Assessing tree crown fire damage integrating linear spectral mixture analysis and supervised machine learning on Sentinel-2 imagery

ABSTRACT Crown fire damage is a mixture of three principal fire-related components: charred material, scorched foliage, and unaltered green canopy. This study estimated the abundance of these physical alterations in two immediate post-fire Mediterranean forest contexts (Portugal and Italy) by applying linear spectral mixture analysis (LSMA) on Sentinel-2 imagery. The tree crowns fire damage was subsequently mapped, integrating fractional abundance information in a random forest (RF) algorithm, comparing the accuracy resulting from the adoption of generic or image spectral libraries as the primary investigative goal. Although image-derived endmembers resulted in more effectiveness in terms of fire-related components abundance quantification (LMSA-derived RMSE < 0.1), the F-scores always were ≥ 90% whether generic endmembers or image endmembers derived information was employed. The environmental heterogeneity of the two study areas affected the fire severity gradients, with a prevalence of the charred (PT) (45–46%) and green class (IT) (44–53%). Post-fire temporal monitoring was initialized by applying the proposed strategies, and the preliminary results showed a positive recovery trend in forest vegetation from the first year following the fire event, with a reduced charcoal predominance and an increasing proportion of green components.


Introduction
In Mediterranean-type ecosystems, forest fires represent one of the main factors causing ecological and socioeconomic changes, both positive (i.e.regeneration, biodiversity enrichment, vegetative activity stimulation) and negative (i.e.environmental degradation, soil exposure, desertification, habitat vulnerability inducement), at varying degrees, depending on the heterogeneity of the event in terms of severity and spatial distribution (Cascone et al. 2020;Catry et al. 2013;De Luca et al. 2021;Häusler et al. 2018;Morresi et al. 2022;Poon and Kinoshita 2018;Saulino et al. 2020).At the Earth's vegetated surface level, the immediate impact of fire results in the alteration of the vegetation cover, both structural (e.g.canopy and biomass consumption) and chemicalphysical, with the decomposition of the organic matter and photosynthesizing tissues, up to their death or partially/totally consumption (Key and Benson 2006;Lentile et al. 2006).Subsequently, in response to fire events, Mediterranean habitats have activated natural mechanisms to recover prefire ecological conditions.The effectiveness of these mechanisms depends on the characteristics of the fire, its direct physical impact on the land surface, and the degree of ecosystem resilience (Catry et al. 2012;Fernandez-Manso, Quintano, and Roberts 2016;Fernández-García et al. 2018;Frazier, Coops, and Wulder 2015;Gouveia et al. 2010;Morresi et al. 2019).In this context, the delineation of generalized protocols for the qualitative and quantitative detection and measurement of the largescale effects and chemical-physical alterations that fire induces on Mediterranean ecosystems, including both immediate and long terms, is crucial for a better understanding of their dynamics and for planning appropriate post-fire management strategies (Chu and Guo 2013;Chuvieco et al. 2019;Quintano, Fernández-Manso, and Roberts 2020).
Remote sensing techniques have been extensively employed to detect and estimate the effects of fires on ecosystems (Chu and Guo 2013;Chuvieco et al. 2019;Corona, Lamonaca, and Chirici 2008;Gitas et al. 2012).Several scholars have applied satellite multispectral optical imagery to detect burned areas (Chuvieco et al. 2016;De Luca, Silva, and Modica 2022;Mpakairi, Ndaimani, and Kavhu 2020;Otón et al. 2019;Pulvirenti et al. 2020;Santos et al. 2020;Silva et al. 2004;Silva et al. 2019;Sousa, Pereira, and Silva 2003), estimate the degree and spatial distribution of burn severity (De Luca et al. 2021;Fernández-García et al. 2018;Morresi et al. 2022;Saulino et al. 2020), and assess other consequences of a fire on environmental biological and structural features, such as biomass consumption (Garcia et al. 2017) and greenhouse gas emissions (Ostroukhov et al. 2022;Rosa, Pereira, and Tarantola 2011).Most of the methodologies presented in these studies relied on the use of optical spectral indices (e.g.normalized burn ratio, NBR) and/or their temporal difference (e.g.ΔNBR), in which fire-sensitive infrared-based bands (near-infrared, NIR, and shortwave infrared, SWIR) are involved in the construction of indirect models for patterning the optical reflectance of burned materials (Daldegan, Roberts, and Ribeiro 2019).Although highly effective for these purposes, the use of spectral indices for the analysis of real field fire consequences is based on empirical non-linear and non-physical relationships that do not allow the direct estimation and quantification of the elements actually present in a post-fire environment, traditionally assessed using standard field monitoring protocols (De Santis and Chuvieco 2007;De Santis and Chuvieco 2009;Hudak et al. 2007; Lentile et al. 2006;2009;Quintano, Fernández-Manso, and Roberts 2013;Shimabukuro and Ponzoni 2019;Smith et al. 2007;Veraverbeke and Hook 2013).For example, Key and Benson (2006) aimed to standardize the measurements of fire effects across space and time in a context characterized by its variability.Scorched tissues, foliage and wood consumed, unaltered green vegetation, and substrate color change are standard visual parameters generally associated with post-fire vegetation conditions and are used for field-based burn severity estimation (Key and Benson 2006;Veraverbeke and Hook 2013).Fire-affected vegetation surfaces are usually composed of a mixture of combustion products (ash and charcoal), scorched vegetation, and green live vegetation (Quintano, Fernandez-Manso, and Roberts 2017).The relative proportion of each of these elements is related to the degree of burn severity (Tane et al. 2018), reflecting the actual heterogeneity of fire effects (Lentile et al. 2006;Smith et al. 2007;Veraverbeke and Hook 2013).Assuming that the image pixel is a surface unit, the assessment of post-fire effects and regeneration constitutes a sub-pixel issue that can be resolved by quantifying the proportion of each of the abovementioned three components (Gitas et al. 2012;Quintano, Fernández-Manso, and Roberts 2013).
The selection of endmembers plays a crucial role in determining the reliability of the SMA approach (Quintano et al. 2012;Quintano, Fernández-Manso, and Roberts 2013;2020;Somers et al. 2011).Endmembers should be representative of the pure components that constitute the spectral fraction inside image pixels (Somers et al. 2011).In the literature, there is a debate regarding the origin of the endmembers adopted in the SMA, concerning the difference in reliability between endmembers taken from generic spectral libraries (laboratory) and endmembers obtained directly from the image under investigation.While the former returns a pure and generalizable spectral signature for a particular component, the latter involves the actual spectral variability of a specific scene (Fernández-García et al. 2021;Quintano, Fernández-Manso, and Roberts 2013;2017;Rogan and Franklin 2001;Tane et al. 2018).
In this framework, the purpose of this study was to quantitatively estimate and map the proportion of the three spectral components indicative of the main physical effects observable immediately after a fire on tree crowns (charred, scorched, and unaltered green), through the application of linear SMA (LSMA), with the primary objective of evaluating and comparing the quality of the result deriving from the adoption of generic or in-scene spectral libraries on multispectral Copernicus Sentinel-2 data.Most SMA applications for post-fire monitoring were developed on medium/low-spatial-resolution images and, to our knowledge, few scholars (Fernández-Guisuraga et al. 2022;Lewis et al. 2021;Quintano et al. 2019;Quintano, Fernández-Manso, and Roberts 2020;Xu et al. 2022) employed this platform for these purposes.Given the free availability of Sentinel-2 images, which guarantee a high temporal and spatial coverage, an essential prerogative for risk management, this is a noticeable shortcoming.In order to test the robustness of our approach, two fire events that occurred in two different Mediterranean study areas, located in Portugal and Italy, were investigated.Furthermore, from the LSMA results, the crown fire damage maps were constructed using a random forest (RF) machine learning supervised non-parametric algorithm, ensuring high accuracy.Particular attention was paid to investigating and characterizing the mixed fire damage pixels, i.e. incorporating both scorched and green components, by analyzing their fractional abundance.Finally, a preliminary temporal analysis was carried out on the post-fire recovery by applying LSMA to other Senitnel-2 images acquired one year after the fire event.This last aspect represents a critical task as an initiator for future in-depth studies.

