Can the National Fire-Danger Rating System (NFDRS)-1978 of the United States be effective in other regions? Israel as a case study

Fire risk assessment

Wildfire activity is projected to increase in many regions worldwide in the coming decades due to climate change (Liu et al. 2010, 2014; Dupuy et al. 2020), as well as human-driven changes in land use and management such as urbanization, agriculture and forestry practices (Butsic et al. 2015; Chas-Amil et al. 2015). The relative importance of these factors can vary depending on specific local conditions, underscoring the need for context-specific wildfire management strategies. Hence, there is a vital need for improvements in fire danger estimation to better understand and predict fire risk (Castel-Clavera et al. 2025) and to inform management decisions aimed at reducing the negative impacts of fire (Jones et al. 2022). Preparedness and alertness of fire and rescue teams can minimize the time it takes them to arrive at a fire and extinguish it while it is still small (Buffington and Ezekoye 2019; Ivanov 2025). Notably, vegetation (fuel) conditions, weather and topography are among the three main natural factors affecting the scale of wildfire propagation and fire intensity (Moritz et al. 2005; Dillon et al. 2011). However, fires have complex dynamics, as the relationships among vegetation, climate, human activities and topographic components are nonlinear and convoluted, and vary across regions and initial conditions (Turco et al. 2017; Jones et al. 2022). To address this complexity, wildfire risk assessment relies on conceptual frameworks and analytical tools that aim to capture the initiation, spread and impacts of fire under diverse ecological and climatic conditions (Moritz et al. 2014; Freeborn et al. 2016; Xofis et al. 2020; Grillakis et al. 2022).

Two primary frameworks are commonly used to conceptualize wildfire risk. The earlier framework (Miller and Ager 2013) emphasizes the likelihood of ignition, potential fire behavior (intensity) and expected effects. The more recent one distinguishes between danger (ignition and propagation), exposure (the extent to which fire may spatially overlap with valued assets) and vulnerability (the potential losses and resilience to those losses) (Chuvieco et al. 2023). Both aim to guide fire management decisions but differ in structure and operational emphasis.

Based on these frameworks, various models have been developed to forecast the likelihood, spread, behavior and effects of wildfires (Rothermel 1972; Bar Massada et al. 2009; Mitsopoulos et al. 2015; Shadrin et al. 2024). Fire danger, in particular, concerns the estimation of the likelihood of fire ignition and the magnitude of its behavior (spread and intensity). Simulation-based approaches have been widely used to quantify fire likelihood and potential behavior, especially when observational data are scarce or when modeling large spatial extents. For example, Carmel et al. (2009) created high-resolution fire-likelihood maps for Israel using Monte Carlo simulations, and Bar Massada et al. (2009) did the same for Northern Wisconsin, United States (US), but accounted for both regular and extreme weather conditions. Moreover, Weise et al. (2010) showed that simulation frameworks can support fire danger estimation under limited meteorological data, as demonstrated in a case study from Hawai’i, US.

Generally, one of the challenges of fire danger (fire occurrence or risk) modeling is to decide which fire model is the most suitable, depending on the characteristics of the study region and data availability (Weinstein and Woodbury 2010). In turn, the model’s accuracy relies on the resolution and reliability of the input parameters. Common sources of input parameters for fire risk models are meteorological danger indices that use a numerical or qualitative indicator derived from a combination of meteorological variables. These indices increasingly rely on forecasted weather data (e.g. temperature, humidity, wind speed, precipitation, dew point), as observed data often arrive too late for operational use (Deeming 1972). In addition, remotely sensed data, thanks to their consistent spatial coverage across broad extents and their relatively frequent acquisition rate, are commonly used as well (Camia et al. 1999; Hamadeh et al. 2017; Szpakowski and Jensen 2019; Laneve et al. 2020). The latter is applicable almost anywhere, due to the global coverage of most satellite-based sensors, whereas the former is applicable only in regions where the relevant meteorological data are readily available. An additional important variable in many studies is live fuel moisture content (LFMC), which describes the physiological state of vegetation and its likelihood to ignite (Yebra et al. 2013; Kang et al. 2023). For example, in chaparral ecosystems, fire activity tends to increase sharply when LFMC drops below critical thresholds (Dennison and Moritz 2009). Recognizing the influence of such factors, a fire danger rating system is a management tool that is based on mechanistic models and utilizes meteorological danger indices (Schlobohm and Brain 2002). It is a broad-scale assessment that elaborates various fire danger factors - both constant and variable, which affect the initiation, spread and difficulty of wildfire control. These factors are combined into one or more qualitative or numeric indices that reflect current or predicted fire potential on an area (Deeming 1972; Schlobohm and Brain 2002). One prominent example of such a system is the National Fire-Danger Rating System (NFDRS) of the US. The NFDRS incorporates meteorological data along with regional fuel and topography information to produce various indices that guide fire management decisions in the US (Schlobohm and Brain 2002). Conceptually, the NFDRS aligns with the earlier fire risk framework (Miller and Ager 2013), which emphasizes the estimation of ignition probability, fire behavior and expected effects, rather than broader components such as exposure and vulnerability.

