FIRE SEVERITY ASSESSMENT OF AN ALPINE FOREST FIRE WITH SENTINEL-2 IMAGERY

Fire is a common phenomenon in many forests and is considered an important ecological tool. Fire severity mapping presents an effective way to assess post-fire management intervention and is helpful in environmental and climate change research. The objective of this study was to determine the severity of a forest fire event that occurred from 24 to 27th October 2019 at Taibon Agordino using Sentinel-2A satellite images and creating a severity map suitable as a decision-making tool for post-fire management intervention. The Sentinel-2A satellite data was classified into the following five classes: Unburned, Low Severity, Moderate Severity, High Severity, and Shadow with the non-parametric Random Forest (RF) classifier, and the resulting classified image was validated using validation sites. The RF classifier was applied first to the ten original band reflectance of Sentinel-2. In a second step, additional variables were added to the classification, namely the digital elevation model (DEM), the slope, and five vegetation indices (i.e., Differenced Normalized Burn Ratio (dNBR), Relative Differenced Normalized Burn Ratio (RdNBR), Differenced Bare Soil Index (dBSI), Global Environmental Monitoring Index (GEMI) and Burn Area Index (BAI)) The inclusion of vegetation indices and DEM-related variables increased the classification accuracy from 99.26% to 99.61% and the overall accuracy from 70.51% to 83.33%. In the classification with the ten original band reflectance, the variable of importance plot ranked the Red-Edge-3, Red, and SWIR 1 band reflectance as the top three most important input features, while for the classification with 17 variables, RdNBR, DEM and dNBR were the top three most important input features.


INTRODUCTION
Forest fires are natural disturbances affecting forest ecosystems in many parts of the world. Fire is an essential component of natural dynamics in fire-dependent ecosystems, while in other ecosystems, it can lead to forest degradation and hinder ecosystem services provision. Fire severity mapping presents an effective way to assess post-fire conditions, providing helpful information for identifying priority areas for management interventions (Morresi et al. 2022). Furthermore, it could provide insights for understanding trajectories of vegetation recovery, biological legacies distribution, and eventually is helpful in environmental and climate change research. Most fire severity mapping studies using remote sensing were performed using MODIS imagery (see the review of Leblon et al., 2016). While MODIS imagery is available daily, the imagery has a too coarse resolution for a detailed fire severity map. Landsat series have also been used for fire severity mapping (e.g., Matricardi et al. 2010., Escuin et al. 2008., Wimberly, Reilly, 2007. The European Space Agency has recently launched the Copernicus Sentinel-2 satellite carrying the Multispectral Instrument (MSI) capable of acquiring images in 13 bands at higher spatial resolutions. Also, given the availability of two satellites (2A and 2B), the revisiting time of the Sentinel-2 satellites allows a suitable temporal resolution (revisit time) of 5 days in the imagery acquisition (Li, Roy, 2017). Some studies tested Sentinel-2 imagery for burned area mapping (Filipponi 2018, DeSimone et al. 2020, Han et al. 2021, Fassnacht et al. 2021, but very few on fire severity mapping in southern Australia (Gibson et al. 2020), in Greece (Mallinis et al. 2018) or Spain (Quintino et al. 2018, and recently in the Western Italian Alps (Morresi et al. 2022).
The objective of this study was to map the severity of a forest fire event that occurred in the Dolomites Mountains (Eastern Italian Alps) from 24 th to 27th October 2018 at Taibon Agordino (Italy) using Sentinel 2A images and creating a severity map suitable as a decision-making tool for post-fire management intervention. Such mapping is quite challenging, given the rough topography of the area.

Study area
The study area is in the northeastern part of Taibon Agordino municipality (Lat. 43.92N; Long. 12.13E) (Figure 1). The study area has a typical dolomitic alpine vegetation mainly comprising of Norway spruce (Picea abies (L.) Karst.), pines (Pinus sylvestris L, P. mugo Turra), and European larch (Larix decidua Miller). Norway spruce is the dominant species. Due to the steepness of the slopes, the forest stands have preeminent protective rather than productive functions. The bedrock is predominantly dolomite, calcium, magnesium, and carbonate compound. The snow in the area begins to accumulate in December and lasts until the end of March in the valleys and up to August at the highest elevation. The area's elevation ranges between 698and 2294m above sea level.

Data
The study uses two cloud-free (<10%) Sentinel 2 Level 1C images. The first image was acquired just before the fire and the second after the fire (Table 1). Both images were orthorectified top-of-atmosphere reflectance images having the UTM 32N map projection and the WGS84 datum. They had a swath width from nadir of 290 km and were acquired using the descending orbit 22. The study also uses a shapefile that delineates the fire perimeter provided by the Veneto regional administration and a Digital Terrain Model (DTM) that was downloaded from the ARPA-Veneto website (http://geomap.arpa.veneto.it/layers/ geonode%3ADTM_5M_GBO). The data has a 5m resolution distributed under the Creative Commons Attribution 3.0. We also used a set of geolocalized field pictures collected in situ that correspond to the different fire severity classes. Samples of ground pictures taken in the Unburned, Low Severity, Moderate Severity, High Severity areas are presented in Figure 2.