Study areas and fire events
The proposed approach was conducted in two Mediterranean study areas (Figure 1).The first area (PT) was located in Algarve (southwestern Portugal, 37°18' N; 08°30' W) at Serra de Monchique (max.elevation 902 m.), occupying almost 26.56 km 2 ; the second area (IT) was in the Province of Reggio Calabria (south Italy, 38°16' N; 15°84' E), at Aspromonte Massif (max.elevation of 1956m), and occupied approximately 18.38 km 2 .Both the study areas have high ecological importance for the respective territory, with the Serra de Monchique being partly included in the European Natura2000 network as a Special Area of Conservation (SAC) (Natura2000 Site Code: PTCON0037), while the Italian Aspromonte area falls inside its homonymous National Park.
Most of the forest cover of the Portuguese site was composed of short-rotation plantations of eucalyptus (Eucalyptus globulus, Labill.1800) and semi-natural evergreen cork oak woodlands composed of Quercus (spp.suber L. and ilex L.), in some cases mixed with secondary native species conforming to the meso-Mediterranean forest ecosystem.Few isolated groups of Mediterranean coniferous species (Pinus pinea L., Pinus pinaster Aiton.)plantations can be observed.The remainder and broader part of the study site is covered by heathlands and/or pastures, composed of typical herbaceous and sclerophyllous shrubby species, in some areas, accompanied by anthropic land cover classes (agriculture, urban) (De Luca, Silva, and Modica 2022;Direção-Geral do Território 2018).
The Aspromonte massif is characterized by high heterogeneity of natural forest flora as a function of the elevation gradient and aspect, represented by typical meso-Mediterranean (Quercus ilex L.) and temperate (Fagus sylvatica L., Abies alba Mill., Quercus petrea (Matt.)Liebl, Pinus nigra supp.laricio, Castanea sativa Mill.) forests species.
In both study areas, fire events occurred during the first fortnight of August (2018 in PT and 2021 in IT).The forest cover was impacted at various degrees of severity in both cases.At the highest severity, only the combustion residues of biomass (ashes, charcoal, and torches) remained, while the bare soil below was exposed.Where the severity of the fire event was lower, the canopy structure was less affected, in part consumed by direct burn or killed by proximal heating, with the dead structural vertical parts (trees and canopies) remaining standing (scorched), while the undercover vegetation was killed and/or partially consumed.Tree canopies, or at least their upper parts, remained green at the lowest severity degrees.

Data and methods
The proposed methodology has been synthesized in Figure 2, in which the main steps have been highlighted.

Satellite dataset and pre-processing
Sentinel-2A Level 2A (Bottom of Atmosphere reflectance) optical imagery was employed to implement LSMA.For each study area, two separate cloud-free images were downloaded, one representing the pre-fire condition (19 July 2018, PT; 28 July 2021, IT) and another immediate post-fire condition (18 August 2018, PT; 16 September 2021, IT).Moreover, additional images were retrieved to conduct a preliminary analysis of the recovery status of the forest vegetation in the first year after the event (13 August 2019, PT; 17 August 2022, IT).

