5.1. Burned Area Discrimination
The high discriminatory power of the NIR spectral band is consistent with findings of Veraverbeke et al.
], Smith et al.
], Pereira [69
], Lasaponara [73
], and López-García and Caselles [76
]. This spectral region has been widely considered as the best spectral region to detect and map burned areas [41
]. The NIR region is strongly reflected by vegetation. As a result, vegetation removal or scorching implicates a post-fire drop of NIR reflectance. The SWIR reflectance increases after burning, due to the fire’s removal of water-retaining vegetation [25
]. This reflectance increase is higher than in the visible wavelengths (VIS) and, hence, the SWIR region outperforms the VIS regions. As stated in literature (e.g., [41
]), the VIS wavelengths have a very low performance in discriminating between burned and unburned areas. A likely explanation is that water, water-rich swamps, dense conifer forests and peat soil types all are relatively dark which causes spectral confusion with burned areas in these wavelengths [69
]. Differences in separability performance among wildfires have been observed by Lasaponara [73
]. Lasaponara [73
] suggests that these observations might be due to the different land cover types affected by the fire. Our results support the hypothesis that differences in separability performance are caused by different vegetation types within the burn scar.
The best performing LSWIR band (central wavelengths 2332, 2339, and 2352 nm, except for Pinus
) corresponds with the findings of van Wagtendonk et al.
], Veraverbeke et al.
], and Pleniou and Koutsias [71
], who found the most discriminating wavelengths at 2370 nm, 2310–2360 nm and 2300–2370 nm, respectively. In the study by van Wagtendonk et al.
], the 788 nm NIR band demonstrated the strongest reflectance decrease, close to the optimal NIR band for Pinus
vegetation (767 nm) and for the pooled dataset (801 nm) found in our study. The optimal SSWIR band in this study (1302 nm), however, differs with Veraverbeke et al.
] (1600 nm) and van Wagtendonk et al.
] (1762 nm).
The NBR had the highest discriminating power in our study, due to its combining of the spectral regions with strongest decrease and increase, the NIR and LSWIR region, respectively [26
]. This index provided also good discriminatory power in South African savannahs [46
]. The CSI, though using the same bands, performs poorly in this study, just as in Veraverbeke et al.
]. In savannahs, however, the CSI attained a high degree of spectral separability [46
]. The higher discriminative power that we found of the NIR spectral band compared to the NDVI is caused by the very low performance of the red (and visible) wavelengths. This is also confirmed by other studies, where the NIR channel alone achieved better separability results than indices based on a combination of NIR and VIS spectral bands [46
5.2. Burn Severity Assessment
Very few studies analyze the correlation between separate spectral bands and field data. Epting et al.
] and Hoy et al.
] only report the NIR and LSWIR spectral region in their analysis. Epting et al.
] analyzed correlations between the NIR-band and the GeoCBI for four forest fires, resulting in correlations (R2
) between 0.06 and 0.62. This is consistent with Hoy et al.
= 0.39 and 0.27) and our study (R2
= 0.40). Hoy et al.
] did analyze the SSWIR band, but did not report numbers due to insignificance of the results. In our study, however, this band demonstrated the highest correlation with the GeoCBI. For the LSWIR spectral region, we found similar results as Hoy et al.
]: in both studies the LSWIR performance is very low. Epting et al.
], however, revealed a R2
between 0.19 and 0.55 for the LSWIR region. The difference in performance between the SSWIR and LSWIR spectral band might be explained by differences in liquid water absorption: water absorption in the LSWIR band is significantly stronger than in the SSWIR band [78
]. The 1302 nm SSWIR band used in this study is situated close to the NIR spectral region. Therefore this SSWIR band is expected to be less sensitive to moisture content than to vegetation structure, which is the most important variable influencing the NIR spectral region [80
]. The inconsistencies between the study of Hoy et al.
], and Epting et al.
], and this study may be due to the different ecosystems and vegetation types observed, i.e.
, Alaskan black spruce forests, interior Alaskan vegetation types (forests, as well as low cover sites) and Atlantic European heathlands respectively.
Indeed, previous research demonstrated that the correlation performance of the GeoCBI depends on the vegetation type [32
]. Stratifying our data per vegetation type also improved the correlation results considerably. Most studies show stronger correlations in forested areas than in sparser vegetation types like shrubs and herbs: Epting et al.
] found strong relations in mixed forests (R2
= 0.83). Herbs and shrubs on the other hand, showed a Pearson correlation coefficient of 0.33 and 0.25 respectively (R2
= 0.11 and 0.06). The authors conclude that the CBI may not be appropriate to assess burn severity in non-forested areas. In addition, Key and Benson [25
], Veraverbeke et al.
], De Santis and Chuvieco [56
], and the review by French et al.
], show stronger correlations in forested areas than in shrub and herb vegetation. In our heathland fire study, the NBR performed rather poorly, while in other studies this index generally acquires the best results when compared to other indices [44
]. This also indicates that the performance of burn severity assessment varies among vegetation types.
