RECOGNITION OF BURNED AREA CHANGE OF DETECTION ANALYSIS USING IMAGES DERIVED FROM SATELLITE SENTINEL-2: CASE STUDIO OF SORRENTO PENISOLA, ITALY

The purpose of this paper is to identify the burned areas that occurred in Italy during the summer of 2017 using change detection analysis techniques. This task is possible thanks to continuous, free and open availability of the multispectral images obtained by Sentinel-2 satellites. Indeed, comparing the satellite images of the same scene recorded at different times, it was possible to evaluate the landscape change. In this paper, the Direct Comparison change detection technique was applied to the analysis and identifi cation of burned area using several Remote Sensing indexes. In particular, in order to achieve this aim, NBR (Normalized Burn Ratio) and NDVI (Normalized Difference Vegetation Index) were used. By case study in South Italy region (Sorrento peninsula), using images derived from Sentinel-2A imagery, it was possible to identify the burned areas in a specifi c period and evaluate the performance of the two indexes. In fact, after having constructed the confusion matrix for the two tested indexes, through the use of methods that indicate the quality of a thematic map (User’s Accuracy, Producer’s Accuracy, Overall Accuracy and Kappa coeffi cient), the percentage values for each remote sensing index analyzed were compared. The analysis of the different methods revealed, from one side the high quality of the results achievable by NBR index, on the other side, it was shown how, in some areas, the NDVI was inadequate for the recognition of burned areas.


INTRODUCTION
During the summer of the year 2017, the Italian territory was devastated by numerous and dramatic fi res which destroyed extended forest areas [01]. In these burned areas also important valuable plant landscapes were involved, such as the mountains of the Vesuvius and the Lattari mountains, located in south of Italy. According to Italian law (Law 21 November 2000, no. 353) on the areas hit by fi res, a building ban is imposed for a period of 15 years (except if the purpose of the project is of public utility). Also, considering the importance of this constraint, the identifi cation of these areas plays an important role for land owners. In order to identify the burned areas, in the Remote Sensing (RS) fi eld, several approaches have been implemented [02, 03]. A widely used method is the socalled "change detection analysis" (CDA) which allows to show the land transformations by comparing images related to the same scene but recorded in different periods. In general, the Change Detection techniques can be divided in two macro areas [04]: Direct Comparison (or Pre-Classifi cation Comparison) and Post-Classifi cation Comparison. For the present study, the fi rst technique, which provides the images comparison before their classifi cation (Image differencing), is adopted. In particular, the fi rst step used in this paper, was the calculation of vegetation index and, subsequently to perform the different maps referred to two specifi c times, i.e. carrying out the so-called Vegetation Index Differencing (VDI).
In the context of the identifi cation of burned areas, a popular approach is based on the creation of the difference of Normalized Difference Vegetation Index (NDVI) calculated on satellite images in two steps: pre-fi re and post-fi re. The NDVI, as is well known in the literature [05,06,07,08], can be defi ned as the normalized ratio of the difference of the near-infrared and red bands, as shown in the equation 1: where ρ is refl ectance, and the subscripts NIR and Red describe the respective spectral regions. The differenced NDVI index (dNDVI), that is the difference between NDVI pre-fi re and NDVI post-fi re , has been successfully adopted in burned land discrimination [09, 10,11]. Based on the potential of the mid-infrared region of the spectrum, in RS fi eld the NBR (Normalized Burn Ratio) index has been developed (Formula 2), whose formula is similar to NDVI and can be defi ned as [12,13,14]: where ρ is refl ectance, and the subscripts NIR means near-infrared and SWIR means shortwave-infrared. In the same way used for NDVI, it is possible to build the differenced NBR index (dNBR) as the difference between NBR pre-fi re and NBR post-fi re [15,16].

2)
Massimiliano Pepe -Recognition of burned area change of detection analysis using images derived from satellite sentinel-2: Case studio of Sorrento Penisola, Italy Observing the formulas (1) and (2), it is possible to note how the NIR, Red and SWIR are the most suitable bands to describe the burned vegetation. To test these indexes, the images from Sentinel-2, an Earth observation mission developed by ESA (European Space Agency) as part of the Copernicus Programme, were taken into consideration in this work.

Study area
The study area concerns part of the "Monti Lattari" located in Campania region, southern Italy. Because of the high temperatures achieved during the summer, the drought and illegality phenomena, great areas belonging to the "Regional Park of Monti Lattari" were devoured by fi res. Indeed, as shown in the Figure 1, the Copernicus Sentinel-2B satellite recorded on 12 July 2017 an image that shows the severity and extent of the vegetation burned in the study area.
In particular, the subset taken in consideration is located approximately between 14°17'28" and 14°45'47" East longitude and between 40°33'31" and 40°44'26" North latitude (reference system: WGS84). In the Fig False color (Red-Green-Short Wave Infrared). In this area, the vegetation landscape varies according to the altitude: in the areas next to the sea there are the carob and the olive, in the intermediate area prevail the holm oak, the fl owering ash, the oak and shrubs like the strawberry tree and the heather while in the greater altitude there are many chestnut and alder, but also hornbeam, ash and beech [17].

