A global fire danger rating system driven by atmospheric model forcing has been developed with the aim of providing early warning information to civil protection authorities. The daily predictions of fire danger conditions are based on the U.S. Forest Service National Fire-Danger Rating System (NFDRS), the Canadian Forest Service Fire Weather Index Rating System (FWI), and the Australian McArthur (Mark 5) rating systems. Weather forcings are provided in real time by the European Centre for Medium-Range Weather Forecasts forecasting system at 25-km resolution. The global system’s potential predictability is assessed using reanalysis fields as weather forcings. The Global Fire Emissions Database (GFED4) provides 11 yr of observed burned areas from satellite measurements and is used as a validation dataset. The fire indices implemented are good predictors to highlight dangerous conditions. High values are correlated with observed fire, and low values correspond to nonobserved events. A more quantitative skill evaluation was performed using the extremal dependency index, which is a skill score specifically designed for rare events. It revealed that the three indices were more skillful than the random forecast to detect large fires on a global scale. The performance peaks in the boreal forests, the Mediterranean region, the Amazon rain forests, and Southeast Asia. The skill scores were then aggregated at the country level to reveal which nations could potentially benefit from the system information to aid decision-making and fire control support. Overall it was found that fire danger modeling based on weather forecasts can provide reasonable predictability over large parts of the global landmass.
Wildfire activity is strongly affected by four factors: fuels, climate/weather, ignition agents, and people (Flannigan et al. 2005). Where fuel is available, weather is the most important factor in shaping fire regimes in many areas of the world (Flannigan et al. 2009). Fires are a global phenomenon extending from the boreal forests of Canada and Siberia down to Amazonia and the central African rain forests. Especially in a savanna ecoclimate, such as the Sahel and west Australia, fires are recurrent hazards because of frequent severe drought conditions but also (as for West Africa) because of agriculture practice (Swaine 1992). Fires also occur in wetter regions such as Southeast Asia (Thailand, Malaysia, and Indonesia) mostly during the dry period before the monsoon onset.
Assessments of forest fire danger in countries with extensive forest cover have for decades relied on a combination of weather information and evaluation of the vegetation state (Taylor and Alexander 2006). Traditionally, fire danger is evaluated at observation stations where weather measurements are routinely available and vegetation status is recorded. The resulting fire danger rating is then extrapolated to a large but undefined area surrounding the observation site. Examples of systems that relay on extrapolation techniques are the U.S. Forest Service National Fire-Danger Rating System (NFDRS; Deeming et al. 1977), the Canadian Forest Service Fire Weather Index Rating System (FWI; Van Wagner 1974, 1987), and the Australian McArthur (Mark 5) rating systems (McArthur 1966, 1967; Noble et al. 1980). These widely used systems provide estimates of fire danger in terms of fire ignition and behavior, energy release, and rate of spread (San-Miguel-Ayanz et al. 2003).
The European Forest Fire Information System (EFFIS; Camia et al. 2006) is currently being developed in the framework of the Copernicus Emergency Management Services to monitor and forecast fire danger in Europe. The system provides timely information to civil protection authorities in 38 nations across Europe (for details, see http://forest.jrc.ec.europa.eu/effis/abouteffis/effis-network/) and mostly concentrates on flagging regions that might be at high danger of spontaneous ignition because of persistent drought. It relies on the calculation of the NFDRS, FWI, and Mark 5 matrices and uses medium-range (1–10-day lead time) weather forecasts instead of observations to extend the advance warning. For some nations and regions, case studies have already shown the advantage of such an approach: Roads et al. (2005) and Mölders (2008, 2010) used regional numerical weather inputs to drive the NFDRS system showing good prediction skill up to one season ahead, and Preisler et al. (2009) used a combination of forecast model outputs and satellite observations to extend the spatial information provided by station-based fire index calculations.
A challenge in building such a warning system is that it requires the availability of global fields such as fuel maps, vegetation characteristics, and topography. Information such as the “greenness” of the vegetation needs to be available in real time and on a global scale at the desired resolution to approximate the traditional human judgment. The quality of such an automated system will ultimately depend on two factors: (i) the accuracy of the modeling components that translate the status of the vegetation into fire danger and (ii) the accuracy of the driving fields in predicting the real atmospheric conditions. A model-based evaluation can be used to define the upper boundary of the achievable skill of such a system. This is often called potential predictability, and its assessment is the subject of this paper.
The driving data need to be from a homogeneous global sample to assess potential predictability, which in turn allows a global comparison of fire index calculations without dependency on forecast skill. This can, for example, be achieved by using atmospheric reanalysis. These datasets are created by combining model and quality-controlled observations for past conditions in a statistically optimal way by means of an assimilation scheme (Tang et al. 2008). A reanalysis provides a dynamically consistent estimate of the climate state at each time step and can, to a large extent, be considered as a good proxy for observed meteorological conditions. Being a model integration, it has the added benefit to also provide a dynamical set of fields, including variables that are not generally observed. How close reanalysis output is to real observations depends on the amount and quality of observations available, the accuracy of the model used, and of the assimilation scheme chosen (Dee et al. 2011). Even with these caveats in mind, fire danger indices calculated from a reanalysis dataset are less affected by uncertainties in the atmospheric forcing when compared with indices calculated from forecast fields that are the result of sole model integrations. Therefore, reanalysis fire indices can be compared with observed occurrence of fire to understand the potential predictability of the modeling components in detecting fire danger and ultimately to highlight the limitations of those components themselves.
