1. Introduction
Fire is one of the main disturbance drivers of ecological change in many terrestrial ecosystems worldwide [
1,
2]. In the Western Mediterranean Basin, around half a million forest hectares are affected by wildfire disturbances annually, with significant implications for terrestrial carbon budgets, land surface energy fluxes and thus for the whole climate system [
3]. At local scales, wildfires have shaped historical landscape pyrodiversity patterns in this region [
4] and may have important consequences on the ecosystems’ resilience, function, and stability [
5], as well as on the structure, composition, and ecological interactions within fire-prone plant communities [
6,
7]. The local to regional feedback can be strongly amplified in the terrestrial ecosystems of the Mediterranean Basin as a consequence of global change drivers [
8], including anthropogenic climate change [
9], conducive to more frequent wildfire events of extreme fire behavior.
Fire severity is one of the fire regime attributes most commonly used worldwide to describe the magnitude of wildfire ecological effects on the ecosystems [
10]. This attribute represents the ecological changes undergone by burned ecosystems with respect to their pre-fire condition [
11] and is usually measured in an operational manner as the above- and/or belowground biomass consumption [
12]. Field-based assessments commonly involve the use of integrative indices, such as the Composite Burn Index (CBI) [
11], the Geometrically structured CBI (GeoCBI) [
13], or their modifications [
14,
15]. The (Geo)CBI is based on the visual assessment of several individual attributes, rather than a single indicator, as proxies for the magnitude of ecological change in several height strata. These attributes are rated in a semiquantitative scale to obtain a (Geo)CBI score per strata of between zero (no change) and three (completely burned), where they are then linearly integrated for an overall plot-level fire severity score.
Compared to the limited representativeness and spatial thoroughness inherent in the exclusive use of field data in fire severity assessments, remote sensing-based techniques provide a cost-effective and appropriate option to retrieve wall-to-wall fire severity estimates, particularly in large, burned landscapes encompassing numerous species assemblages [
16,
17]. Conventionally, broadband optical sensors onboard satellite platforms, such as those of Landsat or Sentinel-2 missions, have been extensively used to retrieve fire severity from moderate spatial resolution imagery (10-20-30 m) at local, regional, or global scales using a wide variety of remote sensing techniques. Previous research has reported that the use of physically based remote sensing techniques, such as radiative transfer models (RTMs) and linear spectral unmixing analysis, rather than empirical models calibrated from spectral indices can be a sounder approach to retrieve generalizable fire severity estimates. For instance, Chuvieco et al. [
18] implemented a coupled leaf (PROSPECT) [
19] and canopy (two-layer Kuusk model) [
20] RTM run in forward mode to simulate the spectral behavior of burned canopies. These RTM simulations were inverted by De Santis and Chuvieco [
16] to retrieve fire severity at the plot-level CBI from Landsat multispectral data. Recently, Fernández-Guisuraga et al. [
15] proposed a fractional vegetation cover (FCOVER) change-detection framework as a biophysical fire severity indicator retrieved from Sentinel-2 multispectral data by the inversion of the PROSAIL RTM [
21]. These studies have shown that the transferability of fire severity retrievals using RTMs is markedly high between different types of plant communities, and that they are sensitive to moderate and high severities. Previous research has also suggested that linear spectral unmixing analysis, particularly advanced methods such as multiple endmember spectral mixture analysis (MESMA) [
22] or weighted MESMA [
23], are sound methods with a physical basis to retrieve sub-pixel image fractions representative of post-fire ground constituents, e.g., char or bare soil, from broadband or narrowband satellite data at moderate spatial resolution [
24,
25,
26]. Therefore, physical-based approaches in conjunction with moderate-resolution satellite data have proven effective for assessing fire severity at large spatial extents. However, wildfire ecological effects on forest ecosystems, particularly in non-stand-replacing events, typically produce a finer patchy mosaic of vegetation burned legacies, bare soil, and recovery responses [
17]. Furthermore, the transition of surface to intermittent crown fire behavior can produce a mosaic of fire effects that are translated into different fire severities [
27] even at individual tree level. As such, satellite data at moderate spatial resolution may not capture the full range of wildfire ecological effects on the ecosystems [
17,
28].