Methodology
The methodology used in the study is described in Figure 3. First, the study area's pre-and post-fire Sentinel-2 imagery were subjected to atmospheric correction using Sen2cor version 2.9. Sen2cor is a Level-2A processor used to correct single-data Sentinel-2 Level-1C Top-Of-Atmosphere (TOA) products from the effects of the atmosphere in order to deliver a Level-2A Bottom-Of-Atmosphere (BoA) reflectance product (Main-Knorn et al. 2017). The atmospheric correction is done using a set of look-up tables generated via libRadtran (Emde et al., 2016). The resulting BoA reflectance images were produced at either 10m or 20m as a function of the band. The images having a 20m resolution were resampled to 10m using the RESAMP algorithm in PCI Geomatica.  [24] postBSI = [10] [24] The post-fire imagery was classified with the non-parametric Random Forest (RF) classifier (Breiman 2001) into the following five classes: Unburned, Low Severity, Moderate Severity, High Severity, and Shadow. Since Random Forest is a supervised classifier, it requires delineating training areas, which was done for each fire severity class. In total, 59 training polygons were delineated for the five classes (Unburned, Low, Moderate, High, and Shadow) with the aid of an aerial photograph acquired in July 2019 over the study area. The training polygons were also used to compute the J-M distances between class pairs with the 10 Sentinel-2 band reflectance. The classifier was applied first to the ten original Sentinel-2 band reflectance. In a second step, additional variables were added to the classification, namely the digital elevation model (DEM), the slope, and the five vegetation indices. The classification was assessed by comparing the resulting classified image with an independent set of validation points extracted from photointerpretation of an aerial photograph taken after the fire and from GPS ground pictures. In the comparison, we used a total of 780 randomly selected validation points.

Class spectral separability
The average J-M distance with all the Sentinel-2 bands is 1.965, indicating an excellent spectral separability between the classes ( Table 2). The lowest separability was between the Unburned and Low Severity classes, with a value of 1.615. Although it falls below 1.9, the magnitude is still good enough considering the degree of similarity between the two classes, making it difficult to differentiate on the Sentinel-2 imagery. Indeed, both classes have similar reflectance, mainly in the red, red-edge, and near-infrared bands. The highest J-M distance (1.9999) was found between two class pairs: High Severity and Unburned classes, Shadow and High Severity classes. These classes share notable differences both physically and spectrally. Areas classified in the High Severity class are distinguished by extreme burns and total loss of foliage against unburned areas with healthy green foliage ( Figure 2). Also, the Shadow class, which is a result of topography and nadir angle of the satellite when the image was taken, shows markedly different spectral characteristics from the High Severity Class. The difference in reflectance was most notable in almost all the bands, especially the green, red and red-edge bands. The reflectance of the High severity class is markedly higher across the bands but is distinctively prominent in the red, green, and blue (RGB) bands. Conversely, the reflectance of the moderate severity class is lowest in the RGB bands but more prominent in the red-edge and shortwave infrared bands. The Low severity and Unburned classes have low reflectance in the RGB bands but high reflectance in the red-edge, near-infrared, and SWIR bands.

Classification
The ten original band reflectance of the postfire Sentinel-2A image were imputed into the RF algorithm. The resulting confusion matrix of  (Table 3). The resulting classified image is presented in Figure 4. In the classification with the ten original band reflectance, the variable of importance plot ranked the Red Edge 3, Red, and SWIR 1 bands as the top three most important input features ( Figure 5). The variable importance plot produced for the classification with all the variables ranked RdNBR, DEM, and dNBR as the top 3 most important variables ( Figure 6).

Validation
When only the ten bands were used in the classification, we achieved an overall accuracy of 70.5% and a Kappa coefficient of 0.42 (Table 4). When the DEM metrics and the vegetation indices were added to the classification, the overall validation accuracy increased from 70.51% to 83.33% (Table 4). The highest Producer's validation accuracy (92.30%) and User's validation accuracy (92.30%) were recorded for the Unburned class (Table 4). The lowest User's validation accuracy (71.42%) occurred for the Shadow class. The lowest Producer's validation accuracy (68.42%) occurred for the Moderate Severity class (Table 4).

CONCLUSIONS
The study assessed the severity of a forest fire event that occurred from 24 th to 27 th October 2018 at Taibon Agordino using Sentinel 2 imagery to create a severity map suitable as a decision-making supporting tool for post-fire management intervention. The fire severity map was produced by applying the Random Forest classifier to a combination of Sentinel-2 raw band reflectance, associated vegetation indices, and metrics. The best accuracies were achieved when the classification was done with the original band reflectance, associated vegetation indices, DEM, and slope data, with a classification accuracy of 99.61% and a validation accuracy of 83.33%. The confusion matrix shows that there is still some confusion between the Moderate and High Severity classes. This study presents preliminary results on the use of Sentinel 2 imagery to map fire severity classes in the case of a fire in an alpine forest at Taibon Agordino, Belluno Province, Italy. Further work may be necessary to test the methodology in other locations in Italy and elsewhere. Additional work is also needed to correct the reflectance for the topography effect.