Imaging Burned Areas and Fire Severity in Mediterranean Fragmented Ecosystems Using Sentinel-1 and Sentinel-2: The Case Study of Tortoli–Ogliastra Fire (Sardinia)

The study aims to explore the added value of the joint use of Sentinel-1 (S1) and Sentinel-2 (S2) data for assessing burn severity in heterogeneous, fragmented ecosystems. The importance of this aim lies in the fact that for both S2 and S1 (as for all the synthetic aperture radar (SAR) C-bands), the impact of fire was found to cause ambiguous effects in complex and fragmented ecosystems. For our investigation, the effectiveness of S1 and S2 fire metrics was first statistically analyzed using ISODATA coupled with field surveys conducted for a fire that occurred on 13 July 2019 in Sardinia. Later, to automatically map burn areas and categorize fire severity, S1 and S2 fire metrics were integrated through a multilevel classification performed at a pixel and feature level. Results were successful (accuracy higher than 94%) compared with independent data sets and in situ investigations.

Imaging Burned Areas and Fire Severity in Mediterranean Fragmented Ecosystems Using Sentinel-1 and Sentinel-2: The Case Study of Tortoli-Ogliastra Fire (Sardinia) Rosa Lasaponara , Carmen Fattore, and Giuseppe Modica Abstract-The study aims to explore the added value of the joint use of Sentinel-1 (S1) and Sentinel-2 (S2) data for assessing burn severity in heterogeneous, fragmented ecosystems.The importance of this aim lies in the fact that for both S2 and S1 (as for all the synthetic aperture radar (SAR) C-bands), the impact of fire was found to cause ambiguous effects in complex and fragmented ecosystems.For our investigation, the effectiveness of S1 and S2 fire metrics was first statistically analyzed using ISODATA coupled with field surveys conducted for a fire that occurred on 13 July 2019 in Sardinia.Later, to automatically map burn areas and categorize fire severity, S1 and S2 fire metrics were integrated through a multilevel classification performed at a pixel and feature level.Results were successful (accuracy higher than 94%) compared with independent data sets and in situ investigations.

I. INTRODUCTION
W ILDFIRES are one of the most important causes of environmental degradation, with an increasing impact on a global scale driven and exacerbated by climate change, which induces prolonged drought and increasing temperatures [1], [2], [3], [4].
To contrast wildfire and limit fire incidence, current and future challenges deal with enhancing social and ecological resilience and improving timely and reliable information on the fire occurrence and caused damage.
Earth observation (EO) technologies, particularly the Copernicus program and Sentinel missions [5], [6], [7] can effectively support new fire monitoring applications from a local up to global scale, from pre-fire risk assessment to fire detection, from the mapping of burned areas (as binary identification to discriminate fire affected from fire unaffected pixels) to the estimation of burn/fire severity (categorization of the degree of consumption of vegetation and surface soil organic matter [8]) considered one of the most critical information needed immediately after the fire to set damage-mitigation strategies.Fire severity metrics are directly proportional to the degree of burn and, in turn, to the level of damage.
In contrast, the mapping of burned areas is generally referred to as their fire perimeter.Severity metrics (including satellitederived ones) are typically based on empirical fixed thresholds directly linked to the fire effects observed in the field within the fire perimeter.Over the years, satellite optical data were widely used to map burned areas and fire severity, providing good accuracy (except in heterogeneous ecosystems) [see, for example [9], [10] and reference therein quoted].In contrast, synthetic aperture radar (SAR) sensors have been less used due to the less availability of free data (before the launch of S1) and the major complexity of data processing, even if SAR systems provide helpful data for discriminating changes and disturbances in vegetation.Even if, recently, SAR-based fire analyses are increasing [11], [12], [13], [14], [15], [16], [17], [18], [19], [20], [21], [22], [23], [24], [25], [26], [27].SAR systems offer incoherent (amplitude) and coherent data to perform change detection (CD) compared to optical data.In coherent CD (CCD), the radar images are compared in amplitude and phase, whereas InCoherent CD (ICCD) is based only on amplitude analyses.ICCD was recently carried out by Mastro et al. [12] for obtaining burned area mapping as a binary discrimination of burned from unburnt areas.Obviously, this binary approach [12] is unsuitable for fire severity estimation that requires mandatory categorization of the degrees of fire effect generally approached using incoherent CD.This is mainly due to the fact that scattering variations occur after a fire event due to the loss in vegetation cover, and these variations are directly proportional to the degree of burn.ICCDs for fire investigations are mainly based on metrics based on fixed thresholds linked to the fire effects observed in the field, as successfully made [14] for the SAR L-band to quantify fire damage levels in different continents, environments, ecosystems, and vegetation types.Whereas, up to now, for the SAR C-band, the impact of fire on the backscattering coefficient was actually found to cause ambiguous effects [10] (as shown in Table I) in different ecosystems because it is influenced by forest structure, which is the main factor influencing the backscatter.
Therefore, Tanase et al. [14] suggested using additional information on forest structures to improve the SAR-based burn estimation.Moreover, SAR backscattering tends to exhibit temporal decorrelation, making identifying fire features complex.On the other hand, in fragmented ecosystems, the © 2023 The Authors.This work is licensed under a Creative Commons Attribution 4.0 License.
For more information, see https://creativecommons.org/licenses/by/4.0/Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.