In this paper we concentrate on the prediction skills of the modeling components of the EFFIS system, that is, the Global ECMWF Fire Forecasting (GEFF) model. The GEFF model provides outputs in terms of fire danger indices that are then distributed to the EFFIS network and used among the European civil protections. Although important for an operational system, we will not detail how the information produced by GEFF are translated into fire suppression actions at the European central level and in the national agencies that are part of the EFFIS network. The objective of this work is to introduce the modeling components and identify regions where the system shows high and low potential predictability of fire danger.
a. Fire danger metrics
Each of the three fire rating systems implemented into the GEFF model provides a comprehensive set of outputs that characterize different aspect of fire conditions. It is therefore important to clarify which aspect related to fire is going to be analyzed in this work. The primary scope of the EFFIS is to provide daily information of fire danger, which is intended as a general term to express an assessment of both fixed and variable factors of the fire environment that determine the ease of ignition, rate of spread, difficulty of control, and fire impact as defined by the National Wildfire Coordinating Group (http://www.nwcg.gov/term/glossary/fire-danger). Fire danger has, therefore, several components, and the rating indices chosen should be a combined metric expressing the probability of ignition, the speed and likelihood of spread, and the fuel availability (sustainability of the event).
Among all the fire indices that are outputs of the GEFF model, the FWI and the fire danger index (FDI) from the McArthur Mark 5 system are selected as good indicators of fire danger containing both a component of fuel availability (drought conditions) and a measure of ease of spread. In the NFDRS, two indices could be used to indicate fire danger: the ignition component (IC) and the burning index (BI). Both of these indices have a component of spread and a component related to the available burning energy. Since these indices are strongly dependent on the fuel type and the local atmospheric conditions, they measure local short time variation of burning conditions as opposed to long-term danger conditions. Looking at the correlation between the BI and the IC and both the FWI and the FDI (not shown), it was found that the IC has a better correlation with the FWI and the FDI and was therefore selected as the best index within the NFDRS to compare with the other two fire danger rating systems.
In this study we limit the analysis of fire predictability to these three indices so as to provide a concise analysis of the generic fire danger predictability. Keep in mind, though, that in the actual operational implementation of EFFIS, which uses real-time forecasts as atmospheric forcings, once dangerous areas are identified, a more detailed analysis can be performed to characterize the event in terms of, for example, expected containment actions required. (In operational practice, this is usually achieved by gathering additional information provided by all of the other indices that are calculated as a part of the fire danger rating system provided in Table 2, which is described in more detail below.)
1) National Fire-Danger Rating System
In the NFDRS, the characteristics of fire danger are functions of fuel type, topography, and weather (Bradshaw et al. 1983; Cohen and Deeming 1985; Burgan 1988). The model explicitly calculates the moisture content of dead and living vegetation.
Dead fuel is divided into classes according to its fast or slow response to the changes in atmospheric temperature and humidity forcing, while live fuel is divided into herbaceous and woody shrubs. The ignition component is a rating of the probability, between 0 and 100, that a firebrand will cause a fire requiring a suppression action. Its value depends on the moisture content of the fastest responding fuel and the available fuel energy stored as a consequence of the lack of moisture in the deeper layers.
2) Fire Weather Index System
Similarly to the NFDRS, the FWI describes the effects of atmospheric temperature, humidity, precipitation, and wind, first on the fuel moisture content and consequentially on the fire behavior and occurrence (Van Wagner and Pickett 1985; Stocks et al. 1989). While the NFDRS allows the user to specify different fuel types, the FWI is specifically calibrated to describe the fire behavior in a standard jack pine stand Pinus banksiana typical of the Canadian forests (Van Wagner 1974). Despite this limitation, the index has successfully been used in countries where vegetation is dissimilar to Canada (Taylor and Alexander 2006) such as Australia (Cruz and Plucinski 2007), New Zealand, and Malaysia (Taylor and Alexander 2006) and can therefore provide useful information worldwide. In the FWI system, the fuel moisture content is calculated with different codes depending on the fuel consistency. Litter and fine fuels occupy the first fuel bed layers. Its moisture content is calculated using the fine fuel moisture code (FFMC). The moisture content of the loosely compacted organic layers of moderate depth is instead given by the duff moisture code (DMC), while the moisture of deep, compact organic layers is evaluated using the drought code (DC). From these diagnostic, weather-driven fuel moisture calculations, the FWI model calculates fire behavior indices in terms of rate of fire spread [initial spread index (ISI)] and fuel available for combustion [buildup index (BUI)]. The FWI integrates current ISI and BUI to produce a unitless index of general fire intensity.