Spectral endmember definition and selection
Four fraction spectral components were defined to unmix Sentinel-2 images, three of which (charred, scorched, and green components) indicated the primary physical effects observable in an immediate post-fire forest environment (Figure 3,top).The charred component (%ch) represented the solid residues from vegetation combustion, with prominent charcoal and light ash at the surface.The scorched component (%sc) was composed of brown dead plant tissues (mostly leaves) and other non-photosynthetically plant material killed by the heat which had been radiated and convected from the fire; most of the vertical structural elements resulted unconsumed by fire (trunk, branches, and foliage).The green vegetation component (%gr) represented the unaffected green foliage cover.The bare soil fraction (%bs) constituted the fourth fire-unrelated component.
Two different types of sources, from which to retrieve the four pure endmembers (%ch, %sc, % gr, %bs) involved in the LSMA, were exploited to assess and compare their debated adequacy for fire effects estimation.The first endmember collection was composed of a set of signatures (unique for both PT and IT) retrieved from three general spectral libraries (laboratory endmembers, LabEnds): the freely-available ASTER v. 2.0 spectral library (Baldridge et al. 2009), the free-available USGS Spectral Library v. 7.0 (Kokaly et al. 2017), and a spectral library provided by the Environmental Remote Sensing and Spectroscopy Laboratory (SpecLab) at the Spanish National Research Council (CSIC), collected in the context of the work of Garcia et al. (2020).In particular, from the ASTER spectral library, the endmembers constituting the green component were distributed as follows: four signatures for cork oak leaves, four for holm oak leaves, and two eucalyptus leaf signatures.Considering that the spectral reflectance of dead vegetation or litter is a valid representative of the scorched component spectrum (De Santis and Chuvieco 2007), this component was accounted for by six signatures of mixed-vegetation leaves litter and grass litter and by three signatures of cork oak litter.From the USGS spectral library, two signatures representing an averaged reflectance burned surface and soils covered by ash and charcoal constituted the charred endmembers; three dry grass and one dry pine needle signatures were representative of scorched vegetation; one signature of oak foliage was taken as a green endmember.From the SpecLab spectral library, 47 signatures of green pine needles were retrieved to form the respective %gr component, and the signatures of the same needles subsequently scorched were considered as %sc endmembers.To ensure strict comparability between all spectral signatures used in the analysis, those retrieved from SpecLab and ASTER libraries were spectrally resampled to the Sentinel-2 spectral response functions v.3.The USGS library was provided already spectrally resampled to Sentinel-2 central bands.
The second group of endmembers was represented by signatures directly chosen from the Sentinel-2 images (image endmembers, ImgEnds), separately for the respective scenes PT and IT.A set of pixels (pixels of interest, POIs) adequately representing the four pure components were visually selected as image endmembers and equally distributed among them (50 pixels for each of the four components).The careful visual interpretation of the POIs was supported using the Esri ArcGIS World Imagery high-resolution satellite map acquired in 2018 (Esri n.d.) concerning the PT study area and the Google Earth Satellite Images acquired in 2021 (Google n.d.) for the IT study area.Moreover, georeferenced photos taken during fieldwork (in both study areas) and direct knowledge of the sites drove the interpretation.
Beyond the endmembers' source, the literature also highlights the importance of using ad hoc techniques to define optimal spectral endmembers.Several automatic endmember selection algorithms have been developed (Boardman, Kruse, and Green 1995;Chang et al. 2010;Dennison and Roberts 2003;Dennison, Halligan, and Roberts 2004;Roberts et al. 2003;Tompkins et al. 1997;Winter 1999) and are commonly used in SMA approaches (Daldegan, Roberts, and Ribeiro 2019;Fernandez-Manso, Quintano, and Roberts 2016;Fernández-Manso, Quintano, and Fernández-Manso 2009;Nascimento and Dias 2005;Quintano, Fernández-Manso, and Roberts 2020;Tane et al. 2018).A pioneering method is represented by the pixel purity index (PPI) (Boardman, Kruse, and Green 1995), which has been widely exploited in literature (Chang et al. 2010;Chang and Plaza 2006;Chang and Wu 2015;Fernández-Manso, Quintano, and Fernández-Manso 2009;Heylen and Scheunders 2013;Nascimento and Dias 2005;Plaza and Chang 2005;Suryoprayogo, Ramdani, and Utaminingrum 2018).Having acquired three multi-source groups of spectral signatures (ImgEnds_PT, ImgEnds_IT, and LabEnds) (Figure 3), the four single endmembers to be used in LSMA (representing the %ch, %sc, %gr, and %bs components, respectively) were thus extracted by implementing the PPI algorithm.It repeatedly projects the spectral information on a large number of random vectors (skewers), counting the number of times each pixel as an extreme for each skewer and choosing the pixels most often recorded as extreme (Chang et al. 2010;Nascimento and Dias 2005).We set the number of skewers to 1000, following Nascimento and Dias (2005), who found no improvements in pixel purity above this threshold.

Spectral unmixing and fraction image extraction
Once the four final spectral endmembers were chosen, the proportion (i.e.relative abundance) of each pure component contained within each pixel of the image was estimated by applying a prebuilt (Therien n.d.) fully constrained least squares (FCLS) model.It adopts both non-negative and sum-to-one constraints by solving quadratic programming, as proposed by Vishnevskiy (2013) and based on the CVXOPT python library (CVXOPT n.d.).The least-squares approach family is a standard mathematical procedure used by the LSMA to estimate abundance.This method estimates the proportion of each endmember within a pixel, minimizing the sum of squared errors (residuals) (Shimabukuro and Ponzoni 2019;Shimabukuro and Smith 1991).For each of the four endmembers searched, LSMA decomposed each pixel of the multispectral input image as a weighted linear combination of the pure component spectral responses and generated an image, commonly defined as a fraction image (FI) or abundance map, whose pixels contain the abundance (the proportion, expressed in grayscale pixel value) of the respective spectral component (Quintano et al. 2012;Quintano, Fernández-Manso, and Roberts 2013;Roberts, Smith, and Adams 1993;Settle and Drake 1993;Shimabukuro and Ponzoni 2019;Tane et al. 2018).For each study area, LSMA was approached separately for LabEnds and ImgEnds.
The performance of the LSMA model was assessed by calculating the root mean squared error (RMSE, Equation 1) from the model residual errors (ε λ ) (Daldegan, Roberts, and Ribeiro 2019): where NB is the number of image bands, and λ indicates the specific single band.Assuming the general SMA model construct (Eq. 2): where p iλ is the reflectance of a single-pixel i in a specific band λ; en jλ is the reflectance of a specific endmember j in a specific band λ; FI enji is the fractional abundance of the endmember j in the specific pixel i; NE is the total number of endmembers.It assumes that the pixel is the sum of the reflectance of each spectral signature that composes it proportionally to their abundance.
The residuals represent the differences between the original pixel mixed signature and the sum of the proportional fraction of each target endmember signature found in that pixel (Daldegan, Roberts, and Ribeiro 2019;Fernández-Manso, Quintano, and Fernández-Manso 2009;Shimabukuro and Smith 1991).

Burned forest area masking
Since the analysis of fire effects was focused on tree crowns, it was necessary to delineate the forest areas contained within the fire perimeter to avoid confusion with non-forest land cover types.Two masks were used in this study to refine the analysis, enabling the generalization of the approach to the entire European territory.First, the burned areas of both study areas were detected by implementing vector maps provided by the European Forest Fire Information System (EFFIS n.d.), which were retrieved by a semi-automatic classification of MODIS and Sentinel-2 satellite imagery.Afterward, the forest cover was isolated using the European forest type map for 2018 (10 m pixel resolution) provided by the Land Monitoring Service of ESA Copernicus (Copernicus Land Monitoring Sevice 2018), where each pixel classified as forest presented a tree cover density above 10%, following the Food and Agriculture Organization (FAO) definition (FAO 2020).

