HIGH RESOLUTION FIRE HAZARD INDEX BASED ON SATELLITE IMAGES*

In December 2015, after 3 year of activity, the FP7 project PREFER (Space-based Information Support for Prevention and REcovery of Forest Fires Emergency in the MediteRranean Area) came to an end. The project was designed to respond to the need to improve the use of satellite images in applications related to the emergency services, in particular, to forest fires. The project aimed at developing, validating and demonstrating information products based on optical and SAR (Synthetic Aperture Radar) imagery for supporting the prevention of forest fires and the recovery/ damage assessment of burnt area. The present paper presents an improved version of one of the products developed under the PREFER project, which is the Daily Fire Hazard Index (DFHI).


Introduction
In December 2015, after 3 year of activity, the FP7 project PREFER (Space-based Information Support for Prevention and REcovery of Forest Fires Emergency in the MediteRranean Area) reached the end (G. Laneve et al., 2016a). However, project partners (see Table I) were requested to continue to provide users with the project products even during summer 2016. The main purpose of PREFER was to set up a common infrastructure to provide mapping tools and services, adapted to the needs of end-users working in distinct stages of forest fire management. Therefore, the project aimed at developing, validating and demonstrating information products based on optical and SAR (Synthetic Aperture Radar) imagery for supporting prevention and recovery/ damage assessment of forest fires. To attain this goal, the project was driven by three main conditions: 1) The development of a common framework regarding fire prevention and recovery, applicable to European Mediterranean countries; 2) The creation of a service available at the operational level and useful for multiple users from different sectors; 3) The timely delivery of easily-accessible cartographic tools, based on harmonized, high-quality and up-todate information.
The present paper aims at presenting an improved version of one of the products developed in the framework of the PREFER project, that is, the Daily Fire Hazard Index (DFHI) (G. Laneve et al., 2016b).
The computation of the Daily Fire Hazard Index (DFHI), is obtained starting from the calculation of the FPI (Fire Potential Index) that is based on the relationship provided by some authors (R. Burgan et al., 1998;A. Lopez et al., 2002;J. San-Miquel et al., 2003).
The estimate processing of the current version of the DFHI involves the following steps: 1) The availability of a fuel map of the area of interest; 2) The computation of the Relative Greenness based on daily revisit frequency sensors (MODIS); 3) The introduction of the meteorological data; 4) Computation of the DFHI for three days by using meteo data of the actual day and the next two days; 5) Correction of the live vegetation contribution to fueling the fire by introducing the vegetation water content; 6) Correction of the DFHI by introducing the effect of the solar illumination conditions.
In Europe, updated information of fire hazards are daily provided by the JRC EFFIS (Joint Research Centre, European Forest Fire Information System) system (J. San-Miguel-Ayanz et al., 2003). However, the current version of DFHI has a spatial resolution of 250 meters, whereas the FWI (Fire Weather Index) provided by EFFIS has a 10 km spatial resolution.
The enhanced version of the DFHI, described in this paper contemplates: an improved spatial resolution using high resolution images (Landsat 8 and Sentinel 2),

PREFER Partners
Short Name/ the introduction of the wind effects in the estimation of the dead vegetation moisture water content, the introduction of further elements in the product processing chain (estimate of the fire Rate of Spread, ROS). The results, in terms of improvement of the high fire hazard areas identification, will be discussed.
After the test period, the JRC has adopted, as method to assess the fire hazard level, the Canadian Forest Fire  Gobron et al., 2000) obtained with space-borne sensor like, for instance, MODIS (on board of the Terra and AQUA satellites) (A. Huete et al., 2002). To carry out the computation, a map of the fuel distribution on the area of interest is needed.
Other needed quantities are: the air temperature and humidity, cloudiness and rainfall. This index, commonly called integrated or advanced, is based on the FPI derived by Burgan (R. Burgan et al., 1998)