3) McArthur’s Forest Fire Danger Meter (Mark 5)
The McArthur’s Forest Fire Danger Meter (Mark 5) was developed to monitor fire danger throughout eastern Australia. Unlike the other two indices in this model, there is not an explicit description of the evolution of moisture in different fuel types. The formulation in Noble et al. (1980) uses atmospheric conditions to evaluate a generic component representing fuel availability called the drought factor (DF). The DF is used to then calculate the FDI, which provides an assessment of fire danger due to the combined effect of drought condition and fuel availability. The drought factor is given a number between 0 and 10 and represents the influence of recent temperature and rainfall events on combustible material availability. It is partly based on the soil moisture deficit, which is commonly calculated using the Keetch–Byram drought index (KBDI; Keetch and Byram 1968), which measures the effects of seasonal drought on fire danger. The actual numeric value of this index is an estimate of the amount of precipitation needed to bring the soil back to saturation. In its original formulation, the KBDI only deals with the top 8 in. (1 in. = 2.54 cm) of the soil profile, meaning the maximum KBDI value is 800.1 KBDI is sometimes used as a fire danger indicator in itself. Its relationship to fire danger is that as the index value increases, the vegetation is subjected to increased stress due to moisture deficiency, living plants die and become fuel, and the duff/litter layer becomes more susceptible to fire (Preisler et al. 2004).
b. Fire danger reanalysis
The analysis system (Fig. 1) was designed using weather information from the ERA-Interim reanalysis system of the European Centre for Medium-Range Weather Forecasts (ECMWF) (Dee et al. 2011). ERA-Interim is the latest of the ECMWF reanalysis products and stretches from 1 January 1979 and is extended forward in near–real time. It employs a sequential 4D-var data assimilation scheme that ensures the optimal consistency between the available observations and the model background (Courtier et al. 1994). The reanalysis data have an original spatial resolution of around 80 km and have been bilinearly interpolated to a regular latitude–longitude grid of 0.25° to be comparable with the available observational dataset of fire activity. Therefore, ERA-Interim has quite a coarse resolution and does not represent small-scale processes that might be responsible in establishing local favorable conditions for small fires. For this reason, most of the following analysis will concentrate on medium to large fire events (>2500 ha).
In addition to these daily data, the GEFF model also relies on constant fields (Fig. 2). There are three tiers of input data to the model. Tier 1 comprises the so-called climatological fields that are precomputed and kept invariant during the runs. Examples are the land–sea mask, the vegetation cover, and the orography. Tier-2 data are daily averaged fields such as 24-h accumulated precipitation, minimum and maximum daily temperature, and humidity. These are calculated from the 3-hourly outputs of ERA-Interim. Tier-3 data are atmospheric fields at a nominal 1200 local time when the condition for wildfire is most favorable. A model integration at any nominal time will simulate the atmospheric conditions at a different local time depending on the location. A temporal and spatial collage of 24-h time model simulations is performed to produce a snapshot at 1200 local time. Thus temperature and relative humidity fields are cut into 3-hourly time strips using the closest 3-h forecast output and then concatenated together so that the final field is representative of the conditions around the local noon within the 3-h resolution available. Using this method, the driving forcings are a composite of forecast outputs at different lead times in a 24-h interval and could therefore have different forecast accuracy. This inconsistency is assumed insignificant given the limited difference in forecast skills in a 24-h lead time range (Buizza et al. 1999).
The schematic in Fig. 2 shows how these three tiers of forcings are linked to the various output components of the fire index system. The summary of the data input needed is given in Table 1, while a detailed explanation of the preprocessing needed to obtain the climatological fields (tier 1) is provided in the appendix. Since the various components of the three fire rating systems are calculated from the same set of forcings, they are directly comparable. Although this work concentrates on fire danger, the system also provides a comprehensive set of fields modeling vegetation stress, probability of ignition, and fire behavior. A list of these output variables of the system is shown in Table 2.
The reanalysis dataset for fire danger was calculated starting from 1 January 1979 and run for 36 years with a daily time step following the availability of ERA-Interim. Since the initial conditions were set using an idealized state the fire variables suffer from the so-called spinup in the first months of the forecasts as the model drifts to its own equilibrium state. The first year of simulation is therefore discarded from further analysis. The resulting fire analysis does not incorporate observations in contrast to atmospheric analysis systems.2
For this reanalysis, all of the fire indices were set to zero (and the moisture content of the fast-responding vegetation classes was raised to 35% in the NFDRS) if snow was on the ground or the daily precipitation is above 1.5 mm day−1. Areas where vegetation fuel is not available were masked out (Fig. A6 in the appendix).