Crown fire damage mapping and accuracy assessment
The well-known random forest (RF) machine learning algorithm (Breiman 2001) was applied to obtain a tree crown fire damage map involving the four FIs and the respective delta images (ΔFIs; the difference between a post-fire FI and the FI of the same component calculated before the event).The optimal RF parameters were set by applying a cross-validated grid-search procedure to a pre-selected group of values.The coordinates of the pixels used to define the ImgEnds (POIs) for IT and PT were used as a reference for the choice of training pixels, in addition to other 150 training pixels for each of the four components, specifically chosen for this purpose with the same methods performed for POIs.The pixel-based classification was performed separately for LabEnds-and ImgEnd-derived FIs after defining four classes: Charred, Scorched, Green, and Soil.The confusion matrices were computed to obtain the single-class producer's and user's accuracies, and their derived harmonic means single-class F-score i (Eq. 3) and multi-class F-score M (Eq.4) metrics (Congalton and Green 2019;Modica et al. 2021;Sokolova and Lapalme 2009).The F-score M was calculated from the single-class producer's and user's values averaged among the four classes (Equations 5 and 6), and describes the overall accuracy of the map: where i represents the single class and n represents the total number of classes.

Crown fire mixed damage characterization
An additional group of 200 pixels representing mixed crown fire severity categories (tree crowns partially scorched and partially green) was visually selected using the same procedure adopted for POIs.The statistical distribution of the fraction abundance of each component falling inside these pixels was plotted and examined to characterize these mixed pixels.

Preliminary analysis of post-fire recovery
The image pre-processing and LSMA described above were expanded on the images acquired in the year following the respective fire events (August 2019 for PT; August 2022 for IT) to initiate a preliminary investigation of post-fire vegetation recovery.Adopting the predictive information (endmembers and training POIs) employed in the classification of immediately post-fire data (2018 for PT; 2021 for IT), the FCLS model was applied, and the respective FIs were retrieved for each observed post-fire year.After that, RF was applied to them to compare the changes in the area distribution of each crown damage class during the observed years.Ternary plots and boxplots were plotted to describe the distribution of FIs among the three crown fire damage classes (Charred, Scorched, and Green) across the post-fire temporal trend.

Selected endmembers
Figure 4 displays the four endmembers (EM1, EM2, EM3, and EM4) selected by the PPI index for each adopted configuration (LibrEnds, ImgEnds_PT, and ImgLibrsIT), spectrally resampled to the Sentinel-2 band centers.Although all EMs in both LibrEnds and ImgEnds configurations present similar spectral patterns, differences are noticeable, especially in the absolute reflectance values and the slope of the signature curves.EM1 represents the charred component of the spectrum (%ch), characterized by a linear and slightly increasing pattern given by the presence of light ash.In contrast, EM4 presents the typical patterns of green vegetation, with a maximum pic at NIR and a sudden decrease at higher SWIR wavelengths.The remaining EMs had a more variable pattern.Comparing the EM2 (representing the %sc component) with the respective green vegetation signature (EM4), the former presents higher reflectance at VIS wavelengths, and it is absent the inflection at the level of the green band (typical of photosynthesizing vegetation spectral behavior) in favor of red band (characterizing the typical color of senescent vegetation); it is perceptible a lowering and consequent flattening of the reflectance in the Red-Edge and NIR wavelengths.The spectral signature of bare soil (EM3) is similar to that of EM1 and has comparable patterns to the EM2 signature in LibrEnds and ImgEnd IT configurations.

Derived fraction images
The FIs generated by employing the FCLS spectral mixture model to the images acquired immediately after the fire (2018 for PT, 2021 for IT) and representing the proportion of each of the four components (%ch, FI-1; %sc, FI-2; %gr, FI-3; %bs, FI-4) are shown in Figures 5 (LibrEnds_PT), 5 (ImgEnds_PT), 6 (LibrEnds_IT), and 7 (ImgEnds_IT).The figures illustrate a portion of the scene where all four components are clearly observable.The value of each pixel is directly associated with the proportions (abundance) of each of the four respective endmembers selected through the PPI index in a grayscale range normalized between 0 (black) and 1 (white) to improve visualization.This is more evident in the RGB false-color image (Figures 5-8, lower left corner), in which the three fire-related FIs were combined (Red = FI-2, %sc; Green = FI-4, %gr; Blue = FI-1, %ch).In PT (Figures 5 and 6), the prevalence of the %ch component in ImgEnds FIs, compared to those retrieved from LibrEnds in terms of proportion and occupied surface, is perceptible by the grayscale palette, where the higher the proportion, the brighter the pixel color.Notably, in LibrEnds, the %ch component is totally excluded from the surfaces covered by unburned and scorched forest vegetation, unlike in ImgEnds, where %ch abundance seems equal to the under-cover bare soil.This was predictable, considering the similarity of the EM3 and EM1 spectral signatures (Figure 4).FI-2, hence the %sc component, is correctly marked on the forest vegetation in ImgEnds, while it appears to be slightly confused with the soil in LibrEnds.The green component appeared moderately distributed even among the scorched forest cover in LibrEnds, while concentrated in specific zones in ImgEnds.
We can observe the same characteristics in the IT study area.The %ch (FI-1) component was concentrated in a few areas in LabEnds (Figure 7), whereas, in ImgEnds (Figure 8), it was slightly distributed on the scene, matching the %bs component.The detected distribution of % gr and %sc was similar in both combinations.However, the former seems slightly (and correctly)     distributed over some areas of moderate fire severity (where the vegetation was not entirely scorched) in LibrEnds.On the other hand, the %sc presents a higher abundance (brighter pixels) in ImgEnds.

LSMA model error maps
The spatial distribution of the per-pixel error (RMSE) resulting from the LSMA model was mapped, and it is shown in Figures 9-12 for each combination, respectively.The RMSE accounts for the difference in reflectance (0, 1) of the original mixed pixel versus the amount of that of the endmembers, proportional to the fractional abundance of each endmember computed by the LSMA.The respective boxplots describe the statistical distribution of the RMSE among the POIs representative of four fire-related classes (200 per class) that had been retrieved as described in Section 3.5.It emerges that the RMSE of the analysis performed with LibrEnds is higher than that performed with ImgEnds, but never actually exceeding 0.3 for PT and 0.4 for IT.Moreover, it is noticeable that the highest RMSE values are reached by vegetation-related classes: Green, Scorched, and Blended (the latter representing the mixed crown fire severity damages).

