16]. In the case of NDSIs, index optimality is defined by the direction and magnitude of pixel displacements

Spectral Index Optimality
Spectral index optimality gauges an index's sensitivity to the change of interest, in this case, fire-induced ecosystem changes and thus fire severity [7,16]. In the case of NDSIs, index optimality is defined by the direction and magnitude of pixel displacements in the bispectral space [39]. This can be illustrated by the displacement from unburned U to burned B in the bi-spectral space (Figure 3). Within the bi-spectral space, the y-axis represents the NIR reflectance for all indices, while the x-axis represents the SWIR reflectance for the dNBR indices, and the MIR reflectance for the dNDVI MID indices. The vector |UB| consists of the sum of the vectors |UO| and |OB|. A spectral index is then sensitive to displacements that are perpendicular to the index isolines. In this case, the spectral index is sensitive to the displacement from U to the optimally sensed O. In contrast, the spectral index is insensitive to displacements along isolines, in this case, the vector |OB|. By measuring the distances of the vectors |UO| and |OB|, the index optimality is defined as:

Spectral Index Optimality
Spectral index optimality gauges an index's sensitivity to the change of interest, in this case, fire-induced ecosystem changes and thus fire severity [7,16]. In the case of NDSIs, index optimality is defined by the direction and magnitude of pixel displacements in the bi-spectral space [39]. This can be illustrated by the displacement from unburned U to burned B in the bi-spectral space (Figure 3). Within the bi-spectral space, the y-axis represents the NIR reflectance for all indices, while the x-axis represents the SWIR reflectance for the dNBR indices, and the MIR reflectance for the dNDVIMID indices. The vector |UB| consists of the sum of the vectors |UO| and |OB|. A spectral index is then sensitive to displacements that are perpendicular to the index isolines. In this case, the spectral index is sensitive to the displacement from U to the optimally sensed O. In contrast, the spectral index is insensitive to displacements along isolines, in this case, the vector |OB|. By measuring the distances of the vectors |UO| and |OB|, the index optimality is defined as: As |OB| can never exceed |UB|, the optimality varies between zero and one. When the optimality equals zero, then the index is completely insensitive to the change of interest. Whereas an optimality score of one suggests that an index performs ideally in monitoring the change of interest, in this case fire severity. All pixels within the fire perimeter were used in this analysis. As |OB| can never exceed |UB|, the optimality varies between zero and one. When the optimality equals zero, then the index is completely insensitive to the change of interest. Whereas an optimality score of one suggests that an index performs ideally in monitoring the change of interest, in this case fire severity. All pixels within the fire perimeter were used in this analysis.

Analysis
In this present study, a non-linear model based on a saturated growth model was used to describe the relationship between GeoCBI and the NDSIs [40]. The saturated growth model has outperformed linear regression models to describe the relationships between the dNBR and GeoCBI. It accounts for the saturation that occurs with the dNBR for high fire severity plots [40]. The regression results of the saturated growth model were compared using two goodness-of-fit parameters: (i) the coefficient of determination R 2 and (ii) the root mean square error (RMSE). The coefficient of determination is an estimate of the proportion of the total variation in the data that is explained by the model. The RMSE is a measure of how much a response variable varies from the model predictions, it is expressed in the same units as the dependent dataset. In addition, optimality statistics of all burned pixels were compared for the different indices, and thus excluding the two km buffer zone around the fire perimeter. The median statistic was used for this purpose due to its robustness in identifying outliers and the non-normal optimality distributions.

Relationships between Field and Airborne Data
The indices were bi-temporally differenced by subtracting the post-fire index values from the pre-fire index values (Step 3 Figure 2). These differenced maps are shown in Figure 4. All spectral indices demonstrated reasonably strong relationships with the GeoCBI (Figure 5a-f). For the pooled dataset over the two fires, all dNBR indices performed better than the dNDVI MID indices. The dNBR 3 index demonstrated the best performance yielding an R 2 of 0.68. For the dNBR indices, the relationship with field data was stronger and different across the Rim and King fire (Table 4). This difference in the form of the regression line is less pronounced in the relationships between the two fires for the dNDVI MID indices (Figure 5d-f). For those indices, regression lines of the individual fire events are very close to each other. Table 4. Fitting parameters, coefficient of determination (R 2 ) and root mean square error (RMSE) of the saturated growth regression models between spectral indices as independent variables and the Geometrically structured Composite Burn Index (GeoCBI) as dependent variable. Note: Rows a-f contain parameter information which corresponds to the figure labels in Figure 5a-f.

Rim
For the individual NDSIs, different RMSE trends can be observed between these fires. The decreasing trend in R 2 statistic of the Rim fire is associated with an increasing RMSE. For the King fire, this decreasing trend results in a decrease in RMSE. The spectral index values of the dNDVI MID,3 displayed the largest value range. The ranges of values were more variable among the dNDVI MID indices compared to the dNBR indices, with the largest range corresponding the dNDVI MID index that was constructed with MIR band centred at 3.91 µm (Figure 5f).
For the individual NDSIs, different RMSE trends can be observed between these fires. The decreasing trend in R 2 statistic of the Rim fire is associated with an increasing RMSE. For the King fire, this decreasing trend results in a decrease in RMSE. The spectral index values of the dNDVIMID,3 displayed the largest value range. The ranges of values were more variable among the dNDVIMID indices compared to the dNBR indices, with the largest range corresponding the dNDVIMID index that was constructed with MIR band centred at 3.91 µm (Figure 5f).