The National Fire-Danger Rating System (NFDRS)-1978

The NFDRS was first implemented nationally in 1972 in the US (Deeming 1972), and in 1978 the system was revised and reissued with improvements such as updated fuel models and refined meteorological relationships (Bradshaw et al. 1984). The 1988 version implemented additional modifications to improve its suitability for eastern US environments (Burgan 1988). Nowadays, this system is being used by most state and federal fire protection agencies in the US and in other countries. In Israel, it had been used until recently by the Israel Meteorological Service (IMS) for predicting fire occurrence risk in Israel. The NFDRS consists of three major parts: (1) the scientific basis, which relies on semi-empirical mathematical models to calculate fire danger; (2) user-controlled site descriptors, which specify the local conditions required to run the system; and (3) weather and non-weather data, which provide essential inputs for daily fire danger calculations (Schlobohm and Brain 2002). The second component of the NFDRS, the site descriptor, aims to represent the local conditions. Hence, the user needs to define and characterize the area examined. First of all, the fire danger rating area needs to be set (the spatial scale and boundaries). It needs to be relatively homogeneous (in terms of weather, topography and fuels). Then, the user specifies site characteristics: climate and slope class, annual precipitation, fuel models and live fuel type (Appendix 1) (Schlobohm and Brain 2002). Lastly, to calculate daily fire danger rating indices (and other components that are listed below), two forms of data are needed. These are daily weather observations and parameters that the user sets (Schlobohm and Brain 2002). Weather data consist of the following components: dry bulb temperature, relative humidity, dew point, wind speed and direction, state of the weather and solar radiation (appearing as codes in Appendix 2), precipitation amount and duration (which occurred within the preceding 24-h) and fuel wet flags (whether fine fuels are wet or not depending on rainfall and state of weather). Besides the above weather inputs, for a better representation of local site conditions, the user must periodically input observations and NFDRS parameters (Appendix 3). For instance, since 2017, the IMS has utilized six fuel-sticks to measure 10-h dead fuel moisture content values in six different meteorological stations across Israel to enhance the accuracy of wildfire hazard alerts (Shmuel et al. 2022). After setting all parameters and variables, the system output generates four main fire danger indices and components (Schlobohm and Brain 2002) (Table 1), though only two, Burning Index (BI) and Ignition Component (IC), are calculated by the IMS.

Several studies examined the efficacy and limitations of the NFDRS in the US. A comprehensive evaluation of NFDRS indices across 5 years in the Western US underscored their effectiveness in predicting fire danger (Roads et al. 2005). These indices – namely, the IC, Energy Release Component (ERC), BI, Spread Component (SC) and Keetch–Byram Drought Index (KB) – quantify fire danger based on the combined effects of fuel, topography and weather conditions. The BI exhibited strong predictive capability in estimating fire danger levels by capturing meteorological conditions that influence fire behavior, particularly the difficulty of fire control. The predictive capacity was particularly strong at seasonal timescales, but also performed well on a weekly basis. However, while these indices (including BI) are useful for fire management planning by capturing fire danger trends and aiding in preparedness and resource allocation, the study identified a weak correlation between them and observed fire characteristics such as fire counts and burned areas. Additionally, this study suggested that the accuracy of these forecasts could be improved by incorporating more detailed meteorological and ecological data. Another study (Freeborn et al. 2015) further emphasized the importance of temporal scaling and local calibration in the use of NFDRS indices. Their decade-long study in the Northern Rockies, US, illustrated that the predictive accuracy of these indices significantly depends on the temporal context – shorter averaging periods revealed stronger correlations with actual fire behavior. This finding suggests that to make NFDRS indices more effective and applicable across different areas, they should be customized not just to the specific regional characteristics but also adjusted to match the local timing and patterns of fire behavior. A further analysis (Walding et al. 2018) which examined the performance of the NFDRS over an 8-year period across the contiguous US, aligns with the findings of Roads et al. (2005), by highlighting regional variability in the performance of the NFDRS. Their research indicated that while NFDRS indices were predictive of the size of fires, especially in regions with consistent fire management practices, there were significant discrepancies in their correlation with fire occurrence and size in southern and eastern US, suggesting a need for regional adaptations of the NFDRS. They recommended enhancements to the NFDRS that incorporate region-specific data to address these inconsistencies and improve overall fire danger forecasting.