Riscos -Associação Portuguesa de Riscos, Prevenção e Segurança
Aeródromo da Lousã -Chã do Freixo | 3200-395 Lousã | Tel. 239 992 251 | Email. riscos@uc.pt The analysis concerns one of the test areas of the PREFER project, namely: Sardinia island. The target geographic area of the PREFER project was composed by all European territories located in the Mediterranean area and where fire occurrence is particularly relevant. To test and demonstrate the products and services developed, 5 smaller areas were selected based on the availability of data required to develop the products, the interest of end-users, the biophysical and social conditions of these areas and their fire occurrence history ( fig. 1).     • Minho region. The region covers 24 municipalities and has an area of 4.700 km 2 . It is characterized by a high amount of rainfall, particularly in the areas influenced by the Atlantic Ocean, contributing to the high level of biomass productivity of the region. The high biomass production, the irregular topographical conditions and the abandonment of the rural area due to recent demographic and social changes (ageing of population and migration to urban centres), favour the occurrence and propagation of forest fires; this high incidence of fires, especially in the inland municipalities, is also due to the presence of human activities (hunting, grazing) and different accessibility structures; • Andalusia area. The area covers 523 km 2 and encloses two municipalities (Cádiz and Malaga), corresponding to the Natural park of Alcornocales, a protected area.
The altitude in the area ranges between 0-500 m. The Alcornocales Park is affected by fires and requires a special effort in prevention due to its outstanding ecological value. Fires affecting the park typically originate in the border with urban areas outside the park, particularly in summer season. During the summer season, the Alcornocales Park is normally considered as "very high fire risk area", but at the same time, the cork oaks silviculture counteracts fire risk; • Corsica region. The region, with an area of 8.680 km² is composed by 2 departments: Corse-du-Sud and Haute-Corse. Its topography ranges from 0 to 2706 m. Its Mediterranean climate, with typical summer drought, is often tempered by altitude. This predominantly rural territory is classified as a French Regional Natural Park, due to its exceptional landscapes, natural habitats and cultural heritage, and benefits from important safeguard measures. On average, more than 500 fires burn over 1000 ha per year throughout the region. In this context, 360 municipalities are exposed to risk of forest fires, hence prevention and fight against this type of hazard is a major challenge, both to protect the public and to preserve biodiversity; • Peloponnese region. The Greek pilot area is sited in south-  and Andalusia (Spain). The following activities are carried out daily to elaborate the DFHI product: • Automatic download of the most recent MODIS image of the area of interest from the USGS website; • Automatic download of the meteorological data from Department of Meteorology of the Italian Air Force; • Retrieval of the NDVI and EWT from MODIS images; • Extraction of the temperature and humidity maps for the areas of interest from meteo data; • Computation of the 15 DFHI maps (3 for each area of interest); • Product upload to the PREFER dedicated ftp address.
The DFHI product has been validated both by the PREFER project team and the end users involved in the project.
Concerning the Sardinia region, the regional Civil The results, shown in fig. 1 • A vegetation fuel map; • A Digital Elevation Model; • A historical (8 -10 years) map of the maxima and minima NDVI for each pixel of the area of interest; • A historical (8 -10 years) map of the maxima and minima EWT for each pixel of the area of interest.
To compute the DFHI by using Landsat8 images a new relationship for retrieving EWT from spectral reflectances have been built. In order to estimate the relationship to be used when Landsat8 images are used we performed an extended series of simulation by using PROSAIL software.
The new relationship has been computed by simulating more than two million of profiles by using PROSAIL.
The results obtained by means of the simulation led to the following relationship between EWT and Landsat8 spectral reflectances:   • in June, in both forest and shrubland areas the temporal change of NDVI in the 80 -90% of the pixels is lower than 5%; • in July, as the season progresses, this value tends to increase, in fact, only in 30% of cases it remains under 5% of change.
Therefore, from the histogram of fig. 4 it seems that a NDVI map computed every 16 days (L8 images refresh frequency) is not enough to catch the variability of the vegetation status during the dry and hot summer season of the South-Italy. However, this problem will be solved when both Sentinel-2 satellites will be operational because in that case the revisit frequency will be improved to 5 days.   been built. In order to estimate the relationship to be used when Landsat8 images are used we performed an extended series of simulation by using PROSAIL software. The new relationship has been computed by simulating  more than two million of profiles by using PROSAIL. The results obtained by means of the simulation led to the following relationship between EWT and Landsat8 spectral reflectances: where a=-0.9573, b=0.4326, c=2.782; e=13.07; f=0.3728 and RNIR, RSWIR1 and RSWIR2 are the reflectances at 0.8 (channel 5), 1.6 (channel 6) and 2.4 µm (channel 7), respectively.
The idea to enhance spatial resolution of the daily fire hazard map assumes that the vegetation status is only slightly changing during the 16-days revisit frequency characterizing Landsat8. This hypothesis is not exactly true during the summer season and, in particular, as the dry season progresses. Of course, the change of vegetation status (reflected in the NDVI values) depends also from the type of vegetation. in June, in both forest and shrubland areas the temporal change of NDVI in the 80 -90% of the pixels is lower than 5%; • in July, as the season progresses, this value tends to increase, in fact, only in 30% of cases it remains under 5% of change.
axima and minima EWT for each pixel of the area of interest. PROSAIL. The results obtained by means of the simulation led to the sat8 spectral reflectances: The FM10hr value, which takes into account wind speed, is computed by using eq. (2) for estimating the temperature needed to produce, with no wind, an evapotranspiration equivalent to the one produced in the actual meteorological conditions (wind > 0).