The increased technology deployment of Unmanned Aerial Vehicles (UAVs) in the last decade has overcome the logistical issues and high acquisition costs of the high spatial resolution remote sensing data [
29] required to assess fine-scale ecological processes on demand [
7,
30]. Apart from the production of orthorectified multispectral and hyperspectral wall-to-wall products at a (sub)centimetric resolution with a high geospatial accuracy and reasonable spatial coverage [
31,
32], UAV imagery allows for a cost-effective alternative to reconstruct 3D surfaces and generate dense point clouds using structure-from-motion (SfM) and multi-view stereo (MVS) pipelines (hereafter SfM-MVS) [
33]. Indeed, SfM-MVS point clouds have been extensively used worldwide to retrieve forest structural traits such as canopy height, cover and volume [
29,
34,
35], tree and stem density [
36], or aboveground biomass [
37], among others. Although SfM-MVS point clouds can eventually provide similar accuracies to UAV-LiDAR data in the retrieval of forest canopy traits [
34,
38,
39], unique challenges such as uniform canopy textures and the inability of SfM-MVS to penetrate to the forest floor under dense canopies, unlike UAV-LiDAR, may prevent an optimal scene reconstruction and the emergence of propagating errors in the estimation of structural traits [
40]. However, the partial or total consumption of canopy foliage and small branches and uneven canopy textures in burned landscapes (i.e., alternance of green, scorched and torched leaves even within the same individual canopy) may alleviate these concerns [
41]. In fact, previous research has leveraged UAV RGB, multispectral, and SfM-MVS (i.e., structural) data to estimate fire severity in a wide variety of ecosystem types [
41,
42,
43,
44]. These studies are based on the use of reference data acquired from the UAV imagery itself or other ground-truth sources. While successful at local scales, these methods do not provide ecologically meaningful information linked to fire-induced changes in the field [
15,
45,
46]. Indeed, spectral indices and a battery of height structural metrics are not intrinsic vegetation biophysical properties that can be generalizable and provide a direct link with field-based descriptors of fire severity. This is particularly relevant if fire severity retrievals are tailored to capture fire-induced changes in several ecosystem compartments as measured in integrative field indices such as the (Geo)CBI [
13].
In this paper, we aim to develop and validate a fire severity retrieval methodology based on UAV-derived structural and spectral metrics with enough physical basis and ecological sense to procure accurate and generalizable estimates. Specifically, we propose a post-fire, mono-temporal framework integrating UAV dense 3D-point clouds, multispectral data, and land surface temperature (LST) at high spatial resolution (
Figure 1). First, the ecosystem compartments as described by the CBI strata configuration are identified through individual tree segmentation and geographic object-based image analysis (GEOBIA) techniques. Second, the wildfire ecological effects are individually estimated for each compartment by means of the following: (i) the structural complexity of the canopy vegetation legacies as a proxy for biomass consumption [
47], and (ii) vegetation biophysical variables retrieved from multispectral data by the inversion of the PROSAIL RTM, along with a direct physical link with the vegetation legacies remaining after canopy scorching and torching [
15,
18]. These structural and spectral descriptors are assumed to procure complementary information on spectral and structural traits of burned vegetation canopies and are important individual attributes in the CBI scheme [
48]. The retrieval of fire-induced impacts on vegetation biophysical variables through physically based remote sensing techniques provide a direct and mechanistic link with field-based descriptors of fire severity [
13]. However, these methods have not yet been implemented to retrieve fire severity by leveraging the potential of high spatial resolution UAV multispectral and SfM-MVS data to discriminate fire effects according to CBI strata. We hypothesize that this method would reflect the ecological changes in the CBI scheme more accurately than the conventional computation of a wide battery of plot-level structural and spectral metrics to feed the statistical or machine learning models [
49,
50]. The fire severity retrieval methodology proposed in this paper was implemented at independent training and validation sites, corresponding to two wildfires in the westernmost part of the Mediterranean Basin.
3. Results
Fire severity in the Foyedo wildfire was strongly decoupled among the vertical strata in the CBI hierarchical scheme (F = 20.187;
p-value < 0.001) (
Figure 4). The lowest fire severity was registered in the substrate stratum (mean CBI score of 0.70). Fire severity increased progressively with strata height (
Figure 4). The highest fire severity was reached in the vegetation stratum of higher than 20 m in height, although it did not significantly differ from that of the 5–20 m stratum. The vegetation stratum higher than 20 m was discarded from further analyses as it was present only in five CBI plots (ratio #observations/#predictors in RF models close to 1:1). Also, this stratum was absent in the CBI plots of the Lavadoira test site. The mean plot-level CBI score was equal to 1.22 (1.20 after discarding the highest vegetation strata).