While the NFDRS has been widely adopted across the US, its applicability in non-US ecosystems remains uncertain without region-specific adaptation. To date, several adaptations of the NFDRS have been implemented and tested in non-US ecosystems, such as its use in South Africa’s fynbos vegetation (Van Wilgen 1984; Van Wilgen and Burgan 1984). These efforts highlighted the importance of tailoring fuel models and danger classification thresholds to local conditions, as differences in vegetation moisture and fire behavior may persist despite superficial structural or climatic similarities (e.g. between fynbos and chaparral shrublands vegetation). Further comparative studies of major fire danger rating systems underscore this point: a comparison of the NFDRS, Canadian Forest Fire Danger Rating System (CFFDRS) and McArthur’s system in China found that the NFDRS performed well for surface fire prediction but required parameter adjustments for crown fire forecasting due to its theoretical limitations (Tian et al. 2005). These findings suggest that the NFDRS was originally designed to estimate potential fire behavior, raising important considerations about its use beyond intended contexts. While its structure allows for broad applicability, misuse – such as interpreting it as a predictive tool for specific fire outcomes – may lead to misleading conclusions or inappropriate management decisions.

Despite its broad applicability and solid mechanistic foundations that make it generalizable to many different ecosystems, it is currently unclear whether a fire danger rating system that was developed for the plethora of ecosystems in the US can provide reliable forecasts of fire danger in other regions with different environmental conditions without considerable re-parameterization. Specifically, the direct application of the NFDRS by the IMS raises the need for robust testing and validation of this system in Israel’s Mediterranean ecosystems (despite the presence of Mediterranean-type ecosystems in California, US). Testing the effectiveness of the NFDRS in capturing fire characteristics in a non-US region such as Israel, may open the door to future utilization of this system in other countries that currently lack an operational system for fire risk assessment. Our study was motivated by this opportunity.

Research aim

Using retrospective analysis, we aimed to test the effectiveness of the BI derived from the NFDRS-1978 for predicting fire characteristics in non-US Mediterranean-type ecosystems, which are present on multiple continents. We were specifically interested in these ecosystems because they are characterized by significant fire activity during the dry season. As a case study, we examined the use of the fire risk index used by the IMS, focusing on how well it explains the characteristics of wildfires in Israel, such as the area burned and fire duration-measured as the time elapsed from the arrival of the first fire truck to the departure of the last, for 2015–2023 (Guk et al. 2023). In addition, we aimed to build an improved predictive model of wildfire characteristics, by combining the meteorological fire risk index with ancillary data.

As part of this study, we tested the following hypotheses:

Study area

Israel has a large Mediterranean region which is characterized by significant fire activity during the dry season (on average, 285 fires larger than 10 ha per year between 2015 and 2022) (Guk et al. 2023). Climate change is expected to increase the frequency, intensity and extent of wildfires as a result of projected prolonged droughts, increased water evaporation and a higher frequency of intense heat waves (Israel Ministry of Environmental Protection 2023, retrieved from https://www.gov.il/BlobFolder/reports/1st-step-mapping/he/climate_change_and_energy_efficiency_1st-step-mapping.pdf). While climate is a major driver of fire activity, human-caused ignitions, such as arson or careless behavior, also significantly contribute to wildfire occurrences in Israel (Levin et al. 2016). Natural ignitions are rare in Israel; only six cases of lightning-induced fires have ever been documented (Paz et al. 2016).