The CSI was the best index in dry heath (Calluna
) vegetation. The CSI is the ratio of the NIR and LSWIR spectral band (Table 2
), a low CSI value means a high burn severity. Bare soil that was exposed post-fire also results in low CSI values, due to similar reflectance values in the NIR and LSWIR region. The CSI is designed to detect the char signal [46
], but it also amplifies the soil cover change signal. This results in good performances in our dry heath vegetation, as well as in the chaparral ecosystems of California [44
]. Wet heath vegetation shows a very poor correlation, considerably lower than the other vegetation types. This might be partly explained by the fact that this is the only vegetation for which we did not sample unburned field plots (Figure 2
), which considerably lowers the range of variability in both the SIs and GeoCBI scores, affecting the strength of the regressions. In wet heath vegetation (mainly Erica
) and grass-encroached heath (Molinia
stands), the MIRBI showed the best correlation with the field data. Our finding that the MIRBI had the highest correlation within the wet heath and grass vegetation types is consistent with Trigg and Flasse [47
]. In fact, this SI was designed for shrub-savannah ecosystems which has similarities with heath ecosystems. For example, grasses are widely present in both ecosystems and they are both characterized by quick post-fire recovery. The relatively high correlations with the SSWIR and especially the LSWIR spectral regions indicate that the post-fire decrease in moisture content varies among and is related to different severity levels. The coniferous vegetation has the highest correlation with the red spectral band and the NDVI. Differences in severity between the coniferous woodland plots were expressed as the amount of scorching of the needles in the crown. In none of the plots the pine needles were completely burned, but the amount of scorching of the crown significantly differed among the coniferous woodland plots. This scorching influenced the photosynthetic activity and the amount of healthy vegetation of the trees, which is measured in the red and NIR region, respectively. This might explain the superior performance of the red region and NDVI for the coniferous woodland plots. The LSWIR region in post-fire environments is especially sensitive to increased reflectance due to the charcoal signal. However, in the coniferous woodland plots, charcoal exposure remained relatively limited, which might explain the lower performance of this region. Consequently, the most important factor in this vegetation type is not the change in water content, such as in Molinia
vegetation types, but the change in photosynthetically active vegetation density.
We observed a discrepancy between the results of the spectral sensitivity analysis for burned area discrimination and the burn severity regression analysis. In the burn severity analysis, the SSWIR band showed the best performance and the LSWIR band the lowest. However, the SSWIR-band resulted in low discriminatory power for burned area discrimination, while the LSWIR was the second best burned area discriminator (compare Table 6
with Table 4
). The results of the spectral indices further confirm this discrepancy: the NBR outperformed the other indices in discriminating between burned and unburned areas, but was for no vegetation type the best predictor of burn severity. The opposite was also true: the CSI revealed very low discriminatory power for burned area mapping, whereas this index performed best to estimate burn severity in dry heath vegetation. Certain individual bands and indices, thus, performed differently for mapping burned area, compared to burn severity. Therefore, it is recommended to optimize the SI selection for each application separately.
We used the optimal regression parameters of the best-performing SI for each vegetation types to predict a spatial burn severity distribution for the entire burned area. An alternative approach could have been to validate the optimal regression parameters with independent field data. Due to the low number of field plots after stratification per vegetation type (e.g., eight Pinus plots), this was not feasible in our research.
In this study, we used spectral indices for their conceptual simplicity and computational efficiency. These spectral indices are widely used, and therefore direct comparison with previous research and other study areas is straightforward. Hence, spectral indices have clear advantages over other burn severity assessment methodologies. However, spectral indices use only a small fraction of the information contained in the spectroscopy images. For example, the spectral indices in this study used two or three spectral bands (see Table 2
), whereas the APEX image acquired data in 233 bands suitable for analysis. Therefore, future research can focus on upcoming but more advanced methodologies that combine all information available from the spectroscopy data cube. Testing multiple band combinations with feature selection methods is one of the methods that may reveal more insights in the relation between severity and spectral response. Another technique that is often applied on imaging spectroscopy datasets is Spectral Mixture Analysis (SMA). SMA may reveal the proportion and physical presence of char within a pixel, which has been proposed as an indicator for severity assessments by several authors [21
]. However, also these techniques have their limitations, e.g., the definition of the endmembers [83
This article is a contribution to spectral imaging spectroscopy research of wildfires. More available spectral libraries (e.g., the ASTER spectral library [84
]), new emerging SMA techniques [83
] and spaceborne imaging spectroscopy missions planned in the near-future (Hyperspectral Infrared Imager, hyspiri.jpl.nasa.gov
; Precursore Iperspettrale (Hyperspectral Precursor), www.isa.it
; Environmental Mapping and Analysis Program, www.enmap.org
), might make spectroscopy data more widely available. This will foster more research on the spectral properties of post-fire landscapes and vegetation recovery in various vegetation types such as the heathlands in this study.