Optimality
The optimality of the dNBR indices was always higher than from the dNDVIMID indices, with only small differences between the different dNBR indices (Figures 6 and 7). All three dNBR indices have similar optimality statistics; however, the highest median optimality for both fires was observed for the dNBR3. The median optimality of the dNBR3 was 0.60 over the King fire and 0.56 over the Rim fires. The optimality of the dNDVIMID varied widely between the indices. The highest optimality of the dNDVIMID indices was obtained by the dNDVIMID,3 index, with a median optimality of 0.46 in the King fire, and 0.49 in the Rim fire. In all indices, a large percentile of pixels recorded an optimality value lower than 0.1. This may be because unburned pixels patches may be included in the mapped fire perimeter. Across both fires, the optimality of the dNBR indices with optimality values higher than 0.1 showed a trend towards more pixels with higher optimality values. This trend was absent in the optimality distributions of the dNDVIMID indices. While median optimality values demonstrated small differences between both fires, the relative optimality performance of the indices was the same for both fires.

Discussion
The overall results of the pooled dataset showed a moderately high correlation between GeoCBI field data and the proposed NDSIs for a temperate environment, such as California (USA). The dNBR indices performed better than the dNDVIMID indices. These model outcomes fall within the ranges observed within previous GeoCBI-dNBR relationships studies [16,23,41]. Large residual variability among each fitted model

Optimality
The optimality of the dNBR indices was always higher than from the dNDVI MID indices, with only small differences between the different dNBR indices (Figures 6 and 7). All three dNBR indices have similar optimality statistics; however, the highest median optimality for both fires was observed for the dNBR 3 . The median optimality of the dNBR 3 was 0.60 over the King fire and 0.56 over the Rim fires. The optimality of the dNDVI MID varied widely between the indices. The highest optimality of the dNDVI MID indices was obtained by the dNDVI MID,3 index, with a median optimality of 0.46 in the King fire, and 0.49 in the Rim fire. In all indices, a large percentile of pixels recorded an optimality value lower than 0.1. This may be because unburned pixels patches may be included in the mapped fire perimeter. Across both fires, the optimality of the dNBR indices with optimality values higher than 0.1 showed a trend towards more pixels with higher optimality values. This trend was absent in the optimality distributions of the dNDVI MID indices. While median optimality values demonstrated small differences between both fires, the relative optimality performance of the indices was the same for both fires.   Potential sources of inaccuracy in our analysis arise from both the field and satellite observations. To fulfil the need for a stratified sampling approach of the GeoCBI, it is