Crown fire damage map classifications and accuracy assessment
The RGB-probability false color maps (Figures 13-16), in which each band contains the probability value of the specific fire-related classes (R, scorched; G, Green; Blue, Charred) retrieved from the RF learning process, represent the spatial distribution of the classified crown fire damages.The classified areas were masked, as described in Section 3.4.
In particular, the distribution of the crown damage classes in terms of area occupied (km 2 ) is reported by the respective ring plots in the Figures.Notably, the level of difference among the endmember source-based combinations in the same study area is low, pointing out a slight Figure 17 shows also the confusion matrices that report the resulting classification accuracy for each configuration.The overall accuracy represented by the multi-class F-score M , derived from the producer's and user's metrics combination, never dropping below 90% in all cases.Some points related to the scorched class were confused with the Charred or the Green class in PT.Some confusion between Charred and Soil was also observed.In IT, although the error relationships between the different classes seemed similar to those for PT, accuracies were higher.Observing the feature importance (Figure 18), reporting the contribution ranking of each input layer in the RF learning process, the %gr fraction (FI-4) and the relative Delta band (ΔFI-4) appear to be the variables that contribute the most to the modeling process in all configurations.The scorched fraction (FI-2) followed in terms of contribution to the learning process for the study area PT, while in IT, the charred fraction (FI-1) was the second variable in terms of importance.Both scorched and charred components for PT and IT reflected feature importance values never less than 50% of the green component importance value.

Mixed crown fire severity damages analysis
The boxplots in Figure 19 describe the statistical distribution of the fractional abundance of the three main fire-related components among a group of 200 POIs representing partially scorched and partially unburned green tree crowns (i.e.mixed fire severity class).Observing the results, it is evident that the abundance of scorched and green components is dominant compared to the charred component, which never reaches a median value higher than 0.1.The scorched component tends to be slightly higher than green in both ImgEnds configurations, while the opposite happens in LibrEnds.
The boxplots are corroborated by the ternary plots (Figure 20), in which the distribution of the three fire-related components' fractional abundance for each POI is expressed as a position in an equilateral triangle.The point clouds were fully allocated to the triangle's left side for all configurations.Most of the points present a high abundance of scorched and green complementary components.The charred component is practically absent.

Temporal analysis
Figures 21 and 22 show the results of the temporal overview of crown fire damage progression relating to both PT and IT study areas for LibrEnds and ImgEnds, respectively.In particular, the RGB false-color maps reporting the proportion of the three fire-related components are reported for each observation period.Below each RGB map, the ternary plot shows each pixel's proportional distribution among the fractional abundance of the three components.
Beyond the expected increase of charred and scorched components immediately after the fire, their decrease is detectable one year after the fire event, in conjunction with the increased distribution of green components.This is confirmed by the point cloud shown in the ternary plots (shifted in the bottom-left corner).This condition is observable in all configurations, although slight distributional differences exist.

Source-related endmember extraction
Endmembers should be spectrally representative of the target features.Several authors (Fernández-García et al. 2021;Quintano, Fernández-Manso, and Roberts 2013;2017;Shimabukuro and Ponzoni 2019) have suggested that the selection of the endmembers directly from the image is more suitable because: i) they are easy to retrieve; ii) they are in the same radiometric scale of the analyzed image; iii) they incorporate the same spectral corrections of the analyzed image; iv) they involve the specific spectral variability of the scene.On the other hand, a higher spectral purity of target components can be obtained from signatures measured in the laboratory, although they do not account for the atmospheric influence and/or image noise generally addressed with calibration and correction techniques (Drake, Mackin, and Settle 1999; Quintano, Fernández-Manso, and Roberts 2013;

2017
).However, it is difficult to determine the ideal conditions in which a pure component fully occupies the image pixels.This is partly observable in Figure 4, where the signatures derived from both lab libraries (LibrEnds) and images (ImgEnds) are compared.The LibrEnds endmembers showed a more marked distinction between the four components, especially in terms of absolute reflectance values.ImgEnds, on the other hand, exhibit apparent phenomena of impurities (more realistic), with the ground and charred or the ground and scorched having very similar shapes and reflectance values in PT and IT, respectively.Nevertheless, the RMSE distribution highlights a discrepancy between the two sources, with the ImgEnds maintaining a low error level across the entire study areas.The higher error returned by LibrEnds may be due to the difference between these and the actual spectral signatures contained in the pixel or to the presence of additional spectral signatures not accounted by the endmembers.Although it is plausible that both factors were equally influential, further investigation may clarify the eventual effects of the different spectral Figure 17.On the left, the classification differences maps show the difference between the two classified maps deriving from the different endmembers-sources of the same study site.When the difference is equal to 0, it means no difference between classifications; when the difference is different from 0, the classification returns a different label for that pixel.On the right, the respective confusion matrices with the derived single-class F-score i and multi-class F-score M metrics are computed for each combination.
scales caused by the different spectral sources used to select LibrEnds.Nevertheless, Tane et al. (2018) reported that using spectra collected at different locations and spatial scales is a common practice to model cover fractions and suggested that it could be easily converted for fire scar assessment, at least at the regional scale.The same authors found some biases when endmembers specific to the target study area were used (at coarse spatial resolution).In Quintano, Fernández-Manso, and Roberts (2013), some other authors are reported affirming that the three fire-correlated components (%sc, %ch, and %gr) are spectrally quite constants across an extensive range of ecosystems, implying the suitability of general endmembers.Vereverbeke and Hook (2013) indicated similar conclusions after analyzing the three main components that highly correlated with field-based measurements (brown, black, and green).Sunderman and Weisberg (2011) affirmed that, although using field-measured endmembers might improve the results, their applicability to other study areas should be excluded.