National inventories of wildfire activities exist in many countries (e.g., Westerling et al. 2006) but they do not have the global coverage and/or the extended record needed for a validation of a fire danger system at a global scale. Satellite observations can supply a valid alternative especially as they cover remote areas where in situ observations are sparse (Flannigan and Vonder Haar 1986; Giglio et al. 2003; Schroeder et al. 2008). Satellite data have been used to monitor biomass burning at regional and global scales for more than two decades using algorithms that detect the radiative emission from active fires at the time of satellite overpass and in the last decade by using burned area algorithms that directly map the spatial extent of the area affected by fires (Wooster et al. 2003; Giglio et al. 2006).
The burned area dataset of the Global Fire Emissions Database (GFED4) combines several satellite products in a homogeneous time sequence of events (Giglio et al. 2013) from August 2000 to the present. Among estimations of fire emission, it provides daily burned area fraction with a 0.25° resolution. GFED4 combines 500-m MODIS satellite burned area maps with active fire data from the Tropical Rainfall Measuring Mission (TRMM) Visible and Infrared Scanner (VIRS) and the Along-Track Scanning Radiometer (ATSR) family of sensors. The daily burned area dataset is used in this paper to validate the relationship between the modeled fire danger and the observed occurrence of fire episodes. Fire events are defined when the burned area fraction is larger than 0.1 (10% of the pixel area). Considering the resolution of the dataset, this choice implies that we “activate” a cell when fires have an extent of at least 2.5 km2 (2500 ha). Hantson et al. (2015) show that fire size follows a negative power-law distribution. Large fires (>104 ha) are orders of magnitude less frequent than small ones (<102 ha). Since the fire mask is constructed on fires larger than 2500 ha in our analysis, we neglect events smaller than this size, which are the most frequent. On the other hand, note that ERA-Interim has an original resolution of 80 km and it is then interpolated on to a 25-km resolution to be comparable with the observed dataset. Therefore, small-scale conditions that can impact fire danger at the local level are not explicitly represented by the weather forcings used. The prediction of fires at these unresolved scales could prove challenging. To understand the impact of fire sizes on the overall prediction skill of the system, another mask is created that takes into consideration only fires smaller than 2.5 km2 by defining the mask when the condition 0 < burned area < 0.1 is met. (This mask for small size fires will be used only for the generation of Fig. 9, which is described below).
Figure 3 shows the number of observed fire days during the 4900 days covering the period August 2000 to December 2013. All fires are included, even the ones induced by human actions, which, despite requiring favorable climatic conditions to be sustained, are not explicitly modeled by the fire indices algorithms used in this paper.
The three indices exhibit different ranges of values, which makes a direct comparison difficult and presents a barrier to integrate this information in an early warning application. The global cumulative distribution function (CDF) for all values above zero in the 11 yr from August 2000 to December 2013 is shown in Fig. 4. It shows, for example, that FWI ≥ 40 occurs in only 20% of the cases while for the same value a frequency of 5% of the cases for the Mark 5 FDI and NFDRS IC is estimated.
To make the numerical values of the indices directly comparable a transformation is performed on the raw index value to obtain a normalized index . This is simply done by means of the inverse of the CDF function:
In this way, all indices are normalized over [0, 1] using the maximum and minimum (above 0) computed over all days of the available time series. Note that Fig. 4 represents a global mean and has been reported for documentation purposes only. At a single site these curves might be different since they the range of possible fire danger conditions for that location.
e. Verification methods
1) Discrimination skill
The IC, FWI, and FDI provide information on the severity of drought conditions and the availability of fuel and thus the probability of spontaneous ignition occurrence. The numerical value should be interpreted probabilistically meaning that high values are no guarantee of actual fire occurrence. A desirable quality for the purposes of fire action planning for any system is its capability to discriminate between fire occurrence and nonevents. Therefore, one would expect high values of the indices to predict fire and low values otherwise. This information can be gathered using a discrimination diagram, that is, plotting the conditional distributions of the forecasts. For binary events (fire/no fire), this diagram plots the conditional distribution of the forecasts given that the event occurred and the conditional distribution of the forecasts given that the event did not occur. Ideally, the two distributions should be separated from one another, becoming two distinct spikes for perfect prediction.
2) Fire detection skill
The calculation of fire detection skill has to take into consideration that fires are events with a low frequency (rare events) (Coles et al. 1999; Ferro 2007; Ferro and Stephenson 2011). The assessment of the quality of the prediction is therefore complicated by the fact that measures of forecast quality typically degenerate to trivial values as the rarity of the predicted event increases (Coles et al. 1999). This is easily understood by considering the contingency table for the observed and predicted fire events in Fig. 5b. Once the occurrence of the event is defined for the forecast (e.g., when ≥ 4th quartile) it is possible to count the number of hits (A), misses (B), false alarms (C), and correct negatives (D) in comparison with what was observed. The resulting table can then be used to derive classic skill scores such as the probability of detection , or the Brier score , to name just two. Coles et al. (1999), Ferro (2007), and Ferro and Stephenson (2011) showed how these common skill scores tend to vanish as the base rate of observed events regardless of the actual forecast skill. Stephenson et al. (2008) and Ferro and Stephenson (2011) proposed a series on nonvanishing measure for rare events. The extremal dependency index (EDI) is less dependent on the base rate and more difficult to hedge [see Ferro and Stephenson (2011) for an extensive discussion on the EDI properties]. EDI provides a skill score in the range [−1, 1]. EDI takes the value of 1 for perfect forecasts and 0 for random forecasts. It is >0 for forecasts that have hit rates that converge to 0, and slower than those of random forecasts, and can be negative in the opposite situation. Therefore, the system beats a random forecast for values > 0 and could be considered skillful.