The wildfire season in Israel coincides with the dry season, which takes place between mid-April and November (Levin and Saaroni 1999). However, the most deadly fire in Israel occurred on 2 December 2010 on Mount Carmel (Shavit et al. 2013). Most fires occur in the summer, but the largest fires (in terms of area) tend to occur in transition seasons (Levin and Heimowitz 2012). During this period, the synoptic systems that cause major forest fires are the North African (‘Sharav’) cyclone and the Red Sea trough system, as they both carry hot and dry air (Levin and Saaroni 1999). There is an especially high probability of large fires in the fall when fuel moisture is especially low after nearly 6 months without precipitation (Bar-Massada and Lebrija-Trejos 2021).

Wildfires in Israel are typically small, with the majority burning less than 10 ha, and are of short duration, often lasting less than a day. This is due to the relatively small and fragmented patches of forest, woodland and shrubland in the Mediterranean part of the country, which cover about 4300 km² (Ben-Moshe and Renan 2022). Temporally, larger burned areas in Israel are associated with rainy years, as precipitation enhances herbaceous vegetation growth, while herbaceous species are the first to dry out at the end of the rainy season (Turco et al. 2017), thereby increasing the coverage of flash fuels. Also, planted pine forests (where recreational activity is common) and military training zones tend to be associated with more frequent fire events (Levin and Heimowitz 2012).

Parameters and components of the fire danger rating as used by the Israel Meteorological Service

The IMS provides the Israel Fire and Rescue Authority as well as other emergency forces with a daily forecast for fire danger in Israel from the NFDRS’ output. The forecast is derived from two indices: the BI and the IC (https://ims.gov.il/en/FireIndexExplanation). From 2010 until recently, these NFDRS-based indices were computed at a national scale in Israel, with a spatial resolution of 1 km and 24 hourly values per day. For the calculation of fire danger indices, the whole study area is characterized by the IMS as herbaceous vegetation (despite the presence of many woodland ecosystems), containing various slope classes, and divided into two climate classes: class 1 (arid) and class 2 (semi-humid). Although later versions of the NFDRS (e.g. 1988 and 2016) introduced modifications and refinements, at the time we conducted our study, the IMS used the 1978 version operationally. Therefore, this study is based on the outputs of the NFDRS-1978, as implemented by the IMS.

Wildfire dataset

The national wildfire database of Israel was generated in a previous study (Guk et al. 2023), and includes 2276 wildfires larger than 10 ha (where 1 ha = 100 m × 100 m) with a total burned area of 142,911 ha between 2015 and 2022. We expanded this dataset and mapped the wildfires for 2023 using the same approach. In 2023, we included 128 wildfires with exact dates and 61 wildfires with recorded durations out of a total of 243 wildfires. To test the IMS fire risk indices, we used 1788 wildfires for which we had an exact occurrence date (Guk et al. 2023). For these wildfires, additional information on fire characteristics and landcover type was extracted from remote sensing and ancillary databases (Table 2). Given the unique context of some of the wildfires in Israel (Besenyő 2019; Levin et al. 2021), some of which result from military training or intentional arson, we tagged these specific cases within the dataset for later exclusion from some of the analyses to avoid introducing bias from non-representative human-caused fires. While the exact causes of all wildfires are not available, fires attributed to intentional arson (e.g. kites) or rockets (Takshe et al. 2012; Zych 2019) were identified by mapping wildfire occurrences near borders during periods of conflict.

Live fuel moisture content

An additional variable that could enhance the prediction of wildfire characteristics is LFMC. A study by Guk et al. (2024) compared three LFMC models (Yebra et al. 2018; Quan et al. 2021; Cunill Camprubí et al. 2022) and determined that the Australian model was the most suitable for Israel’s Mediterranean region. This model utilized a national land cover dataset comprising three vegetation classes (woodland, grassland and shrubland) to enhance the accuracy of LFMC mapping. Hence, we utilized LFMC data at 500 m resolution derived from this Australian MODerate Resolution Imaging Spectrometer (MODIS)-derived physical model for our analysis. For each fire polygon, we derived two LFMC variables: first, the mean LFMC value 8 days prior to the wildfire, and second, the delta LFMC value, which represents the change in LFMC from 16 to 8 days before the wildfire. These time frames were chosen to capture the dynamic nature of LFMC changes, reflecting the rapid drying of vegetation and its transition into a more flammable state. We selected these specific periods because the MODIS reflectance data are based on a 16-day composite, which aligns with the temporal resolution of the model (Guk et al. 2024).