Discussion
The overall results of the pooled dataset showed a moderately high correlation between GeoCBI field data and the proposed NDSIs for a temperate environment, such as California (USA). The dNBR indices performed better than the dNDVI MID indices. These model outcomes fall within the ranges observed within previous GeoCBI-dNBR relationships studies [16,23,41]. Large residual variability among each fitted model indicated relatively high uncertainties in the predictability of fire severity ( Figure 5). Similar asymptotic behaviour is exhibited across all NDSIs in relation to the moderate to high GeoCBI values. The inability and underestimation of the dNBR indices to distinguish among high severity sites highlighted in previous GeoCBI-dNBR models have been the driver of attempting non-linear asymptotic modelling as suggested by Hall [20,40]. In our study, similar saturation of spectral indices for high severity plots occurred justifying the use of a saturated growth model [16]. Previous studies based on the spectral index theory, reported that the dNBR outperformed other dNDSI with regards to optimality distributions [6,16]. In these studies, the dNBR outperformed the differenced NDVI (dNDVI) and differenced normalized difference moisture index (dNDMI). Our dNBR results approximated earlier reported values for the dNBR and fall within ranges recorded in six Alaskan fires by Murphy et al. [41].
In addition, our results reaffirm the strength of the dNBR as being the most sensitive to fire-induced changes in spectral reflectance, and less so to noise. Over both fire scars, the dNBR's optimality was consistently higher than what was observed for the dNDVI MID . The optimality statistics for the dNDVI MID resulted in a large variety of mean values ranging from 0.167 to 0.46. Roy et al. [7] conducted an optimality analysis on the dNBR of MODIS-sensed fires in the boreal forest in Russia, African savannas and the South American tropical forest. Roy et al. [7] included both unburned and burned pixels in their analysis, and roughly a third of their analysed pixels had optimality scores lower than 0.1. Including unburned pixels resulted in lower mean optimality scores (between 0.24 and 0.33) than in studies which excluded the unburned pixels [6,7,16]. Our analysis reports a similar percentage of pixels with an optimality score lower than 0.1. These pixels may be unburned patches that were included in the fire perimeter.
Studies assessing the correlation between several spectral indices and GeoCBI or CBI field data identified the NBR as the best index in the bi-temporal differencing approach [16,22]. Moreover, this study confirms that the dNBR outperforms the dNDVI MID . Despite this relative success, our analyses confirm the same large variability in the form of relationships across fires as mentioned by previous studies [22,40,41]. This variability may partly be caused by the spectral sensitivity of the SWIR spectral region to soil variations [5,7]. The dNBR approach is therefore most likely affected by variability in background conditions. This hampers comparison of dNBR across fires with different soil conditions. Miller and Thode [42] have proposed the relative dNBR (RdNBR) in which the dNBR is divided by the square root of the pre-fire NBR. The RdNBR as such accounts for differences in pre-fire vegetation cover and enhances the intercomparison over heterogeneous landscapes. The dNDVI MID indices displayed a reduced variability in regression forms across the Rim and King fires. This improved intercomparison across different fires from the dNDVI MID is in line with previously reported improvements in discriminating between unburned and burned pixels [26,43]. This suggests that the dNDVI MID bi-temporal differencing approach may result in a more comparable fire severity proxy over large regions.
Based on the dNDVI MID pooled data-GeoCBI correlations, the indices that were constructed with MIR band centred at 3.76 µm (dNDVI MID,2 ) and 3.91 µm (dNDVI MID,3 ) provided similar results regarding their coefficients of determinations (R 2 = 0.61) and outperformed the index centred at 3.61 µm (dNDVI MID,1 ). The GeoCBI-dNDVI MID,2 model yielded the lowest RMSE of 0.48, which is approximately one-sixth of the total GeoCBI variation ( Table 3). The NDVI MID,2 also demonstrated the smallest variation in regression from across the two fires ( Figure 5d). Furthermore, the dNDVI MID,2 also displayed a wider spectral range for the 85 samples than the dNBR indices. This holds also true for the dNDVI MID,3 . This first spectral optimality analysis of the dNDVI MID displayed large variation in performance dependent on the wavelengths used for the MIR bands. The dNDVI MID index that integrated the MIR band with the longest wavelength yielded the highest optimality scores, in which the dNDVI MID,3 index approached an optimality of 0.5.
Potential sources of inaccuracy in our analysis arise from both the field and satellite observations. To fulfil the need for a stratified sampling approach of the GeoCBI, it is required that the number of plots for each fuel type are in proportion to the total area burned of each fuel type. Because of accessibility limitations, we were unable to fulfil these conditions. Part of the observed divergence between field and remotely sensed fire severity estimates can also be explained by the fact that both field and satellite data are imperfect proxies of fire severity. The GeoCBI is based on a semi-quantitative judgement approach, which is inherently prone to some degree of subjectivity, especially when different people perform the field assessments. This was the case in our study in which different field teams surveyed the Rim and King fires. Several noise factors can impede remote sensing image analyses. The amount of noise in all spectral indices was substantial, as the median optimality scores are considerably lower than the optimal score of one. Noise within remote sensing analysis can result from imperfections in image pre-processing and slight deviations in phenology between near-anniversary date image acquisitions. Garcia et al. [44] recently developed a new metric, the waveform area relative change (WARC), from light detection and ranging (LiDAR) imagery. The WARC proxy captures changes in vegetation structure as well as changes in soil and leaf colour and was successfully validated over the King fire. Synergistic use of such LiDAR derived metrics with optical indices like the dNBR and dNDVI MID could provide insights on how differences in vegetation structure influence the performance of spectral indices. Our findings can contribute to improvements in the remote sensing of fire severity, as it proposes an index with similar performance as the current dNBR approach. Our preliminary findings of the dNDVI MID approach suggest that use of the dNDVI MID index for fire severity assessment may lead to more comparable remote sensing assessments among fire scars than with the dNBR. We evaluated the dNDVI MID thanks to the unique availability of pre-and-post-fire MASTER imagery over the Rim and King fires. The potential of the dNDVI MID for assessing fire severity needs further evaluation when consistent MIR imagery acquisition will become available at a spatial resolution of less than 100 m, allowing comparison with field measurements of fire severity.

Conclusions
The results of our comparison between field and remotely sensed proxies of fire severity and the spectral optimality analysis confirmed the dNBR as a strong fire severity predictor. Overall, the dNBR yielded better results than the dNDVI MID in estimating fire severity over two Californian fires. The field data analysis, however, also identified that the dNBR approach suffered from significant variations in the form of regression lines across the two fires. The saturated growth fit between field and dNDVI MID data resulted in more comparable regression lines across the two fires, despite a slightly lower overall performance than the dNBR. This suggests the dNDVI MID bi-temporal differencing approach may be a more robust method in estimating and comparing fire severity over large regions. Our study leveraged unique pre-and post-fire airborne acquisitions over the Rim and King fires in California from the MASTER sensor. As such, our study is the first to assess the NIR-MIR bi-spectral space for assessing fire severity and calls for a continued evaluation of dNDVI MID for assessing fire severity. This will require pre-and post-fire imagery to be available for the MIR spectral region at a spatial resolution that allows comparison with field measurements.