3) Fire localization
The accuracy in the localization of an event might not be the most important aspect for an early warning system. High fire danger forecast in the neighborhood of an observed event still adds useful information for the scope of emergency planning and should therefore be positively rated. To take this aspect into account the prediction score calculation follows a fuzzy approach using a spatial filter that aggregated a matrix of 3 × 3 neighboring pixels around the verification point (Fig. 5a). One contingency table is created from the nine time series. This means that any hits in the neighborhood will count toward the EDI value in the verification grid box.
a. Comparing indices
The combined probability density function of any index combination (Fig. 6) shows the different behavior of the FWI, Mark 5, and NFDRS systems. The FWI and FDI variables are calibrated to represent the inflammability of a specific vegetation condition (i.e., boreal forest). Conversely, the NFDRS explicitly represents the evolution of the vegetation moisture content in different ecosystems, specifically the 20 fuel moisture classes given in Table 3. The ignition component of the NFDRS is calculated for different vegetation regimes and is then compared with the two other indices representing a constant vegetation. This difference is probably responsible for the larger spread in the joint probability function observed when NFDRS is used. The contour lines show the same probability density function but using a reduced dataset where at least one fire event was recorded in the GFED4 dataset (see Fig. 3). Apart from the need of rescaling, the indices show a very good correlation among each other when a linear model is used to calculate the correlation coefficient r. This proves that overall they are representative of the same phenomenon.
b. Global predictability
The potential predictability evaluates the capability of the system to effectively flag regions and periods as at high (low) danger when actual fire events (nonevents) were observed. The discrimination diagram in Fig. 7 shows the conditional distributions of the forecasts whether the event did or did not occur. If the two distributions are well separated, the indices have a good discrimination capability and the information can be useful in an early warning system. Only two distinct spikes would be present in the case of a perfect forecast.
In roughly 80% of the cases, the indices are able to flag as above normal (i.e., >50% of climatological value) cases that developed in fire events and flag as below normal (i.e., ≤50% of climatological value) situations that did not develop into fire events. Remarkably, all the fire indices are also able to predict more severe events. A fire developed 60% of the time in which values were in the upper percentile (above 75% of all cases). FWI and Mark 5 are extremely similar in their statistical behaviors because of the similar nature of the modeling components. The IC is slightly more cautious in predicting fire events but, in contrast to the other indices, provides a lower false-alarm rate.
c. Regional fire danger detection
The global assessment in the previous section is useful to gain an appreciation of the average potential capability of the system. A regional analysis is necessary to understand in which countries/regions the automatic system has enough predictive skill to be useful to plan fire control actions. The regional analysis is performed using the fuzzy pixel logic described before. Two verifications are performed when forecast events are defined for the index being larger than the first upper quartile (≥50%; Fig. 8): 1) using the standard fire mask (fires > 2500 ha) and 2) using the fire mask for small fires only (Fig. 9).
Figure 8 shows that the predictions are very good in regions of the planet covered by boreal forests (taiga ecosystems that cover nations such as Russia, Canada, and the Nordic countries). Since the boreal forest zone consists of a mixture of conifers (white and black spruce, jack pine, tamarack, and balsam fir) it is not surprising that the FWI performs best among the indices in these areas as it is specifically calibrated for this vegetation cover. Prediction skill is higher in the Canadian boreal shield west ecozone (Stocks et al. 2002) where large fires occur frequently when compared with the Canada’s montane cordillera ecozone (Stocks et al. 2002) where fires are numerous but tend to be smaller. Northern boreal Eurasia and Siberia in particular present a very similar range of fire weather conditions as in boreal Canada. In these regions the vegetation is quite homogeneous and the values of the indices are controlled mostly by the weather forcings. It has to be noted, though, that fire regimes can be very different, with the Siberian fires tending to be not as large as in Canada, relatively frequent, and having moderate to high intensity (de Groot et al. 2013).
In Australia, fires can develop in two very different environments. They can either burn in mountainous or alpine areas, which are usually densely forested, or they can start along flat plains or areas of small undulation, predominantly covered in grasses or scrubland. In the first case fire episodes can be extremely intense and long lived, while in the second case fires move quickly, fueled by high winds in flat topography, quickly consuming the small amounts of fuel/vegetation available (Bradstock et al. 2002). The predictive skill of the GEFF system seems to be able to distinguish between these two regimes and performs better in forested areas. Indeed, Luke and McArthur (1978) suggest a modification of the FDI index for grassland that could improve the prediction in this ecoclimate, although it is not implemented in the GEFF system.