The distribution of the UAV 3D-point clouds and the individual segmented trees closely resemble the observed ecological fire effects in the field in representative CBI plots dominated by
Pinus pinaster communities with the same pre-fire structure, i.e., mature trees and a well-developed understory (
Figure 5). Under low-intensity surface fire, the 3D-point cloud clearly depicts completely unaffected tree canopies and intermittent foliage consumption in the understory (
Figure 5). At moderate fire severity, the 3D-point cloud captures an almost total biomass consumption in the understory and partially scorched canopies with incomplete foliage loss, particularly in the upper canopy (
Figure 5). In high-severity plots, biomass consumption is almost total in all the strata and only the tree trunks and thick branches remain (
Figure 5). Importantly, there was no CBI plot in which the Dalponte and Coomes [
67] ITS algorithm identified and segmented individual trees associated to a CBI stratum that was not present in the plot.
The UAV canopy density and vegetation biophysical variables with ecological sense, retrieved from the 3D-point clouds and PROSAIL-D inversion, respectively, varied markedly across the fire severity categories at the plot level (
Figure 6). Canopy density, FCOVER, and C
w in the understory and overstory strata gradually decreased as the severity of the fire increased, while C
brown showed the opposite behavior, together with the char/ash cover in the substrate stratum (
Figure 6). Differences for all structural and biophysical descriptors between the fire severity categories were gradually stronger with the decreasing strata height. The char/ash cover, directly retrieved from GEOBIA, was the variable with the strongest response to fire severity. The biophysical variables showed a stronger variability between the fire severity categories than the canopy density, particularly C
brown and C
w (
Figure 6).
The canopy density metric and vegetation biophysical variables featured a consistent and strong linear relationship with the CBI scores aggregated by strata (
Figure 7). The char/ash cover was highly correlated (R
2 = 0.799) with the substrate CBI measured in the field. The canopy density metric featured a moderate correlation (R
2 = 0.645 ± 0.112) with understory and canopy CBI scores by strata. Among the vegetation biophysical variables, the strongest relationships were evidenced for FCOVER (R
2 = 0.761 ± 0.066) and C
w (R
2 = 0.759 ± 0.098). The relationship between C
brown and the CBI scores in the 5–20 m vegetation strata was markedly higher (R
2 = 0.716) than in the understory (R
2 = 0.460).
The retrieval of CBI scores aggregated by strata from the structural and spectral UAV-derived metrics with ecological sense using RF regression models in the Foyedo wildfire featured very high overall fit (R
2 = 0.860 ± 0.036) and low predictive error (RMSE = 0.314 ± 0.018) (
Figure 8). The CBI score retrievals were closely tailored to the 1:1 line, with no apparent under- or over-estimation effects for each stratum throughout the whole CBI range (MBE < |0.006|). From these results, it can be assumed that UAV-derived canopy density and vegetation biophysical variables at the CBI strata-level provide complementary information in the retrieval of fire ecological effects because of the higher performance of the RF models compared to the univariate relationships. The RF variable importance metrics (%IncMSE) followed the same pattern as the performance of the univariate linear relationships depicted in the
Figure 7.
The predicted CBI score aggregated at the plot level with observed data in the Foyedo wildfire using ecologically related metrics previously retrieved at the strata level (internal model validation) featured a very high fit (R
2 = 0.910) and a low predictive error (RMSE = 0.219) with respect to the field-measured values (
Figure 9A).
The RF model extrapolation from the Foyedo to the Lavadoira wildfire with unobserved data (external model validation) provided encouraging results. The R
2 decreased and the RMSE increased only by 0.1 and 0.04 units, respectively, and the under- or over-estimation effects were still negligible (
Figure 9B).
The RF extrapolation ability from the Foyedo to the Lavadoira wildfire, by directly retrieving the fire effects at the plot-level CBI using ecologically related UAV metrics (benchmark method #1), was penalized by a significant underestimation of the CBI field-measured values (
Figure 9C), particularly at high severity, resulting in MBE = −0.109. The overall fit and predictive error were lower (R
2 = 0.751 and RMSE = 0.314) than in the extrapolation approach using the ecologically related metrics previously retrieved at the strata level (R
2 = 0.807 and RMSE = 0.260).
The worst-case scenario corresponded to the prediction of the CBI scores aggregated at the plot level in the Lavadoira wildfire by using a battery of UAV structural and spectral metrics (benchmark method #2). The retrieval resulted in a relatively weak performance (R
2 = 0.677 and RMSE = 0.349) and similar underestimation error (
Figure 9D) compared to benchmark method #1.
