Linking crown fire likelihood with post-fire spectral variability in Mediterranean fire-prone ecosystems

Introduction

Wildfires are a frequent disturbance in terrestrial ecosystems of Mediterranean-type climate regions around the world (Catry et al. 2013; Xofis et al. 2020). In the western Mediterranean Basin, the abandonment of traditional extensive agricultural practises and forestry uses in past decades. Together with proliferation of unmanaged forest plantations and fire suppression policies, this has led to an extensive and continuous accumulation of dense and flammable fuels (Moreira et al. 2011; Fernandes 2013). In addition, anthropogenic climate change has increased the occurrence of prolonged droughts and heat waves (Tripathy et al. 2023), leading to dryness conditions of fuel conducive to extreme fire behaviour and intense crown fires (Dimitrakopoulos et al. 2007; Pickering et al. 2023). These events may have unprecedented ecological consequences (Lasslop et al. 2019) and associated losses of human lives, infrastructure and properties (Mansoor et al. 2022).

Extreme wildfire events involving crown fires usually occur under severe fire weather conditions, resulting in hazardous and erratic fire behaviour (Mitsopoulos and Dimitrakopoulos 2007) over a wide variety of forest ecosystems throughout the western Mediterranean Basin (Fernández-Guisuraga et al. 2019, 2023a). A crown fire in forest spreads much faster (up to two to four times) than a surface fire burning in the same conditions (Fernandes et al. 2004; Perrakis et al. 2023), and its control efforts by direct attack are ineffective (Erni et al. 2020; Frost et al. 2022). These characteristics, together with increased fire intensity and frequent long-range spotting (Albini et al. 2012), may entail a serious threat to property and life in the wildland–urban interface (Fiorini et al. 2023). Long-lasting effects of high-intensity crown fires may include near total tree mortality (Woolley et al. 2012), abrupt shifts in species composition and structure leading to alternative stable states threatening ecosystem resilience (Allen et al. 2002; Knox and Clarke 2012) and biogeochemical cycles (Varner et al. 2021). Such fire impacts reflect the type and extent of heat transfer processes to which vegetation and soils were exposed (Fernández-Guisuraga et al. 2023b). Altogether, predicting the occurrence of crown fire behaviour is of utmost importance for planning pre-fire fuel treatments and fire suppression strategies (Scott and Reinhardt 2001).

Wildfires rarely behave only as crown fires due to changing fuel, topography or fire weather and thus, the spatial variability in the fire type entails major implications for the structure, successional dynamics and functioning of fire-prone ecosystems (Erni et al. 2020; Pérez-Izquierdo et al. 2021; Taylor et al. 2021). Therefore, it is not only important to assess the likelihood of crown fire behaviour, but also the spatial variability of the fire type (e.g. surface, crown) after wildfire extinction. Such knowledge would be of value, for example, in (1) testing the accuracy of fire behaviour models (Alexander and Cruz 2012), (2) providing insights on pre-fire fuel conditions conducive to extreme fire behaviour and severe ecological impacts (Dimitrakopoulos et al. 2007), (3) evaluating fuel treatment effectiveness (Cruz et al. 2004; Hu et al. 2019), and (4) estimating post-fire ecosystem responses including delayed tree mortality (Shearman et al. 2023). Indeed, several authors suggested that physically meaningful variables of actual fire effects that can be used as a proxy for fire type, such as crown fraction burned (CFB), can be more readily translated into management applications (Woolley et al. 2012; Hood et al. 2018), as opposed to integrative fire severity indices (e.g. the Composite Burn Index; Key and Benson 2005) (Morgan et al. 2014; Fernández-Guisuraga et al. 2023c).