Fires in Southeast Asia—comprising the countries of Thailand, Malaysia, and Indonesia—are usually human caused for the purposes of gathering nontimber products and agriculture or, as in Indonesia, due to deforestation for establishing plantations. The fire seasons in this region are controlled by rainfall seasonality associated with the monsoon, which produces an annual, or in some regions semiannual, wet–dry cycle. The NFDRS system requires quite an accurate knowledge of the vegetation cycle. Vegetation green-up and curing are, in the original implementation, human triggered, while in GEFF they are read from a mean climatological database (see the appendix). This highlights the shortcoming of a global implementation for a model that highly relies on local knowledge. The NFDRS would probably be much more locally accurate if real-time data for the vegetation stage were to be provided.
In the Mediterranean region, vegetation is dominated by a combination of shrublands and low forests. Persistent dry climatic conditions in summer favor the establishment of intense fire seasons that are only limited by the availability of fuels (Pausas and Paula 2012). Despite the different types of vegetation than are found in the boreal forests, all the indices perform remarkably well in marking fire events in these ecosystems, especially in the southern part of Spain, Greece, and Italy. The climate in Central Europe is subcontinental temperate, and the vegetation is characterized mostly by deciduous broad-leaved forests. The peak of the fire activity tends to be just after snowmelt and before leaf flushing and is driven more by short-term dryness of surface soil layers than long-term drought (Wastl et al. 2013). The potential predictability of fires is lower than in the Mediterranean region.
Finally, the large forests of South America and central Africa are characterized by large seasonal fires mostly initiated by agricultural burning. The peak of the fire season is between August and September, which coincides with the end of the dry season. Given the vast availability of fuel, the main cause of fire danger is the establishment of drought conditions, which are well predicted by all indices.
The global analysis in Fig. 8 is repeated using the mask for small fire (Fig. 9) to highlight that the EDI skill score is degraded everywhere on the planet by about 20% on average. The coarse resolution of ERA-Interim cannot resolve small-scale variations in the weather parameters that might result in the development of smaller fires. The relationship between weather anomalies and fire events is therefore weakened.
d. Aggregation at country level
EFFIS (Camia et al. 2006) has been designed with the main aim of providing decision support for the management of fire danger at European and national level. Wildfires are clearly a cross-border phenomenon, but for decision-making support, an aggregation of forecast skill to a national level can be useful to understand in which nation GEFF products could also be employed. For every nation, Fig. 10 shows the surface that was affected by at least one fire during the observational period as a percentage of the nation’s ground area. The country inventory and the surface-area data are taken from Esty et al. (2008) and are not available for a few nations such as Afghanistan and Serbia. For very small countries such as Denmark and Belgium, the coarse 25-km2 resolution of the observed dataset can also lead to misleading overestimations of the percentage of burned area. Fires might have occurred with smaller extent but are then aggregated to a larger grid at the resolution of the dataset and in some cases the grid is too coarse to “resolve” small nations. However, for most of the countries, the plot provides a reliable measure of representation by providing system skill at a country level. For very low coverage (≤10% of the country surface), the given average fire index skill might not be a meaningful indication of the real performance of GEFF in that country. This is especially true for nations with very heterogeneous vegetation covers and fire activities.
Figure 11 shows the mean skill of the three fire rating systems aggregated at a country level. Keeping in mind the information provided by Fig. 10, it shows that the system is able to beat the random prediction in almost all nations. All indices perform better where fire is moisture limited, such as the boreal forests, where fires are fuel limited (e.g., in the savanna).
In this paper, we have presented an operational global fire danger system that relies exclusively on forcings from atmospheric modeling. The Global ECMWF Fire Forecasting system was developed from the collaboration between the European Joint Research Centre (JRC) and the ECMWF. GEFF is the modeling component of the European Forest Fire Information System (Camia et al. 2006), which is being developed in the framework of the Copernicus Emergency Services to provide a platform of shared information on fire danger to civil protection authorities of 38 nations across Europe. GEFF calculates daily prediction up to 10 days ahead of all the indices from the U.S. Forest Service NFDRS, the Canadian Forest Service FWI, and the Australian McArthur (Mark 5) rating systems using the ECMWF forecasting system as the atmospheric forcings.
This study introduced the modeling components of the GEFF system and has assessed its maximum potential skill to identify conditions of fire danger. To this end the analysis has employed meteorological forcings from the latest ECMWF reanalysis dataset (Dee et al. 2011), ERA-Interim, instead of the more uncertain forecast fields. Reanalysis products can still be biased relative to observations, but they have the advantages over raw weather station data by providing spatially and temporally continuous records consistent with the physical constraints imposed by a numerical weather prediction model and an optimal data assimilation algorithm. Reanalysis fire indices can therefore be compared with observed occurrence of fire to understand the potential predictability of the modeling components and where GEFF could provide useful information for fire control. The analysis has been performed interpolating the original 80-km ERA-Interim grid to a regular latitude–longitude grid of 25 km to match the available burned area dataset. ERA-Interim has therefore quite a coarse resolution and does not represent small-scale processes that might be responsible for establishing local favorable conditions for fire danger. For this reason, the comparison has been limited to medium to large fire events (≥2500 ha).