Fig. 4 -Variação do NDVI em duas semanas (resolução temporal das imagens Landsat8) em dois tipos diferentes de vegetação, e dois períodos da época de incêndios (durante o verão)
Therefore, from the histogram of fig. 4 it seems that a NDVI map computed every 16 days (L8 images refresh frequency) is not enough to catch the variability of the vegetation status during the dry and hot summer season of the South-Italy. However, this problem will be solved when both Sentinel-2 satellites will be operational because in that case the revisit frequency will be improved to 5 days.    , 1998;G. Laneve et al., 2014). Such a quantity can be computed by using the meteorological parameters and lationship described by Lopez (A. Lopez et al., 2002). In the Lopez relationship the FM10hr parameter depends itly by the air temperature and humidity. In order to introduce, in such parameter, the effect of wind, as y done for introducing the effect of the exposition to the sun (topography) in the estimate of the fire hazard, ploit the evapotranspiration (ET0) relationship provided by the Penman-Monteith formula (R. G. Allen et al.,
conditions (wind > 0). Fig. 6 compares the maps computed by using the day of August 2016, with those obtained by assuming a wind speed of 2 where the difference (actual wind vs constant value of 2m/sec) affects very high fire hazard.

Conclusion
The paper aims at describing the progress made in the development of a daily fire hazard index capable to capture at the best the meteorological and vegetation conditions which, when a fire has been ignited negligently or deliberately, determine its propagation. The evolution of the DFHI index developed in the framework of the SIGRI and PREFER projects concerns the analysis of the possibility to produce accurate maps by using high resolution satellite images and devising a way to introduce the wind speed maintaining the definition of the fire hazard as given in R. Burgan, 1998 andA. Lopez, 2002. Taking into account the high variability of the NDVI (vegetation greenness) during the summer period and the time span needed to completely cover, using high resolution images (20 -30 m), a region as large as  Verifica-se, como esperado, uma certa correlação. Por razões de visualização gráfica, o total de área ardida (em hectares) foi dividido por 50.

Riscos -Associação Portuguesa de Riscos, Prevenção e Segurança
Aeródromo da Lousã -Chã do Freixo | 3200-395 Lousã | Tel. 239 992 251 | Email. riscos@uc.pt computed with the temporal frequency and spatial resolution allowed by the exploited satellite sensor (daily at 250 m with MODIS, bi-weekly at 30 m with LANDSAT/OLI and every five days at 10 m with the Sentinel-2 constellation). Now, more specifically, going to the better analyze the potential advantage of using high spatial resolution images (Landsat 8 in the present case), fig. 8 shows a detail of the DFHI map computed by using a Landsat8 image acquired on 3 rd of September 2016. The map computed using L8 image is compared with the one computed by using MODIS image of the same day. The MODIS based map has been resampled to 30 meters by using a nearest neighbor sampling method, that is all the 30 m pixels which fall in the 250m pixel assume the same DFHI value of the MODIS pixel.  Sardinia (24090 km 2 ) it seems unsuitable to use a high spatial resolution DFHI. Having a 30 m spatial resolution DFHI could be useful to monitor/forecast fire hazard in limited areas as natural parks, protected areas and WUI zones. Possibly, using Sentinel-2 images will solve the problem of image refresh frequency, thanks to the larger sensor swath and orbit revisit frequency. However, in that case a new problem related to the dimensions (in MB) of the maps (10 m spatial resolution) to manage will arise.

Riscos -Associação Portuguesa de Riscos, Prevenção e Segurança
Aeródromo da Lousã -Chã do Freixo | 3200-395 Lousã | Tel. 239 992 251 | Email. riscos@uc.pt   The effort needed to compute daily maps of DFHI for a whole region like Sardinia could be unsuitable for the use of high resolution images. In fact, in that case the area of interest could request several days to be covered by the sensor ( fig. 9) and the revisit frequency sometime could be incompatible with the rapid change of the NDVI of natural vegetation areas in summer.
• A high resolution DFHI map could be useful for monitoring/forecasting fire hazard conditions in critical areas ( fig. 8) like natural parks, wild urban interface (WUI) areas, etc.

Conclusion
The paper aims at describing the progress made in the development of a daily fire hazard index capable to capture at the best the meteorological and vegetation conditions which, when a fire has been ignited negligently or deliberately, determine its propagation. The evolution of the DFHI index developed in the framework of the SIGRI and PREFER projects concerns the analysis of the possibility to produce accurate maps by using high resolution satellite images and devising a way to introduce the wind speed maintaining the definition of the fire hazard as given in R. Burgan, 1998and A. Lopez, 2002. Burgan, R. E., Hartford, R. A. (1993. Monitoring Vegetation Greenness with Satellite data, General