Field methods enable assessing the type of fire based on both direct observation of fire behaviour (Cruz et al. 2003) and crown scorch and consumption (e.g. Pollet and Omi 2002; Morgan et al. 2014), but are labour-intensive and unable to thoroughly cover extensively burned landscapes. In this sense, the synoptic nature of remote sensing earth observations may be more appropriate for wall-to-wall estimation of fire type in post-fire landscapes (Fernández-Guisuraga et al. 2023a). Most remote sensing-based research to date have been focused on assessing fire severity drivers (e.g. Parks et al. 2018; Fernández-Guisuraga et al. 2021, 2023d; Fernández-García et al. 2022) or on developing accurate wall-to-wall fire severity estimates (e.g. De Santis and Chuvieco 2007; Miller et al. 2009; Quintano et al. 2013; Fernández-Guisuraga et al. 2023a). Conventionally, multispectral remote sensing data acquired from broadband sensors have been used to compute spectral indices, such as the differenced Normalised Burn Ratio (dNBR; Key 2006) or its relativised variants (Miller et al. 2009; Parks et al. 2014), as a proxy for the spectral signal of fire effects through empirical models. In this context, many previous studies (e.g. Roy et al. 2006; Fernández-Manso et al. 2016; Delcourt et al. 2021) and operational programs such as the Monitoring Trends in Burn Severity (MTBS; Picotte et al. 2020) in the United States or the Rapid Damage Assessment (RDA) module of the European Forest Fire Information System (EFFIS), have extensively used spectral indices computed from Landsat and Sentinel-2 imagery to obtain wall-to-wall fire severity estimates. Physical-based models applied to broadband or narrowband multispectral data, such as multiple endmember spectral mixture analysis (MESMA; Roberts et al. 1998), have also been considered in previous research (e.g. Quintano et al. 2013, 2023; Meng et al. 2017) to decompose sub-pixel reflectance signal and retrieve the fractional cover at pixel level of post-fire ground components (e.g. char or green vegetation) representative of wildfire ecological effects (i.e. fire severity) on vegetation and soils. This advanced pixel unmixing technique is considered to be more robust, scalable and generalisable than spectral indices due to its physical basis (Quintano et al. 2013).

Previous studies have seldom considered the potential of post-fire spectral variability to provide fire type estimates, particularly in Mediterranean fire-prone ecosystems (Mitri and Gitas 2006; Collins et al. 2018). These studies leveraged either pixel-based or object-based classification schemes and spectral indices computed from multispectral satellite data for mapping fire type or predicting fire severity categories as proxies for fire type. Other authors used a change detection framework for identifying fire type from airborne laser scanning (ALS) data in Sierra Nevada, California (Hu et al. 2019). As a bi-temporal change-detection framework, this method requires the acquisition of pre- and post-fire ALS datasets, which is a constraint because of the limited ALS data availability (Fernández-Guisuraga et al. 2022), contrary to multispectral satellite data.

To the best of our knowledge, physical-based and generalisable remote sensing approximations applied to broadband multispectral data, together with extensive field assessments, have not been used to detect crown fire in post-fire environments. Accordingly, we explore for the first time in the literature the potential of multispectral satellite data and advanced pixel unmixing models to provide meaningful and generalisable fire type estimates in Mediterranean ecosystems. Specifically, we examined how the post-fire spectral signal variability of MESMA fraction images computed from Sentinel-2 multispectral data reflects the likelihood of distinct fire types (surface, intermittent crown, continuous crown fire; Forestry Canada Fire Danger Group 1992) based on CFB, using spectral indices as a benchmark. See Scott and Reinhardt (2001) and National Wildfire Coordinating Group (2023) for fire type definitions. We selected four wildfires for this purpose that burned different types of broadleaf and conifer forests across the western part of the Mediterranean Basin. We hypothesise that MESMA fraction images, specifically the char fraction, would outperform spectral indices due to their higher physical sense and the intrinsic relationship between CFB and char signal in post-fire landscapes.