In large areas of the planet all of the selected indices from the three fire danger rating systems are able to identify dangerous conditions for fire events. Where fuel availability is not limited—such as in the boreal forests, the Mediterranean region, South America, and central African regions—fire events are mainly controlled by persistent drought conditions, and GEFF prediction skill is high (well above the random forecast). Conversely, in temperate regions, such as the mountainous regions of Central Europe, fuel availability is limited. Here fire can depend on highly stochastic conditions, such as the short-term superficial drying of the available organic matter on the ground. In these conditions, the skill of the system degrades although it is still above the skill of a random forecast because of the small-scale nature of the processes involved.
The FWI and Mark 5 FDI indices have very similar behavior and seem to outperform the NFDRS IC index in some regions of the planet. It is fair to recall that while the FWI and FDI only rely on weather inputs for their calculations, the IC highly depends on the knowledge of fuel conditions such the vegetation state and its annual cycle. This information has been provided in a mean climatological way and might not be accurate enough to exploit the potential for this index. Since the GEFF system is intended to provide information on a regional to global scale, some compromises had to be made, the most important being the substitution of human judgment with mean climatological conditions for the vegetation evolutions that penalized the NFDRS system.
The very promising results show that climate model simulation may usefully extend the early warning available from environmental monitoring. It is important to emphasize that this study is a first step and is limited to identifying the potential skill in such a system. Actual skill of the operational forecasts extending up to 10 days will be lower because of the use of more uncertain weather forcings. In another sense, the assessment also represents a lower threshold of potential skill, since improvements in the forecast modeling systems will increase skill over time. A full analysis of the prediction skill will be the natural next step.
This research was funded by the EU Project ANYWHERE (Contract 700099) and the Global Fire Contract 389730 between the Joint Research Centre and ECMWF. We are grateful to three anonymous reviewers for the many suggestions that greatly improved the manuscript. We are in debt to David Lavers for reading and improving the manuscript.
Invariant Fields in the GEFF System
a. Vegetation stages
An annual climatology of vegetation stages is derived starting from observations of leaf area index (LAI) recorded by the MODIS instrument (Myneni et al. 2002). LAI is defined as the one-sided green leaf area per unit ground area and has been used in land surface modeling as an indicator of the vegetation state (green-up, green, transition, and curing processes) (Knote et al. 2009). The dataset used here has been processed from the products MOD15A2 and spans the period 2000–13 with a time resolution of 10 days, as described in Boussetta et al. (2013). Quality control and temporal and spatial aggregation have been performed on the original MODIS data to generate an annual climatology of LAI that is now employed operationally as climatological LAI field in the ECMWF’s integrated forecasting system (IFS) model (Boussetta et al. 2013).
Starting from this LAI dataset, we have calculated vegetation stage intervals (examples for two days are shown in Fig. A1). For each location a low-bandpass filter is applied to the LAI climatological annual cycle to remove data variability at scales lower than the seasonal. The resulting curve is processed in terms of quartiles. Days with LAI values larger than the upper quartile are marked as “green,” while days with LAI value lower than the lower quartile are flagged as “cured.” All days remaining are checked against the derivative of the LAI annual curve. If a point belongs to the up-slope part of the curve (i.e., positive derivative) it is then flagged as “pregreen,” otherwise it is marked as “transition.” This approach leads to very realistic climatological vegetation stage in locations where vegetation is characterized by a clear annual cycle. In the extratropics where precipitation is associated with frontal systems and is in general seasonal, this is usually a good approximation (Figs. A2a,b). In other regions, mostly in the tropics where precipitation is driven by, for example, monsoon systems, a single peaked annual vegetation cycle might not represent fully the interannual variability. In East Africa, for example, there are two monsoon-driven rainy seasons. They are responsible for two germination phases in spring and autumn (Fig. A2c). Only the green-up and green phases associated with the long rain in April are picked by the automatic procedure after the filter smoothing. From the point of view of the fire danger assessment, this limitation of the system in local points is nevertheless deemed acceptable since it goes in the direction of increasing the false alarms while underestimating the missed events.
b. Orographic parameters
The model orographic parameters needed to initialize GEFF are taken from IFS model’s invariant fields for consistency. The orography is based on an interpolated version of the terrain elevation dataset GTOPO30 at 30-s resolution everywhere except in Greenland where the Kort-og Matrikelstyrelsen (KMS) DEM also at 30-s resolution is used.A1 The orographic high-resolution map is also used to derive the land–sea mask, which provides the percentage of grid box covered by land. The GEFF model converts this parameter into a mask where grid points that have more than zero land are considered as land points. This is a very conservative approach that nevertheless allows the inclusions of all coastal points and small islands into the fire danger calculations.