Crown fire damage map
The LSMA applied to Sentinel-2 imagery successfully characterized the effects of fire on Mediterranean forest crowns.The satellite spectral information was translated into physical information as sub-pixel fractions of the three fire-related components (charred, scorched, green) and one fireunrelated class (soil).
The RF algorithm was a high-performing predictor, as confirmed by the accuracy levels achieved (> 90%).This is also confirmed by the extensive literature that uses the RF for classification purposes, with almost always superior results in accuracy and time-consuming efficiency, compared to other classification approaches.Furthermore, despite the mathematical complexity of machine learning, these models are now easy to implement thanks to the availability of numerous frienduser libraries, toolkits, and software.The probability score could maintain the physical meaning given by the fractional proportion of crown fire damage classes, and, at the same time, it enables the integration of auxiliary layers, such as the delta layers (Δ).This type of information allows the vegetation state to be accounted for following the definition of fire severity (Key and Benson 2006).
Despite the low error levels, the confusion matrix shows that some pixels belonging to the green class were mistakenly classified as scorched.This is admissible considering the 'contaminations' that might occur, for example, when both components contribute to the spectral signature of the pixels (mixed pixels) due to the presence of still green foliage (or undergrowth) among scorched canopies or when fire-induced stress affects the surrounding live green vegetation, weakening its spectral response, or inversely (Smith et al. 2007).Similarly, the confusion between Green and Charred, or between Scorched and Charred classes, can be explained by the possible presence and exposure of a burned undergrowth char layer covering the ground, overtopped by a partially (or completely) green canopy, or by scorched tree crowns (De Santis and Chuvieco 2007;Rogan and Franklin 2001), or litter.Additional considerations regarding mixed pixels are discussed in Section 5.3.
As reported in the literature, the charred component tends to predominate over the other firerelated components (scorched and green), making them negligible for modeling fire severity  (Quintano et al. 2019).Rogan and Franklin (2001) affirmed that char exerts such a strong influence on the surface, drastically reducing its brightness that its spectral signature clearly distinguishes it from other components.The charred component is more directly correlated to the biomass portion lost/consumed and carbon emissions (Quintano, Fernández-Manso, and Roberts 2020), and it is considered an effective quantitative indicator of the actual physical effects of fire on vegetation ( Lentile et al. 2006).For this reason, several studies have proven that it can be used as a unique estimator of burn severity, enabling higher accuracy than vegetation indices (Fernández-Manso, Quintano, and Fernández-Manso 2009;Hudak et al. 2007;Lentile et al. 2006;2009;Quintano, Fernández-Manso, and Roberts 2013;2017;2020;Smith et al. 2007;Veraverbeke and Hook 2013).Quintano, Fernández-Manso, and Roberts (2020) observed that the charcoal proportion and the evapotranspiration driver contributed the most to the burn effect model.However, in the present analysis, a high contribution from the green fraction was estimated during the learning process.Low-severity fires affected the study areas mainly (IT) or partially (PT), which could explain the higher contribution of the green fraction.The characteristics of the environment should also be considered to understand the differences between the study areas better.Most of the PT area was formed by forest cover, surrounded by shrubs and/or herbaceous vegetation, and other undergrowth was settled under the forest cover, inducing the formation of higher severity and intensity fires.The IT study area, on the other hand, was less heterogeneous and less covered by shrubby, reducing the spread of high severity fire and the consequent consumption of vertical structural forest layers.What is analogous to the study of Quintano et al. (2019) is the poor influence of soil fraction on fire severity characterization.In ImgEnds retrieved for the PT area, for example, the %bs component is partially redundant because of its close signature to the %ch component.This can be explained by citing the study of Tane et al. (2018), in which the authors noted that separating soil and charred areas was particularly difficult, even with images at a higher resolution than Sentinel-2.They claimed this was partially due to spectral impurities caused by a mixture of soil, ash, and charcoal.Qualitatively, the confusion between charcoal and soil in an area previously occupied by forest vegetation can be a minor error, as in both cases, it would describe the partial or total destruction of the epigeal structure of the trees.
The difficulty of correctly unmixing images without impurities is as significant as its resolution (Tane et al. 2018).Nevertheless, it is probable that the higher spatial resolution provided by Sentinel-2 inevitably contributed to the good results obtained in this study.The main positive factors of this platform are its free availability and high temporal resolution (which guarantees a revisit time of 3-5 days), substantial privilege for risk management.

Mixed severity classes
Through SMA, it was possible to characterize the mixed fire severity classes.These are characterized by scorched vegetation, heavily affected by fire (especially at the under-and/or lower layers of the canopy), but with a part of the unaffected green structure still present (generally at the upper layers of the canopy).According to the well-known fire severity classification protocol proposed by Key and Benson (2006), the just described scenario can be categorized as a moderate-severity fire.Figures 19 and 20 show that the fractional abundance related to the pixels falling into this category is characterized by variable values that alternate between the scorched and green components.The precise identification of an explicit crown mixed severity class seems difficult; however, the results obtained here will be helpful for future implementation.

Post-fire recovery
Temporal monitoring revealed an evident incremental recovery activity from the first year after the fire vent (Figures 21 and 22).In particular, the Charred class was almost entirely reduced, owing to the combined effect of the absorption and degradation of this component by the environment.At the same time, an increase in the class representing regrown vegetation (Green) was observed.Most of the post-fire green recovery is due to pioneer herbaceous and shrubby vegetation, which grow rapidly after fire occurrence and can be confused with tree canopies.Several authors have stated that charred fraction could be an optimal indicator of post-fire effects (Fernández-Manso, Quintano, and Fernández-Manso 2009;Hudak et al. 2007;Lentile et al. 2006;2009;Quintano, Fernández-Manso, and Roberts 2013;2017;2020;Smith et al. 2007;Veraverbeke and Hook 2013).In Smith et al. (2007), instead, the charcoal component gave more consistent results compared to green and similar to those obtained by ΔNBR, analyzing the post-fire conditions of the scene one year later.They assumed this was caused by the delayed mortality of trees, which may take months to become apparent.This also makes scorch components more controversial to interpret.However, Fernandez-Manso, Quintano, and Roberts (2016) represented post-fire recovery as a temporal series of green vegetation fraction images in a more specific context.
During the examined timeframe, the green class did not reach the pre-fire levels regarding spatial distribution.This is because of the presence of scorched tree areas, whose brown cover remains on the surface and continues to have a characteristic spectral signature.Presumably, a minor contribution might have been given by the delayed vegetation mortality, expected for such higher intensity fires, which causes both reduction of green tree cover and increased soil exposure (Smith et al. 2007).Although they need refining, the results of the temporal analysis show the effectiveness of SMA in tracking temporal post-fire recovery trends and improving their physical interpretation.However, the preliminary nature of these analyses does not allow for a more thorough discussion.More investigations must be conducted, analyzing a wider temporal frame (medium-long term) and validating the observations with ground data.