Material and methods

Study sites

We selected four wildfires that affected forests, shrublands and grasslands in the western Mediterranean Basin (north-western and central Spain) under extreme fire weather conditions in the summer seasons between 2017 and 2022 (Fig. 1), with unprecedented prolonged droughts and heat waves in the months prior to the wildfires. The sites encompass wide environmental gradients are in Table 1. We chose as target conifer and broadleaf forests within the wildfires. Maritime pine (Pinus pinaster Ait.) dominated conifer forests in lowlands, whereas Scots pine (Pinus sylvestris L.) dominated in highlands. Broadleaf forests were mainly dominated by Pyrenean oak (Quercus pyrenaica Willd.) and holm oak (Quercus ilex L.). Wildfires were selected on the basis of available canopy fraction burned (CFB) data acquired by the same observers.

Table 1.Location and characteristics of the four studied wildfires. We provide the number of field plots established within each ecosystem type in the wildfires (Qp, Quercus pyrenaica; Qi, Quercus ilex, Pp: Pinus pinaster; Ps, Pinus sylvestris).

Wildfire	Sierra de Cabrera	Villapadierna	Navalacruz	Sierra de la Culebra
Location	NW Spain	NW Spain	C Spain	W Spain
Wildfire size (ha)	9940	82	24,444	28,046
Wildfire date	21 August 2017	22 August 2019	14 August 2021	11 June 2022
Elevation range (m)	836–1938	922–1027	939–2157	747–1205
Slope range (%)	0–149	6–27	0–649	0–155
Average annual precipitation (mm)	850	761	758	750
Average annual temperature (°C)	9.0	10.7	9.6	11Confusion matrix and accuracy metrics
Plant communities (# plots)	Qp (21)	Qp (15) Pp (5) Ps (5)	Qi (8) Ps (18)	Qp (9) Qi (14) Ps (21) Pp (13)
Pre-fire Sentinel-2 image date	13 August 2017	15 August 2019	9 August 2021	26 May 2022
Post-fire Sentinel-2 image date	2 September 2017	30 August 2019	8 ctober 2021	15 July 2022

Fig. 1.

Location of Sierra de Cabrera (a), Villapadierna (b), Navalacruz (c), and Sierra de la Culebra (d) wildfires in the western part of the Mediterranean Basin (north-western and central Spain). A Sentinel-2 false colour composite (R = band 12; G = band 8A; B = band 4) is displayed at the background.

Results

The CFB of conifer forests was significantly larger than that of broadleaf forests (Mann–Whitney P-value < 0.01) in the study sites (Fig. 2). Char fractional cover estimated from MESMA was significantly higher in conifer than in broadleaf forests (P-value < 0.01), with NPVS fraction exhibiting the opposite behaviour (Fig. 2). Both spectral indices (i.e. dNBR and RBR) and the GV MESMA fraction did not significantly differ (P-value > 0.05) between forest types (Fig. 2).

Fig. 2.

Boxplots depicting the relationship of canopy fraction burned (CFB), spectral indices (dNBR and RBR) and MESMA image fractions (char; GV, green vegetation; NPVS, non-photosynthetic vegetation and soil) with ecosystem type (BF, broadleaf forests; Cf, conifer forests). Mann–Whitney U test results are shown in the upper part of the plots.

All spectral indices and MESMA fractions differed significantly between fire types (χ > 46.96; P-value < 0.001), with a significant increase from surface to crown fires in the case of spectral indices and the char MESMA fraction, and the opposite behaviour in the case of the GV and NPVS fractions (Fig. 3). The strongest relationship with fire type corresponded to the char MESMA fraction.

Fig. 3.

Boxplots depicting the relationship of spectral indices (dNBR and RBR) and MESMA image fractions (char; GV, green vegetation; NPVS, non-photosynthetic vegetation and soil) with fire type. We show Kruskal–Wallis test results in the upper part of the plots. Lowercase red letters denote significant differences between fire types at P = 0.05.

In broadleaf forests, the highest performance in fire type classification was obtained by the multivariate RF model calibrated from MESMA fractions (OA = 82.09% and Kappa index = 0.71). The Producer’s and User’s accuracy of the model were balanced and no remarkable confusion was observed among all classes (Table 2). Classification accuracy was notably lower for the spectral indices (OA < 71.96% and Kappa index < 0.55), particularly for the dNBR (Table 2). The greatest confusion was observed for both indices between surface and intermittent crown fires, and between intermittent and continuous crown fires. Continuous crowning was never misclassified as a surface fire for all remote sensing products (Table 2).