Satellite platform
In order to identify burned areas, images derived by Sentinel-2 satellite platform, were taken in consideration. The Sentinel-2 is provided of multi-spectral instrument (MSI) able to acquire with 13 spectral channels in the visible/near infrared and short wave infrared spectral range. As concerning the spatial resolution, the MSI instruments acquires in different mode. In particular, four bands at 10 m  [18,19]. In the case of Sentinel-2 image, the formulation of NDVI and NBR indexes are (Formula 2-3):

Datasets
The Sentinel-2A datasets used in this study are the standard Level-1C products, were produced by radiometric and geometric corrections, including ortho-rectification and spatial registration on a global reference system with sub-pixel accuracy. The Sentinel-2 datasets have been downloaded from ESA (European Space Agency) website (https://scihub.copernicus.eu/). The two best datasets suitable for the study, pre and post fi re of the 12th July 2017, are: In both datasets, the coordinate system is WGS84/UTM 33 and, from the radiometric point of view, the pixel values are refl ectance. If the pixel values were not refl ectance, the transformation from Digital Number (DN) to refl ectance would have been applied; it is widely explained in several papers [20,21,22,23,24] and can be summarized in the following steps: • Atmospheric Correction and cloud masking; • Geometric resampling and geographic registration; • BRDF (bidirectional refl ectance distribution function) normalization; • Band pass adjustment.

Behavior of the burned vegetation
Based on the analysis of the spectral signature of the pixel in the burned areas (pre and post-fi re), it is possible to extract information using Red, Near Infrared and Shortwave infrared band. In the case of healthy vegetation, the Red band shows a high contrast and at the same time provides a clear separation between vegetated areas and uncovered soils. In addition, this region is also extremely important for its sensitivity to chlorophyll concentration in vegetation. In the NIR region, the absorbance of vegetation is low whereas refl ectance and transmittance are high.

) Color Infrared (Red-Green-Near Infrared, composite bands); c) False color (Red-Green-Short Wave Infrared composite bands)
In the SWIR region, the refl ectance and transmittance of the vegetation are low and the absorbance is very high [25]. In the post-fi re zone, the recently burned areas show a relatively low refl ectivity in the near infrared, a very low refl ectance in the red band and a high refl ectance in the short wave infrared band. The behavior of the burns and healthy vegetation region, regarding the refl ectance values in the electromagnetic spectrum, can be sketched in the following graph ( Figure  2). In addition, in the following graph are reported, varying the wavelength, the band number of Sentinel-2 used for the calculation of remote sensing indexes.

Use of the SWIR band in Sentinel-2 dataset
While in the case of images derived from Landsat 8 OLI, an American Earth observation satellite developed in collaboration between NASA and the United States Geological Survey (USGS) the Red, Infrared and Shortwave-Infrared bands have the same geometric resolution (30m), this does not occur in the case of Sentinel-2 images. Indeed, the SWIR band has a geometric resolution of 20m while the NIR and Red bands have one of 10m. Some authors used the pan-sharpening methods in order to increase the geometric resolution of the SWIR band. However, this method, widespread in order to increase the geometric resolution in visible range thanks to the acquisition of the pan-chromatic band, cannot be used in this way. Indeed, the panchromatic band covers only a visible part of the spectral wavelength range and does not cover the part beyond the NIR band. Also, it is not correct to use pan-sharpening method to increase the geometric resolution of SWIR band. In fact, in order to preserve the original physical value, the calculation of NBR was performed by sample ratio between the bands setting the spatial resolution of the output raster with the highest geometric resolution, i.e. that Red or Infrared band (10m).

Method
In order to identify the burned areas, a different procedure in relation to remote sensing index considered was adopted. In particular, for the identifi cation of the burned areas by NDVI, the fi rst step was the classifi cation of the pre and post-fi re images in three classes: soil, vegetation and water. The identifi cation of the thresholds was carried out empirically using the criterion of Maximum Likelihood (ML). This task was carried out in ArcGIS software using the tool called "Maximum Likelihood Classifi cation" which it is based on two principles: the cells in each class sample in the multidimensional space are normally distributed and Bayes' theorem of decision making. The result of the classifi cation is shown in the Figure 3. After the classifi cation, the direct comparison technique was applied to the vegetation class according the Formula 5: The resulting map of the change detection analysis on the vegetation class obtained by NDVI index, is shown in Figure 4 where the red pixels are potential burned areas. As concerns the NBR index, once calculated the index on the pre and post-fi re images, the difference map was calculated using the Formula 6: This task was carried out by ArcGIS software using raster calculator tool. The difference map obtained by these indexes is shown in the Figure 5. Once the difference dNBR map was generated, it was necessary to create a severity fi re map. This means to identify a suitable threshold value in order to discern the burned areas. For example, for dNBR map, the USGS FireMon program [26], a National Burn Severity Mapping Project of the U.S. Geological Survey, indicates severity layer variable [27], as shown in the Table 1. The levels referred to the burned areas are divided in different classes.