4. Discussion
The potential of high spatial resolution UAV multispectral and SfM-MVS data to provide ecologically meaningful fire severity estimates was examined in this paper for the first time. Our results shed light on the importance of using physically based remote sensing techniques with ecological consistency to process UAV data and procure accurate estimates of integrative fire severity measurements in the field (R2 = 0.910 and R2 = 0.807 for internal and external model validation, respectively), such as those provided by the CBI protocol.
Although the estimation of CBI data with structural and/or spectral UAV metrics has not been conducted to date, previous studies addressing fire severity classifications with remote sensing techniques applied to UAV data, ranging from post-fire RGB and multispectral indices to a battery of structural metrics, have reported overall accuracies of 38–91% without independent model validation [
41,
42,
43,
44]. However, the target classes and reference data in the classification methods of these studies show a high disparity and therefore are not directly comparable to our approach. In addition, they evaluate ecological change exclusively at the top-of-canopy level in broad classes because the reference data are usually extracted from high spatial resolution RGB orthomosaics, neglecting ecological change in the understory and substrate. Nevertheless, when the CBI thresholds proposed by Miller and Thode [
99] were applied in the plot-level CBI predictions of the internal and external model validation in our study for comparison, the procured overall accuracy was 95% and 89%, respectively.
For the sake of comparison, airborne LiDAR data have also not been widely implemented to estimate an integrative measure of fire severity, but rather to predict individual ecological impacts such as canopy consumption [
103], changes in canopy and understory structure [
96,
104], or LAI and basal area [
66], among others. So far, only a few authors have leveraged airborne LiDAR data to predict the fire impact on the ecosystems by means of the CBI. For instance, Montealegre et al. [
49] used a wide battery of post-fire LiDAR metrics at the plot-level to estimate the variability of categorized and continuous CBI data in the conifer and mixed forests of Northeast Spain. The authors reported an overall accuracy of 85.5% in the classification scheme and an R
2 = 0.63 in the regression model, which do not exceed the performance procured by the Landsat-derived dNBR index in their study and are in line with our benchmark #2 performance (battery of UAV structural and spectral metrics at the plot level). Recently, Gale et al. [
105] implemented the profile area change (PAC) method developed by Hu et al. [
66] to predict CBI variability in Southeastern Australia through the estimation of fire-induced differences in the mean height percentile profile area between pre- and post-fire LiDAR acquisitions. The authors attributed the relatively low PAC performance (R
2 = 0.64 vs. 0.91 in our study under the same internal validation conditions) to the metric inability to represent nonstructural ecological changes, which can be especially problematic in areas burned at low to moderate fire severity [
106]. Indeed, we found that, apart from the consideration of all ecosystem strata to be consistent with the integrative nature of the CBI, it is important to account for the following: (i) the fire-induced ecological impacts on the substrate, (ii) the green vegetation fraction that remains after canopy scorching and torching, and (iii) the spectral signal of damaged and dry leaves. In this context, post-fire biophysical variables retrieved from the PROSAIL-D RTM inversion typically showed a higher contribution in each stratum to explain CBI variability than the density of the vegetation legacies, while both groups provided complementary information about the spectral and structural canopy traits [
7]. Previous research has also determined that biophysically meaningful variables depicting changes in soil and foliage color, as well as in vegetation cover, are important indicators of fire damage, particularly at lower severities [
15,
18,
75], where less-damaged trees may have preserved the leaves at the moment of the UAV survey and thus show no major structural changes [
107].
Our results are in line with the findings of García et al. [
106], who proposed the waveform area relative change (WARC) as a new metric that allows for the comprehensive assessments of fire severity, not only accounting for all ecosystem strata but also for target reflectance and thus substrate affectation and foliage consumption. The WARC metric showed high consistency and strong correlations with simulated full waveform LiDAR data through the FLIGHT RTM [
108] execution in forward mode for a range of fire severity effects in GeoCBI terms, representative of conifer and mixed forests in Central Spain. The GeoCBI retrieval performance by the WARC metric in Sierra Nevada, California (external validation), was markedly high (R
2 = 0.78) and comparable to that of our study in the same conditions with unseen data (R
2 = 0.807). As a bi-temporal change-detection framework, WARC requires the acquisition of pre- and post-fire LiDAR datasets, which is a constraint for any airborne sensor because of the limited data availability [
109], contrary to the proposed UAV mono-temporal framework. In this context, our experimental design, clearly compartmentalized by CBI strata in an object-based nature [
104], can be leveraged as a cost-effective method to derive ground-truth fire severity data and augment UAV-based samples with wall-to-wall coarser satellite imagery [
110,
111], given the accuracies reported here. Indeed, the predictive error in the external model validation scheme (RMSE = 0.260 in CBI units) is around 13%, considering the CBI range in the test site, well below the 25% threshold for qualifying a remote sensing product as not scalable for operational implementation in post-fire assessments [
16].