In conifer forests, the overall accuracy provided by spectral indices and MESMA fractions in the RF fire type classification, as well as the Producer’s and User’s accuracy of the models (Table 3), followed the same pattern as in broadleaf forests (MESMA accuracy > RBR > dNBR). The very low confusion rate in the MESMA-based classification is worth noting (Table 3).

All remote sensing products performed better in conifer (OA = 76.88% ± 8.11% and Kappa = 0.62 ± 0.13) than in broadleaf (OA = 72.71% ± 8.99% and Kappa = 0.56 ± 0.15) forests.

The likelihood of continuous crowning relative to the spectral variability of burned areas captured by the MESMA fractions was very similar in conifer and broadleaf forests (Fig. 4). Occurrence of continuous crowning is linked in both ecosystems to post-fire char fractions greater than 0.5, the maximum likelihood being reached when char fraction is equal to 0.8. When GV and NPVS fractions are greater than 0.25–0.30, the likelihood of continuous crowning is very low. The MESMA char fraction was the most important variable to explain fire-type likelihood in broadleaf and conifer forests (Fig. 4). Consistency of continuous crowning likelihood between spectral indices and ecosystem types was lower than in the case of MESMA fractions (Fig. 4). dNBR values above 500 seem to be associated with maximum crowning probability in conifer forests, while in broadleaf forests the probability continues to increase progressively above such dNBR threshold. The same behaviour in both forest types was observed with RBR values above 300.

Fig. 4.

Partial dependence plots depicting the relationship in broadleaf and conifer forests between crown fire likelihood and the variability of MESMA image fractions and spectral indices in the Random Forests (RF) classification algorithm. The red line is a LOESS smooth curve. Mean decrease in accuracy (MDA;%) metric is shown for each MESMA image fraction (multivariate RF models).

When pooling data from conifer and broadleaf forests, the RBR index shows poor generalisation in fire type estimates (OA = 62.88% and Kappa index = 0.42) (Table 4) when comparing results with RF classification models by single ecosystems (OA = 73.84% ± 2.79% and Kappa = 0.58 ± 0.04) (Tables 2 and 3), with even greater underestimation of continuous crowning and higher confusion between surface and intermittent crown fires. Conversely, the performance loss of the global RF model was negligible (OA = 82.95% and Kappa index = 0.73 vs OA = 83.79% ± 2.40% and Kappa = 0.74 ± 0.04) when calibrated with MESMA fractions (Table 4).

These results are consistent with wall-to-wall fire type maps computed from RF model objects at landscape scale in broadleaf and conifer forests within the Sierra de la Culebra wildfire (Fig. 5). The area of continuous crowning is much smaller in RBR than in MESMA estimates due to the high confusion between intermittent and continuous crown fire types in the former product. Wall-to-wall estimates of the RF models calibrated in conifer and broadleaf forests separately are quite consistent with those of the RF model calibrated from global data in the case of MESMA fractions. In contrast, wall-to-wall estimates of RBR-based models are much less consistent between the two model calibration strategies.

Fig. 5.

Location of the field plots and wall-to-wall fire type predictions in broadleaf and conifer forests within the Sierra de la Culebra wildfire (RF model in conifer and broadleaf forests separately). We also present the map for the RF model calibrated from global data (conifer and broadleaf forests pooled together).

Discussion

Despite the extensive literature on fire severity assessments through remote sensing techniques using integrative field measurements (e.g. Fernández-Guisuraga et al. 2023a; Miller et al. 2023) or physically meaningful variables such as crown scorch height or crown consumption (e.g. Lentile et al. 2009; Lydersen et al. 2016; Arkin et al. 2023), the potential of physically-based and generalisable remote sensing approximations to provide wall-to-wall fire type estimates representative of wildfire behaviour is examined for the first time in this study. Our results showed the importance of generalisable remote sensing techniques to procure accurate crown fire likelihood estimates that align with post-fire land management needs in Mediterranean burned landscapes (Keeley 2009). This is particularly relevant because, although fire severity is closely related to fire behaviour (Finney 2005), the latter is what really determines the effectiveness of fire suppression efforts and size-dependent fire impacts (Fernandes et al. 2010).