Final findings and outstanding issues
The findings of this study show the high effectiveness of SMA methods in delineating the high heterogeneity that characterizes crown fire damage (Lentile et al. 2006).Veraverbeke and Hook (2013) observed that using SMA is necessary rather than complementary information when quantitative physical parameters relative to post-fire effects need to be examined (e.g. the biomass lost).For example, estimating biomass consumption or carbon emissions requires the quantification of complex physical indicators (e.g.charred component), which could be difficult to obtain only through the empirical relationship offered by vegetation index-based models.
The most significant contribution of this study was comparing the results derived from two different sources from which the endmembers were obtained.The derived findings could follow two separate strands: for the physical assessment of the components related to fire damage on the trees' crowns, the spectral signatures obtained from the image itself have been shown to return lower errors (RMSE), which is a symptom of a higher correlation with the actual situation on the Earth's surface.On the other hand, if the objective is to classify and subsequently map the spatial distribution of the fire-related components, both sources achieve satisfactory accuracy values when modeled using advanced machine learning algorithms.For this last use, generic libraries speed up the workflow and make it generalizable to multiple environments (at least on multiple study areas in the Mediterranean environment), to the detriment of capturing local realistic variability.The generic origin of the spectral library made it spatially and temporally generalized, a fundamental prerogative to reduce the slight variability in reflective characteristics of a specific cover across the scene, which could be caused by different environmental and site-specific variables (sunlight angle, topographic effects, site-specific soil characteristics, canopy exposition, and plant-plant specific characteristics) (Quintano, Fernández-Manso, and Roberts 2013;Shimabukuro and Ponzoni 2019;Somers et al. 2011).Some authors have effectively exploited the advantages of the combined use of endmembers derived from different sources (e.g.Rogan and Franklin 2001;Tane et al. 2018).Fernandez-Manso, Quintano, and Roberts (2016) incorporated both image and reference endmembers to build an initial spectral library and then defined the optimal spectra using selection indices.
The study confirmed the suitability of LSMA on Sentinel-2 multispectral data for the tree crown fire damage characterization, given the high spectral and temporal resolution.In particular, using a simple LMSA made it possible to speed up and simplify the workflow with good accuracy.One of the main objectives, in fact, was the construction of a generalizable workflow for mapping the physical damages on the forest canopy (posing the problem of the endmembers choosing and of the mixed class identification).However, other methods can also be explored when more complex analyses (such as post-fire recovery) need to be carried out.Some scholars have proposed advanced SMA techniques, such as multiple endmember SMA (MESMA), which has proven to be very robust because of its ability to account for the variability existing within endmember classes (Fernandez-Manso, Quintano, and Roberts 2016;Fernández-García et al. 2021;Quintano, Fernández-Manso, and Roberts 2013;2017;2020;Tane et al. 2018;Veraverbeke and Hook 2013).On the other hand, the integration of either multi-source (e.g.SAR and LiDAR) and higher spectral resolution information (e.g. the incoming CHIME by ESA) and the assessment of their contribution to improving the results might be considered.

Conclusions
This study investigated the integration of LSMA and machine learning on multispectral Sentinel-2 satellite imagery to analyze, quantify, and map three main physical components observable at the forest tree crown level as effects of fire occurrence: charred, scorched, and unaltered green foliage components.The results complement the findings of previous studies that supported the use of SMA in mapping fire effects and fire severity.The effectiveness of a post-fire effect analysis method depends on its ability to capture the high variability of the affected environmental features.The LSMA fraction products are more suitable for dealing with this issue, taking advantage of the direct relationship between the abovementioned sub-pixel fraction components and the fire severityrelated parameters revealed using well-known protocols, such as the composite burnt index protocol (Key and Benson 2006).
The main challenge for SMA is defining an adequate group of endmembers.In this study, two sources were employed and compared.Although the error achieved in the LSMA process resulted in higher for endmembers derived from generic libraries, complicating their physical quantification, no significant differences were observed when the machine learning-based classification process was conducted to retrieve the crown fire damage maps.
Further analyses are needed in different Mediterranean study areas to evaluate their effectiveness better and to encourage the development of improved techniques involving the increasing availability of free images at finer spatial and temporal resolutions (e.g.Sentinel-2) and open-source software.Considering this last aspect, the advantageous operability and increased sharing capacity are guaranteed by Python-based libraries.

Disclosure statement
No potential conflict of interest was reported by the author(s).

Figure 1 .
Figure 1.Top figure: the location of the study areas in Portugal and Italy, respectively.Below is the overview of the study areas (Portugal, left; Italy, right) on a false-color RGB map (Red = SWIR 2202 ; Green = NIR 833 ; Blue = Red 665 ), with the burned area highlighted by the red areas.On the bottom, a panoramic view of Portuguese and Italian study area landscapes shows the heterogeneity of the crown damage level on the forest vegetation.

Figure 2 .
Figure 2. Flowchart schematizing the data and processes that characterized the present study.

Figure 3 .
Figure 3.The four main fraction spectral components and the respective endmembers as retrieved from two different sources: laboratory libraries (USGS, ASTER, and SpecLab) and image pixels.Three components are typically observable in a forest environment after a fire event (charred, %ch; scorched, %sc; unaltered green vegetation, %gr); one component is fire-unrelated (bare soil, %bs).

Figure 5 .
Figure 5. Fraction images (FIs) for PT study area generated by ImgEnds endmembers, illustrating a representative portion of the study area, the first four upper square panes report: FI-1 (charred component, top left), FI-2 (scorched component, top right), FI-3 (soil component, middle left), FI-4 (green component, middle right).Below, on the bottom left, the RGB false color map is reported (Red: FI-2, Green: FI-4, Blue: FI-1); on the other side, there is the satellite (Esri, n. d.) base map (BASE) showing the immediate post-fire damage in true colors.

Figure 6 .
Figure 6.Fraction images (FIs) for PT study area generated by ImgEnds endmembers, illustrating a representative portion of the study area, the first four upper square panes report: FI-1 (charred component, top left), FI-2 (scorched component, top right), FI-3 (soil component, middle left), FI-4 (green component, middle right).Below, on the bottom left, the RGB false color map is reported (Red: FI-2, Green: FI-4, Blue: FI-1); on the other side, there is the satellite (Esri, n. d.) base map (BASE) showing the immediate post-fire damage in true colors.

Figure 7 .
Figure 7. Fraction images (FIs) for IT study area generated by ImgEnds endmembers, illustrating a representative portion of the study area, the first four upper square panes report: FI-1 (charred component, top left), FI-2 (scorched component, top right), FI-3 (soil component, middle left), FI-4 (green component, middle right).Below, on the bottom left, the RGB false color map is reported (Red: FI-2, Green: FI-4, Blue: FI-1); on the other side, there is the satellite (Esri, n. d.) base map (BASE) showing the immediate post-fire damage in true colors.

Figure 8 .
Figure 8. Fraction images (FIs) for IT study area generated by ImgEnds endmembers, illustrating a representative portion of the study area, the first four upper square panes report: FI-1 (charred component, top left), FI-2 (scorched component, top right), FI-3 (soil component, middle left), FI-4 (green component, middle right).Below, on the bottom left, the RGB false color map is reported (Red: FI-2, Green: FI-4, Blue: FI-1); on the other side, there is the satellite (Esri, n. d.) base map (BASE) showing the immediate post-fire damage in true colors.