Accuracy workfl ow
The verifi cation of the thematic map accuracy produced by two indexes tested is usually expressed through the error matrix (also known as the confusion matrix or contingency table). The error matrix compare, on a class basis, the relationship between (known) reference data, and the corresponding results of the classifi ed image. In this case study, the true data has been carried out by photo-interpretation using Sentniel-2 image and images derived from very high resolution optical satellites.
The further data available that confi rming the exact burned areas was provided from the register (database) of the fi res drafted by local public administration. In order to build the confusion matrix, through the use of a tool developed in ArcMap environment called "Zonal Statistic as Table", it was possible to extract the statistical parameters. Therefore, the workfl ow implemented in ArcGIS software needs to obtain the accuracy value of the thematic map produced by each remote sensing used, is shown below ( Figure 6). Lastly, once build the confusion matrix, using Matlab software, the several parameters accuracy processing were performed. In order to analyze the quality of the remote sensing indexes tested, several accuracy methods were used [28,29]: User's Accuracy (UA), Producer's Accuracy (PA), Overall Accuracy (OA) and Kappa coeffi cient (k ).
The fi rst two indexes provide an estimate of the thematic accuracy of each class considered, while the last two provide a global assessment of the degree of accuracy achieved by the classifi cation [30]. The User's Accuracy of the class i (UAi), can be defi ned as the ratio between the number of pixels correctly classifi ed in the burn class and the total number of pixels assigned to that class while the Producer's Accuracy of the class j (PAj), can be defi ned as the ratio between the total of the correct pixels of the class j and the total number of pixels of that class present in the reference (column total). The Overall Accuracy of the classifi ed image compares how each of the pixels is classifi ed in or not buns area obtained from their corresponding ground truth data and it can be defi ned as the ratio between the total number of correctly classifi ed pixels (the elements present on the diagonal) and total number of pixels in the error matrix. The last method used for experimentation is the Kappa coeffi cient (also known as Kappa hat or K-hat coeffi cient), compares the number of pixels in each cell in the error matrix with the possibility to distribute pixels as a random variable and can be calculated by Formula 7 [31]: where N total number of pixels; m number of classes;

7)
D ij total diagonal elements of an error matrix; R i total number of pixels in row i; C j total number of pixels in column j.

RESULTS AND DISCUSSIONS
In the area under investigation were recognized 35 burned areas distributed over a region of over 700 hectares. The accuracy parameters achieved by NBR index are shown in the following Table 2. The accuracy parameters achieved by NDVI index in the study area, are shown in the following table (Table 3). From the observation of the Table 2, it is possible to note the high performances achievable by dNBR in order to recognize the burned areas. Instead, from the analysis of the Table 3, it was highlighted the poor performance of the NDVI index in the recognition of burned areas. Indeed, in a large and defi ned area, due of a type of vegetation that had a low NDVI index both pre and post-fi re, the use of the dNDVI was not able to recognize this area as burned. Therefore, the use of the NDVI index in the identifi cation of the burns area should be used carefully.
In addition, the difference NDVI map show only if a pixel belong the vegetal layer is change or not, but without specifying if such the change was caused by a fi re or other factors. Therefore, if the works present at moment in the literature showed a little difference in terms of accuracy in the recognition of the burned areas between the NBR and NDVI index [09, 10], in this paper it was showed that strong differences between these indexes can result in particular cases. In other word, it was established that the use of the index NDVI in presence of some type of vegetation as well as mixture vegetation/soil, can be unsuitable for the recognition of the burned areas.
Massimiliano Pepe -Recognition of burned area change of detection analysis using images derived from satellite sentinel-2: Case studio of Sorrento Penisola, Italy Figure 6: Workfl ow developed in order to build the confusion

CONCLUSIONS
Using change detection analysis on Sentinel-2A images and using NBR index, it was possible to identify the burned area in the Sorrento Peninsula and, as consequence, to build with elevated accuracy the map of this region.
Regarding the continuity of acquisition of the Sentinel-2 sensor between two dates of interest, it was possible to get images of the study area pre and post-fi re and, of consequence, to document, analyze and investigate a specifi c phenomenon. Indeed, the Sentinel constellation satellites allows covering all surfaces of the Earth between the latitudes 84°S and 84°N every fi ve days. Therefore, the ability to continuously acquire represents a valid and excellent tool for monitoring burned areas through satellite images. Lastly, considering the high geometric resolution achieved by Sentinel-2 dataset, it was possible to identify even the small burned areas. Indeed, in this case study, it was possible to obtain even the recognition of areas with an extension of 1000m 2 . This characteristic becomes very important especially in urban and semi-urban landscape, both in order to impose urban planning constraints and to establish/plan sustainable interventions on burned areas.

ACKNOWLEDGMENT
This research was part of the "Change detection techniques applied to satellite images for the identifi cation of expansions of the built territory", a research project supported by University of Naples "Parthenope". In addition, we want to thank the anonymous reviewers for constructive comments concerning our manuscript.