Typically, remote sensing techniques with a physical basis, including MESMA, perform better than empirical methods based on spectral indices for retrieving biophysical properties of burned landscapes (e.g. De Santis and Chuvieco 2007; Fernández-Guisuraga et al. 2021, 2023a; Quintano et al. 2023). Accordingly, MESMA image fractions also procured here a remarkably higher performance than spectral indices in the fire type RF classifier, particularly when pooling field data from several ecosystems. Mitri and Gitas (2006) used an object-based classification relying on spectral indices computed from a post-fire IKONOS scene, together with object contextual information, for mapping fire type within a single, small wildfire that affected a Mediterranean pine forest in Greece. The authors reported a classification accuracy (overall accuracy and Kappa index) comparable to that obtained here with MESMA fractions. However, they did not consider intermittent crown fire in the fire type classification, which may be responsible for high patchiness of fire effects over the landscape (Scott and Reinhardt 2001) and thus for mixed spectral responses. In contrast to our study, only one ecosystem type within a single wildfire was examined, and it has been previously reported that spectral indices (1) suffer from generalisation issues between different vegetation types and environmental conditions due to the lack of physics in the retrieval of fire effects (e.g. De Santis and Chuvieco 2007; Lentile et al. 2009; Fernández-Guisuraga et al. 2023a), and (2) have suboptimal sensitivity to complex spectral responses (e.g. Roy et al. 2006; Mallinis et al. 2018).

Despite this, the higher accuracy of MESMA image fractions versus spectral indices used as benchmark within the same experimental design for estimating fire-type spatial variability may be related to the sound physical meaning of sub-pixel image analysis techniques in post-fire environments (Quintano et al. 2013). In this context, capturing the variability in the fire type spectral signal can be a subpixel matter when using moderate spatial resolution multispectral imagery (Quintano et al. 2013; Fernández-Manso et al. 2019), such as Sentinel-2. Specifically, the high spatial variability in surface and canopy fuels expected in Mediterranean fire-prone ecosystems (Fernández-Guisuraga et al. 2023a), together with fine-scale variations in terrain and fire weather, may result in fire behaviour alternation between different fire types at small spatial scales (Scott and Reinhardt 2001). MESMA image fractions also represent intrinsic biophysical characteristics of burned landscapes (i.e. char or photosynthetic vegetation), not only a secondary proxy for these constituents. Indeed, the SWIR region of broadband remote sensing data involved in dNBR and thus RBR calculation is not as sensitive to the spectral variability of char, ash and soil, nor to their complex mixture in post-fire scenarios as traditionally assumed (see Lentile et al. 2009), unlike the NIR region to vegetation vigour or amount (Hudak et al. 2007). Several authors also stated that spectral indices such as dNBR were originally conceived to map burned areas, and not to estimate the variability in their biophysical properties (Roy et al. 2006).

The ability of MESMA to resolve complex and mixed spectral responses of scorched and burned canopies due to the use of full available spectra (Quintano et al. 2023), rather than a limited number of bands and thus spectral information as in the case of spectral indices (Mallinis et al. 2018), may have prevented the high confusion between fire types and the underestimation of continuous crowning. This effect, evident in dNBR and RBR estimates and highly dependent on vegetation type (Lentile et al. 2009), was probably also attributable to the SWIR reflectance saturation and steady NIR reflectance decrease at high char and ash cover (Soverel et al. 2010). Indeed, high crown fire likelihood in conifer and broadleaf forests is linked to a post-fire MESMA char fractional cover of about 0.8 (Fig. 4), close to the CFB defining a continuous crown fire (CFB > 0.89). Consistent with previous fire severity research (Tane et al. 2018; Fernández-Manso et al. 2019), the char fraction was the most important ground constituent to retrieve fire type likelihood in broadleaf and conifer forests.