Figure 9 .
Figure 9. Root mean square error (RMSE) spatial distribution for PT study area generated by LSMA when LibrEnds were employed.On the top-left, the map of the entire study area is shown, while two detailed overviews are displayed on its sides (on the right).The box plot reports the distribution of RMSE among four groups of pixels (POIs) representing three pure crown damage classes (Green, Scorched, Charred), one mixed class representing partially scorched and green crowns (Blended), and the Soil class.

Figure 10 .
Figure 10.Root mean square error (RMSE) spatial distribution for PT study area generated by LSMA when ImgEnds were employed.On the top-left, the map of the entire study area is shown, while two detailed overviews are displayed on its sides (on the right).The box plot reports the distribution of RMSE among four groups of pixels (POIs) representing three pure crown damage classes (Green, Scorched, Charred), one mixed class representing partially scorched and green crowns (Blended), and the Soil class.

Figure 11 .
Figure 11.Root mean square error (RMSE) spatial distribution for IT study area generated by LSMA when LibrEnds were employed.On the top-left, the map of the entire study area is shown, while two detailed overviews are displayed on its sides (on the right).The box plot reports the distribution of RMSE among four groups of pixels (POIs) representing three pure crown damage classes (Green, Scorched, Charred), one mixed class representing partially scorched and green crowns (Blended), and the Soil class.

Figure 12 .
Figure 12.Root mean square error (RMSE) spatial distribution for IT study area generated by LSMA when ImgEnds were employed.On the top-left, the map of the entire study area is shown, while two detailed overviews are displayed on its sides (on the right).The box plot reports the distribution of RMSE among four groups of pixels (POIs) representing three pure crown damage classes (Green, Scorched, Charred), one mixed class representing partially scorched and green crowns (Blended), and the Soil class.

Figure 13 .
Figure 13.The RGB-probability map generated for the PT study area using a LibrEnds-derived dataset.It reports the probability proportion of the three crown fire damages classes (Red: probability score for Scorched class; Green: probability score for Green class; Blue: probability score for Charred class) false color combination.On the top-left, the map of the entire study area is shown, while two detailed overviews are displayed on its sides (on the right).The ring plot reports the surface distribution of the four main classes resulting from the classification using the random forest algorithm.

Figure 14 .
Figure 14.The RGB-probability map generated for the PT study area using an ImgEnds-derived dataset.It reports the probability proportion of the three crown fire damages classes (Red: probability score for Scorched class; Green: probability score for Green class; Blue: probability score for Charred class) false color combination.On the top-left, the map of the entire study area is shown, while two detailed overviews are displayed on its sides (on the right).The ring plot reports the surface distribution of the four main classes resulting from the classification using the random forest algorithm.

Figure 15 .
Figure 15.The RGB-probability map generated for the IT study area using a LibrEnds-derived dataset.It reports the probability proportion of the three crown fire damages classes (Red: probability score for Scorched class; Green: probability score for Green class; Blue: probability score for Charred class) false color combination.On the top-left, the map of the entire study area is shown, while two detailed overviews are displayed on its sides (on the right).The ring plot reports the surface distribution of the four main classes resulting from the classification using the random forest algorithm.

Figure 16 .
Figure 16.The RGB-probability map generated for the IT study area using an ImgEnds-derived dataset.It reports the probability proportion of the three crown fire damages classes (Red: probability score for Scorched class; Green: probability score for Green class; Blue: probability score for Charred class) false color combination.On the top-left, the map of the entire study area is shown, while two detailed overviews are displayed on its sides (on the right).The ring plot reports the surface distribution of the four main classes resulting from the classification using the random forest algorithm.

Figure 18 .
Figure 18.Feature Importance resulted from RF classification and calculated for the four different combinations in terms of study area and endmembers source, respectively: LibrEnds_PT (top left), ImgEnds_PT (top right), LibrEnds_IT (bottom left), ImgEnds_IT (bottom right).

Figure 19 .
Figure 19.Statistical distribution of the fractional abundance of a group of 200 pixels of interest (POIs), representing tree crowns partially scorched and partially green (mixed fire severity pixels) among the three fire-related components (charred, scorched, and green).The four plots represent the four different combinations in terms of the study area and endmembers source, respectively: LibrEnds_PT (top left), ImgEnds_PT (top right), LibrEnds_IT (bottom left), ImgEnds_IT (bottom right).

Figure 20 .
Figure 20.Ternary plots reporting the proportional distribution of a group of 200 pixels (POIs), representing partially scorched and partially green trees crowns (mixed fire severity pixels), among the fractional abundance (FIs) of the three fire-related components (green, scorched, and charred).The four plots represent the four different combinations in terms of the study area and endmembers source, respectively: LibrEnds_PT (top left), ImgEnds_PT (top right), LibrEnds_IT (bottom left), ImgEnds_IT (bottom right).

Figure 21 .
Figure 21.The results of the temporal overview of crown fire damage progression calculated for both configurations LibrEnds (top) and ImgEnds (bottom), relative to the PT study area are displayed.The RGB-probability false color images (R, probability image for the Scorched class; G, probability image for the Green class; B, probability image for the Scorched class) are reported for each observation period (pre-fire, left column; immediately post-fire, middle column; one-year-after post-fire, right column).The ternary plots express the proportional distribution of each pixel, classified in one of the three-crown fire damage classes, among the fractional abundance (FIs) of the three fire-related components (green, scorched, and charred).

Figure 22 .
Figure 22.The results of the temporal overview of crown fire damage progression calculated for both configurations LibrEnds (top) and ImgEnds (bottom) relative to the IT study area are displayed.The RGB-probability false color images (R, probability image for the Scorched class; G, probability image for the Green class; B, probability image for the Scorched class) are reported for each observation period (pre-fire, left column; immediately post-fire, middle column; one-year-after post-fire, right column).The ternary plots express the proportional distribution of each pixel, classified in one of the three-crown fire damage classes, among the fractional abundance (FIs) of the three fire-related components (green, scorched, and charred).

Funding
Giandomenico De Luca was funded by the European Commission and the Regione Calabria with the POR Calabria FESR FSE 2014-2020source [CUP C39B18000070002].João M. N. Silva was funded by the Forest Research Centre, a research unit funded by Fundação para a Ciência e a Tecnologia IP (FCT), Portugal (UIDB/00239/2020), and by the project FireCast -Forecasting fire probability and characteristics for a habitable pyro environment, funded by FCT (PCIF/GRF/0204/2017).