TABLE I ECOSYSTEMS/GEOGRAPHICAL AREAS WHERE AMBIGUOUS EFFECTS HAVE BEEN OBSERVED IN THE C-BAND BACKSCATTERING COEFFICIENT FOR FIRE-AFFECTED AREAS ADAPTED
FROM TANASE et al. [14] fixed thresholds commonly used for S2 metrics to categorize fire severity seem unsuitable.
To contribute to this issue, the effectiveness of several S1 and S2 fire severity metrics was assessed in a fragmented ecosystem of Sardinia (Italy).Finally, S1 and S2 fire metrics were integrated through a multilevel classification to map burned areas and fire severity automatically.

II. STUDY AREA AND DATA SETS
The fire event that occurred in Sardinia on 13 July 2019 near Tortoli, in Ogliastra, was selected as a test case.The fire affected around 800 hectares of Mediterranean shrubs (mainly brush, bush, and pasture) classified in the Copernicus EFFIS system [1] as chaparral fuel, i.e., vegetation primarily consisting of tangled shrubs and thorny bushes.For the purpose of our investigation, for both S1 and S2, the most common indices were adopted; in particular, for S2, the normalized burn ratio (NBR) index [12] was computed using the following equation: NBR is particularly sensitive to the changes in the amount of live green vegetation, moisture content, and some soil conditions that may occur after a fire, such as ash, the reduced transpiration of vegetation, and the increased surface temperature due to the loss of vegetation cover.All these effects increase the reflectance in mid-infrared [22] and reduce surface reflectance in near-infrared and are better reflected considering the pre/post-burn NBR index difference (dNBR) in the following equation: For S1, we adopted the radar burn ratio (RBR) computed as in (3), the radar burn difference (RBD) as in (4), and the Radar Forest Degradation Index (RFDI) computed as in ( 5) (5) Both RBR and RBD are calculated for a specific polarization [18], averaging several pre-and post-fire scenes acquired in dry conditions as in [15] to reduce SAR noise and preserve spatial detail.The selection of the number of SAR images to average is based on the meteorological conditions to avoid misclassification due to precipitations and soil moisture.
The RFDI was analyzed [see (6)] as the difference between pre and post-fire maps dRFDI = RFDI pre-fire − RFDI post-fire .
III. METHOD

A. Analyses of S1 and S2 Fire Severity Metrics
The analyses of the effectiveness of fire severity metrics for both S1 and S2 were carried out using ISODATA unsupervised classification (Fig. 1) in order to overcome the issue linked to the ambiguous effects observed in S2 [9], [10] and S1 [16] fire metric in a heterogeneous, fragmented ecosystem.The use of a cluster-based classification (as ISODATA or other similar approaches) maintains the clear and immediate physical meaning of the investigated fire metrics.It makes the interpretation easier without using fixed thresholds that the local and geographical conditions can obviously limit.ISODATA was also adopted to limit the human intervention in setting up the algorithm parameters as much as possible.The importance of applying unsupervised classification is that: 1) it is an automatic process, which requires only a minimal amount of initial input compared with supervised data processing; 2) classes do not have to be defined a priori; and 3) unknown feature classes may be discovered.Several unsupervised classification algorithms are commonly used in remote sensing, including: 1) K-means clustering and 2) ISODATA, which are pretty similar.In both, the user has only to indicate: 1) the number of predefined classes (clusters) and 2) the number of iterations to be carried out.
Considering that the EFFIS system and Copernicus EMS Rapid Mapping utilize five and three classes of fire severity, respectively, wherein the number of the predefined classes (clusters) was set at 6 and 4 to make comparable the S1 and S2 outputs with the Copernicus products.This enabled us to have Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.discriminating fire severity levels?An effort to answer this question is provided in this letter.