The orography slope is an important factor to characterize fire danger behavior. This parameter is calculated at the model target resolution from the high-resolution GTOPO30 orographic dataset (Baines and Palmer 1990; Lott and Miller 1997). The slope value in IFS is transformed into slope classes as in NFDRS implementation (Cohen and Deeming 1985) following a simple linear rescaling (Fig. A3).
c. Vegetation cover
The vegetation cover is an additional mask to exclude points with no vegetation from the fire danger calculation. In IFS, vegetation is represented by six climatological parameters: vegetation cover of low vegetation, vegetation cover of high vegetation, low vegetation type, high vegetation type, leaf area index for low vegetation, and leaf area index for high vegetation. These parameters are built from the Global Land Cover Characteristics (GLCC) dataset, which was derived from 1 year of Advanced Very High Resolution Radiometer (AVHRR) data, digital elevation models, ecoregions, and map data (Loveland et al. 2000). The nominal resolution is 1 km, and the data come on a Goode homolosine global projection. In GEFF, the fractional covers for low and high vegetation are summed to obtain a total vegetation cover. Following a conservative approach, any point with vegetation cover greater than zero is assigned vegetation mask true.
d. Mean cumulative annual precipitation
One of the inputs needed in the calculation of the Keetch–Byram drought index is the climatological total annual expected precipitation in a location. The Keetch–Byram drought index (Keetch and Byram 1968) is essentially a bookkeeping record of water deficiency in the soil. In these terms it describes anomalies relative to a defined mean state. Therefore, to avoid inconsistency (biases) between the background status defined by the expected climatological rain availability and what is predicted daily by the forecast, all terms of the water balance at the surface should be derived in a consistent way. The annual accumulated climatology for rainfall is therefore calculated using the 1980–2014 ERA-Interim (Dee et al. 2011) daily dataset by first accumulating the precipitation over every year and then calculating the average (Fig. A4).
e. Climatic classification
In the NFDRS, the response of fuel moisture to environmental conditions is influenced by the mean climate. Therefore, a climate class must be specified for each grid point. In the 1978 implementation of the NFDRS (Deeming et al. 1977) four classes were used to characterize the U.S. climatic regions: 1—arid/semiarid (e.g., desert and steppe), 2—subarctic (e.g., taiga), 3—humid (e.g., forest), and 4—wet (e.g., rain forest). This classification was mostly based on the climate of temperature and precipitation in a given location, which is very similar to the approach adopted by W. Köppen for his noted world climate classification (Kottek et al. 2006; Chen and Chen 2013). To provide a numeric input to the GEFF system, we have digitalized the five class maps available in Chen and Chen (2013) with the following mapping conversion: class 4 (wet) to tropical, class 3 (humid) to mild/temperate, class 2 (subarctic) to snow, and class 1 (arid/semiarid) to dry (Fig. A5). In the polar regions, vegetation and thus fire fuel are almost negligible, and consequently the associated danger of wildfire is null. The polar climate class is therefore discarded.
f. Fuel model
The global fuel map used in the study (Fig. A6) is a product developed by JRC and derived from the Global Land Cover 2000 (GLC2000; http://forobs.jrc.ec.europa.eu/products/glc2000/glc2000.php) database (Bartholomé and Belward 2005) and related regional products for Africa (Mayaux et al. 2004), Asia (Bartalev et al. 2002, 2003), and Europe. The GLC2000 was overlaid on the NFDRS fuel models (Deeming et al. 1977; Bradshaw et al. 1983) as mapped on the conterminous United States by Burgan et al. (1998). It contains 20 vegetation species listed in Table 3 plus a series of classes for marsh, water bodies, urban, and agriculture. The derived co-occurrence matrix of NFDRS fuel models and GLC2000 land-cover classes was used as initial reference to derive, through expert judgment and interpretation, the relationships of fuel model classes with the rest of the world and regional land-cover legends. The resulting map has 1-km spatial resolution and is intended to provide a first approximation of the spatial distribution of fuel complexes throughout the world, as classified according to the NFDRS (Deeming et al. 1977; Bradshaw et al. 1983).
8.00 in. of precipitation would be needed to bring the soil back to saturation.
While the FWI and Mark 5 systems are diagnostic formulations that could not benefit from the assimilation of observation, the NFDRS includes prognostic equations for the fuel moisture content that could, in theory, ingest available measurements provided an assimilation scheme were available. Nevertheless, direct observations of fuel moisture are only collected in isolated research projects thus not available at the spatial scale required for global applications.
The GTOPO30 dataset, as used in the IFS, was completed in 1996 through a collaborative effort led by the U.S. Geological Survey Earth Resources Observation and Science Center (https://lta.cr.usgs.gov/GTOPO30) and was derived from a variety of information sources. Extensive information can be found in ECMWF (2014).