Geodaysit 2023

Luca Pulvirenti


A burned area database for Italy from Sentinel-2 images and ancillary data
Luca Pulvirenti, Giuseppe Squicciarino, Dario Negro, Silvia Puca

The damages generated by fire events on vegetation structure and its evolution and the economic impacts on human activity, life and infrastructures have led the scientific interest to develop tools and algorithms able to support the detection and monitoring of burned areas (BA).
The possibility of monitoring the fire evolution and mapping the BA has been strongly supported in last decades by the opportunity to use a significant quantity of satellite observations. The freely and timely availability of remote sensing data has grown so faster in the last years as well as a higher spatial resolution that makes the earth observation derived data the key component in supporting both government agencies and local decision-makers in monitoring natural disasters such as wildfire or floods.
The Copernicus Sentinel-2 with 20-m spatial resolution and a 5-day return period is a great candidate for near real-time (NRT) applications of change detection based on spectral indices. An automatic near-real time (NRT) burned area (BA) mapping approach designed to map BA using Sentinel-2 (S2) data was proposed in [1] and recently updated in [2]. The AUTOmatic Burned Areas Mapper (AUTOBAM) tool was originally designed to respond the need of the Italian Department of Civil Protection in monitoring spatial distribution and numerousness of BA during the fire season (June- September) over the Italian territory. The atmospherically corrected Level-2A(L2A) surface reflectance products from S2 are used: the automatic chain downloads and processes the most updated L2A products available on Copernicus Open Access Hub over the studied area. At the three spectral indices estimated (Normalized Burn Ratio, the Normalized Burned Ratio 2, and the Mid-Infrared Burned Index) a change detection approach is applied. AUTOBAM compares the values of these indices acquired at current time with the values derived from the most recent cloud-free S2 data. The procedure for BA mapping is based on different sequential image processing techniques such as clustering, automatic thresholding, region growing that conduce to a final BAs map with grid pixel size of 20m. Finally, a quality flag is included for each AUTOMAB BAs to certify a temporal and spatial correspondence with ancillary data, such as derived active fire detections from MODIS, VIIRS and national fire notifications.
The daily run of AUTOBAM allowed us to produce a burned area database for Italy. To evaluate the quality of the database, the AUTOBAM-derived BAs have been compared with the burn perimeters compiled by Carabinieri Command of Units for Forestry, Environmental and Agri-food protection. These perimeters represent the official burned area data for Italy. A validation procedure based of both a pixel-based confusion matrix and a set object-based accuracy metrics has been set up considering the whole Italian territory and years 2019-2021. Good results have been obtained by AUTOBAM in terms of detection capability (the Correctness parameter) and overlap factor (both larger than 60%). However, quite high values of the commission error were obtained, especially in 2019. Through a per land cover analysis, it was found that this error mostly occurred in cultivated land. Excluding the latter target, the commission error was always less than 35%, the omission error was less than 27% and the Dice Coefficient was larger than 69%. Moreover, starting from 2021, the Lazio region is providing AUTOBAM with accurate fire notifications derived from its SOUP (Italian acronym of Permanent Unified Operations Room). An experimental activity has been performed to verify whether these notifications can be used as trigger for the burned area mapping algorithm to reduce the number of false positives.


[1] L. Pulvirenti et al., “An automatic processing chain for near real-time mapping of burned forest areas using sentinel-2 data,” Remote Sens., vol. 12, p. 674, 2020.
[2] L. Pulvirenti, G. Squicciarino, E. Fiori, D. Negro, A. Gollini, and S. Puca, “Near real-time generation of a country-level burned area database for Italy from Sentinel-2 data and active fire detections,” Remote Sens. Appl. Soc. Environ., vol. 29, 2023.

AIT Contribution
Sala Biblioteca @ PoliBa
A Preliminary Investigation of the PRISMA Hyperspectral Sensor Potential for Burned Area Mapping in an Operational Context
Luca Cenci, Luca Pulvirenti, Giuseppe Squicciarino

In the past, the scarcity of hyperspectral Earth Observation (EO) data hindered the development of operational applications based on such technology. Considering the current increasing availability of this kind of data (e.g., PRISMA, EnMap), that it is expected to further grow in the future (e.g., Copernicus CHIME, PRISMA Second Generation), it is important to evaluate the potential retained by hyperspectral remote sensing for EO applications that could provide operational services in the next few years. Within this context, this work was conceived to perform a preliminary investigation of the capabilities of the PRISMA hyperspectral sensor for burned area (BA) mapping in an operational context (e.g., civil protection applications).