The better performance of MESMA-based models (and spectral indices) in conifer than in broadleaf forests may be associated with the typical lower canopy closure in Mediterranean conifer forests than in broadleaf forests (e.g. García et al. 2010; Sheffer et al. 2015), which may cause the char spectral signal from the lower canopy strata to be better captured by the sensor. This behaviour has been previously reported by Gibson et al. (2020) in Mediterranean, semi-arid and subtropical burned areas of eastern Australia. The authors evidenced that high canopy closure may result in the underestimation of the fire severity spectral signal sensed by moderate spatial resolution satellites. Conifer forests were more prone to elevated CFB and thus to high-intensity crown fires than broadleaf forests according to fire hazard generic expectations and previous reports of crowning potential in the Mediterranean Basin (Fernandes 2009; Fernandes et al. 2010; Fernández-Guisuraga et al. 2023c), and in other biomes elsewhere (e.g. Scott and Reinhardt 2001; Epting and Verbyla 2005). This may be linked to higher accumulation of flammable litter and fuel loading of fine-fuel rich species in the understory of unmanaged Mediterranean conifer forests than in those dominated by broadleaf species, coupled with increased vertical fuel continuity and ladder fuels (Safford et al. 2012; Fernández-Guisuraga et al. 2021).

MESMA fraction images provided not only more accurate fire type estimates than spectral indices in broadleaf and conifer forests, but also featured greater generalisation ability. This could be related to the analogy between the variation in crown scorch and consumption estimated through MESMA fraction images and the fire type definition (CFB) used in the field. In addition, the variability of the background reflectance signal corresponding to bare soil can affect the discrimination of ecological fire effects and generalisation ability when relying on spectral indices (Meng et al. 2017; Fernández-Guisuraga et al. 2023a). Conversely, the endmember collection of all representative ground constituents at each site to build the definitive spectral library in the MESMA algorithm could be accounted for the minimisation of the background influence in the estimation of fire effects (Quintano et al. 2023).

Overall, the results of this study could be leveraged by land managers to reliably infer tree damage, mortality and ecosystem responses in Mediterranean post-fire landscapes, and develop accordingly appropriate post-fire management plans and restoration strategies. In particular, the potential implementation of physical-based algorithms such as advanced spectral mixture models in consolidated geospatial processing platforms in the cloud; e.g. Google Earth Engine (GEE; Gorelick et al. 2017), could emerge as a valuable resource for minimising data acquisition and processing efforts in wildfire management (Costa-Saura et al. 2022). Moreover, future research could leverage the potential of recently-available spectroscopic spaceborne data together with advanced image analysis techniques like MESMA (Quintano et al. 2023) to determine whether wall-to-wall fire behaviour estimates reported here can be further refined. Recently, the use of unmanned aerial vehicles (UAVs) has enabled accurate predictions of physically meaningful variables as proxies for fire effects at the level of individual trees (Moran et al. 2022; Arkin et al. 2023), and can therefore be a reliable tool in future research to characterise the fire type particularly when fire effects vary greatly at fine spatial scales.

Conclusions

We examined for the first time the potential of broadband satellite data to provide meaningful fire type estimates representative of wildfire behaviour in Mediterranean ecosystems. A key result of this study is that intrinsic biophysical characteristics of burned landscapes, such as the fractional cover of char or photosynthetic vegetation, retrieved from sub-pixel image analysis techniques with a physical basis, are more accurate at assessing fire type (e.g. surface or crown fire), given the direct physical interpretation, than commonly-used spectral indices. For example, post-fire char fraction estimates computed by MESMA can be used by forest managers directly to estimate the CFB and thus determine fire type in distinct ecosystems; i.e. without the need for calibration with field data, unlike spectral indices. Such estimates would enable evaluating both the performance of fire behaviour models and pre-fire treatments as moderators of extreme fire behaviour.