First, the high transferability of our CBI retrieval scheme can be associated with the representativeness of the canopy density metric on the post-fire fuel distribution and openness in intermediate parts of the canopy when estimated from dense 3D-point clouds [
109]. Additionally, the division of each CBI stratum into the same number of static height bins may be more representative of fuel consumption throughout the vegetation vertical profile, regardless of the pre-fire stand composition or growth stage [
112], than using predefined height bins for the whole plot, e.g., understory, intermediate, and overstory layers [
113]. Second, the retrieval of the variability inherent to the fire-induced effects on representative vegetation biophysical traits [
114] meets the definition of a generalizable biophysical indicator of fire damage from an ecological standpoint [
98,
99]. Therefore, the relationships procured for each CBI strata are expected to be more transferable between the distinct species assemblages [
16] while accounting for the complex mixture of vegetation legacies and background soil/char signal in post-fire landscapes [
24] through reference endmembers in the PROSAIL-D retrieval scheme. The use of reference endmembers has also been shown in previous research [
15] to improve the retrieval performance from Sentinel-2 data in a wide variety of Mediterranean plant communities at the plot-level CBI. Although not validated with field data, the behavior of the retrieved biophysical variables has been shown to strongly depict the expected ecological changes with fire severity variability in the CBI strata [
18], assuming a proper PROSAIL-D parametrization. Our goal here was not to validate the FCOVER, C
brown, and C
w retrievals (already conducted extensively in burned areas) but to evaluate their performance to be implemented operationally in fire severity assessments with high spatial resolution UAV data, as in many previous remote sensing-based research [
115,
116,
117]. There is also room for improvement in this study by leveraging geometric RTMs that provide a more realistic description for complex vegetation canopies in burned areas with non-homogeneous strata [
75], in comparison to widely implemented RTMs with turbid-medium assumptions such as PROSAIL. However, the more complex structure of geometric RTMs such as GeoSAIL [
118] or FLIGHT [
108] should be regarded as unrealistic parameterization without field data, typically unavailable in the short term after a fire [
46], may produce misleading simulations and introduce large uncertainty in the retrieval of vegetation biophysical variables to be used operationally [
119]. Also, lightweight narrowband sensors are increasingly deployed as UAV payload to retrieve biochemical and biophysical plant traits from high spatial resolution imagery [
120]. Progressing along this line, narrowband UAV data could be leveraged to reduce the RTM inversion uncertainty associated with important absorption feature regions not commonly sampled on multispectral sensors [
25,
121] and thus better characterize fire effects on vegetation biophysical traits. Notwithstanding, RTMs have not yet been applied to spectroscopy data in the fire severity field. Future studies should also consider the inclusion, in the CBI scheme, of other individual fire severity indicators that have been shown to be relevant in estimating post-fire ecological processes [
48], such as changes in soil and ash color in the substrate stratum [
122,
123], or the minimum tip diameter of remaining branches in the lower vegetation strata [
124].
The significant underestimation in the retrieval of fire effects using ecologically related UAV metrics at the plot-level CBI (benchmark #1) entailed a 20% higher error than when the same metrics previously retrieved at the strata-level were used. The underestimation may be associated with the typical mixture of low, moderate, and high fire severity signals within multiple strata framed within an CBI plot [
17], for instance, under intermittent crowning fire behavior, leading to an aggregation effect of mixed fire effects which cannot be properly resolved at the plot scale [
28]. Typically, this would be the result of the higher contribution of low to moderate fire effects’ signals in the understory than the actual one at the plot level, considering the CBI scheme. The use of a battery of UAV structural and spectral metrics (benchmark method #2) showed the worst performance, probably related to the lack of physics and low generalization of spectral indices between distinct species assemblages [
16], together with their limited ability to discriminate fire effects when the background signal dominates surface reflectance [
125]. In addition, height distribution and variability metrics from 3D-point clouds may obscure relevant relationships with CBI field data in conifer forest ecosystems because of the common presence of height-invariant snags, i.e., dead standing trees [
106].
Altogether, our results suggest moving towards more ecologically based fire severity assessments through remote sensing techniques. Further research efforts are needed to assess the transferability and validate the UAV-based methodology proposed in this study in a wider variety of oceanic plant communities, as well as in other climate-type regions, namely Mediterranean ecosystems.