One of the most common approaches used for BA mapping via EO data is based on the Differenced Normalized Burn Ratio (dNBR) index, which detects the fire-induces alterations to vegetation and soils by taking advantage of the spectral information acquired in the Near InfraRed (NIR: 0.7-1.2 µm) and Short-Wave InfraRed (SWIR: 1.2-2.5 µm) bands of two images: one acquired before the fire event, one after [1]. Multispectral imagery commonly used for performing BA mapping for operational applications (e.g., Sentinel 2, Landsat) have specific NIR and SWIR bands that can be used for dNBR computation [2]. Hyperspectral images, instead, allow for several bands combinations of data acquired in the NIR and SWIR spectral regions, thereby generating numerous (and, in some cases, slightly) different definitions of dNBR maps. Amongst these bands’ combinations, the more reliable ones shall be identified (i.e., the ones capable of producing BA maps more accurate). At the same time – since the dNBR is also sensible to non-fire induced spectral alterations [1] – the less reliable ones shall be avoided.

The aim of this study was to set up an experiment in which it was prototyped an automatic methodology of operational BA mapping based on PRISMA Level2D products (i.e., orthorectified, surface reflectance imagery; GSD: 30 m). The wildfire that occurred in Pantelleria Island (Italy) on 17/08/2022 was used as a case study. For this event, there were available two PRISMA images acquired on 06/08/2022 (pre-event) and 16/07/2022 (post-event). An ancillary shapefile produced by the Copernicus Emergency Management Service (EMS) and representing the extent of the BA on 19/08/2022 (ca. 28 ha) was used as a reference layer to validate the analysis results.

The methodology that was set up – conceptually similar to the one developed by [2] – produced more than 7600 dNBR maps (obtained from the combinations of the PRISMA NIR and SWIR spectral bands), from which the pixels corresponding to the BA were mapped by using the Otsu approach for automatic threshold selection. The analysis was carried out over the whole Pantelleria Island territory, where water bodies, clouds and clouds’ shadows were masked out (as well as poor quality PRISMA bands). Then, the accuracy of the classification was quantified (as a percentage) by means of the Dice Coefficient (DC) [3], which was calculated by using the Copernicus EMS reference BA layer. According to the DC, the best bands combination for mapping the BA of the Pantelleria 2022 wildfire corresponds to the 0.903 (NIR) and 2.253 µm (SWIR) wavelengths. The DC associated with this BA map was 89.4%.

In an operational context, ancillary information (i.e., BA reference layers) are often not available to identify the most reliable bands for BA mapping. Therefore, an image-based selection criterion useful to achieve this objective shall be used. Indeed, for every NIR/SWIR bands combination used during the analysis, the spectral separability [3] of the pixels classified as BA – from the neighbouring ones classified as not BA – was computed. Then, the bands combination characterized by the highest separability value was used for identifying the best dNBR map to use for BA mapping. For this specific exercise, this combination corresponds to the 1.038 µm (NIR) and 2.245 µm (SWIR) wavelengths. The DC associated with this BA map was 88.8%. This value is very similar to the one identified via the ancillary reference BA layer.

The details of the methodology will be presented at the conference, where the analysis results will be also thoroughly discussed.


[1] van Gerrevink M.J. & Veraverbeke S. (2021). Evaluating the Hyperspectral Sensitivity of the Differenced Normalized Burn Ratio for Assessing Fire Severity. Remote Sensing. 13(22):4611.

[2] Pulvirenti L. et al. (2023). Near real-time generation of a country-level burned area database for Italy from Sentinel-2 data and active fire detections. Remote Sensing Applications: Society and Environment. 29.

[3] Roteta E. et al. (2019). Development of a Sentinel‐2 burned area algorithm: Generation of a small fire database for sub‐Saharan Africa. Remote Sensing of Environment. 222, 1–17.

AIT Contribution
Sala Videoconferenza @ PoliBa