B. Fire Severity Categorization
The approach devised to map burned areas and assess fire severity is based on results previously obtained by the same author group using LISA coupled with ISODATA to process S2 [9] in a fragmented ecosystem or S1 [16] in a homogeneous area.In both of these studies, using LISA enabled the enhancement of fire-induced changes, made more evident by enhancing texture variations and/or by the presence of edges and clusters.In this study, ISODATA has been selected to "manipulate" fire metrics, "preserving" their physical meaning linked to the geophysical parameters under investigation.
To exploit the complementarity of S1 and S2 sensors, a data fusion approach was adopted at pixel and feature levels.Both S1 and S2 fire metrics were stacked together to be reclassified at pixel and feature levels.
The classification at the pixel level is obtained from the joint processing of the together stacked S1 and S2 fire severity metrics (all resampled at 10 m of pixel resolution).As a whole, three different classifications were run obtained based on the joint use of: 1) S2 dNBR and S1 RBD [( 2) and ( 3)]; 2) S2 dNBR and S1 RBR [( 2) and ( 4)]; and 3) S2 dNBR and S1 difference of the RFDI obtained from the difference of pre and post-fire maps [( 2) and ( 6)].
Feature-level data are generally adopted for decision analysis, and fusion results are expected to provide new features more suitable for decision analysis.The feature-level classification is obtained from the feature-level data fusion approach, based on the joint use of fire features obtained from the classification of each metric obtained from each sensor.So, as in the case of the pixel-level classification, the final products are three.
IV. RESULTS Outputs from all the classifications were overlapped on Google Earth and compared with the Very High-Resolution optical satellite Pléiades S1A/B imagery acquired on 18 July 2019, immediately after the fire occurrence and used by Copernicus EMS [28] for mapping burned areas and fire severity.The Copernicus EMS provides support (if institutional authorities request) immediately following a disaster.In the current case, the service was requested and activated on 17 July.It provided the mapping of burned areas and different levels of fire damage as obtained using Pléiades S1A/B (see Fig. 4).EMS estimated that the area affected by the fire was around 665.5 ha and identified the diverse levels of fire damage categorized as destroyed (in red), damaged (in orange), and possibly damaged (in yellow).
Results from our analysis pointed out that as follows.
1) The use of LISA enabled a better identification and delineation of the fire-affected pixels, not only for S2 (thus confirming the already published output [12]) but also for S1-based fire metrics.2) For the mapping of burned areas (fire perimeter), the best performance was obtained from the classification of Moran applied to RBR in the VH polarization.Fig. 4 shows (on the top) the zoom of the results from the joint classification of Moran of S2 dNBR and S1 RBD VH polarization (herein indicated as feature-level classification).3) For the fire severity categorization, the best performance was obtained from the classification of Moran applied to RBR in the VH polarization performed at the pixel level.As a whole, the satellite sentinel-based results (at 10 m of spatial resolution) were successfully compared with the independent fire map provided by the Copernicus EMS (from Pléiades S1A/B data at 2 m of spatial resolution) (Table III) [27].
V. DISCUSSION AND CONCLUSION The study illustrates the potential of the synergic use of S1 and S2 for speditive and effective fire severity mapping, exploiting the complementarity of the two sensors.
The satellite sentinel-based results (at 10 m of spatial resolution) were successfully compared with the independent Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.
fire map provided by the Copernicus EMS (from Pléiades S1A/B data at 2 m of spatial resolution) [28].The comparison clearly showed the excellent match of the burned areas and fire damage mapped from the Copernicus EMS Rapid Mapping map based on Pléiades S1A/B and the classification obtained from S1 and S2.These results are highly satisfactory considering that the whole area is sparsely vegetated and characterized by a dominant presence of Mediterranean shrubs (mainly brush, bush, and pasture) classified as chaparral fuel [1].As a whole, results from our analyses clearly pointed out that both of the two proposed approaches (i.e., the joint ISODATA classification of S1 RBD of VH polarization and S2 at pixel and feature level) can be used to detect burned areas and estimate fire severity providing a speditive and accurate characterization with the diverse degrees of fire damage.
Future efforts will be addressed to refine the classification further using the enhancement of fire features (herein obtained by LISA) and artificial intelligence (AI) approaches, which require a significantly large data set to obtain successful results.

Fig. 1 .
Fig. 1.General flowchart of the proposed data processing.

Fig. 2 .
Fig. 2. Images show satellite-based fire severity metrics: (on the left) S2 dNBR and (on the right) S1 RBD.Brighter pixels are related to the fire event that occurred in Sardinia on 13 July 2019 near Tortoli, in Ogliastra.

Fig. 3 .
Fig. 3. Images show the ISODATA results from (on the top) S2 dNBR and (on the bottom) S1 RBD.

Fig. 4 .
Fig. 4. On the top: joint classification of Moran of S1 RBR and S2 dNBR (herein denoted as feature level classification) 10 m spatial resolution.On the bottom right, ISODATA classification from the joint use of S1 RBR and S2 dNBR at the pixel (10 m spatial resolution) level at 1:100 000 spatial scale.On the bottom left, the Burned area (pixel at 2 m) from the EMS Rapid Mapping [27].

TABLE III RESULTS
OBTAINED FROM THE PIXEL-LEVEL CLASSIFICATION OF S1 RBR AND S2 DNBR (BOTTOM RIGHT IN FIG. 4) WERE COMPARED WITH THE DAMAGE ASSESSMENT PROVIDED BY THE COPERNICUS EMS RAPID MAPPING [25] FOR THE TWO CLASSES DENOTED AS DESTROYED AND DAMAGED.IN CONTRAST, THE CLASS DENOTED AS PROBABLY DAMAGED WAS NOT CONSIDERED BECAUSE OF ITS UNCLEAR PHYSICAL MEANING