Examining the correspondence between fire danger indices and wildfire characteristics at different temporal scales

There was a strong correlation between BI and IC at a national spatial scale for all days in 2015–2023 (n = 2181; Spearman’s ρ = 0.92, P < 0.001). This correlation was computed solely to assess redundancy between the indices, and not to model their temporal relationship. We therefore included only BI in all subsequent analyses. We calculated the mean value of the daily BI index across all Mediterranean areas in Israel and the West Bank (Palestinian authority) at five temporal scales: event-based, daily, weekly, monthly and yearly (April–November, 2015–2023) and at two spatial scales: the entire study region (Fig. 1) and within the boundaries of individual wildfire events. For each wildfire, we calculated the mean BI value on the same day of the event by averaging the BI values of all grid cells (1 km resolution) intersecting with the burn scar polygon. Each fire danger index used here represents the daily maximum over the last 24 h. Then, we calculated Spearman’s correlation coefficient between wildfire characteristics (burned area and wildfire duration) and the fire danger index across different temporal scales (daily, weekly, monthly and yearly). To account for possible effects of fire size on these correlations (which are related to lower accuracy of ignition-date estimates for smaller fires), we set two threshold filters on wildfire size: 300 and 500 ha. Wildfires without a precise occurrence date were included in our analysis only at weekly, monthly and yearly scales. We further analyzed the Spearman’s rank correlation coefficient between BI and wildfire characteristics for four vegetation groups: Woodland, Shrubland, Herbaceous and Agriculture. Finally, we used quantile regression, implemented in XLStat (www.xlstat.com), to explore relationships across different wildfire size groups, focusing on quantiles less than or equal to 0.75. Higher quantiles (e.g. 0.9, 0.95 and 0.99) were excluded as the residuals were not randomly distributed around zero, potentially compromising the validity of the regression results.

Fig. 1.

Land cover map (2019–2020, Israel Nature and Parks Authority) and wildfires burned scars (2015–2023) of the study area within the Mediterranean portion of Israel and the West Bank (Palestinian Authority).

Predictive models of wildfire characteristics

To explore the retrospective relationship between mean BI and wildfire characteristics at the individual fire level, we built Generalized Linear Models (GLMs) with an exponential response distribution to predict two wildfire characteristics: burned area and wildfire duration. Predictor variables included fire risk index (BI), vegetation type (expressed as the percentage of woody and herbaceous cover within the burned area) and vegetation continuity (within a 3 km buffer around the wildfire centroid), to represent the potential of the area that can be burned (Guk et al. 2024). Additional predictors included two LFMC variables (Guk et al. 2024), a binary variable indicating whether the fire occurred inside or outside a military training zone, distance to the nearest city and the distance to the country’s border (Table 2). All continuous predictor variables were standardized to a zero mean and unit variance prior to model fit to facilitate comparability of effect sizes. The binary variable was retained in its original form (0 or 1) to preserve interpretability. Additionally, to account for potential threshold responses of fire characteristics, we re-ran the models using filtered subsets of the data, focusing on specific wildfire size thresholds and geographic locations.

Time series and variability of the fire risk index

The time series of the BI and total burned area from wildfires at the national level, represented at a daily scale, are presented in Fig. 2. Despite some fluctuations in the BI index, larger burned areas were not consistently observed on days with higher BI values and the correlation between BI and burned area was low (Spearman’s ρ = 0.18). However, both the median and minimum sizes of burned areas were generally greater in the higher BI groups (Fig. 3). When analyzing the distribution of distinct calendar days across different BI bins, categorized by wildfire size, days without wildfires predominantly occurred at lower BI values (below 60) mainly in April, October and November, while larger wildfires were increasingly observed on days with higher BI (Fig. 4). The histogram in Fig. 4 underscores the tendency for larger wildfires to occur during periods with elevated fire risk, though smaller wildfires also occur at higher BI values.

Fig. 2.

Mean of daily national maximum Burned Index (BI) value and total burned area size (ha) by day at the national scale, for nine fire seasons between the years 2015 and 2023.

Fig. 3.

Box plot of burned area scars between 2015 and 2023 (n = 1697) in logarithmic scale, categorized by groups of National Fire-Danger Rating System’s Burned Index (BI). Each wildfire was assigned to the mean maximum index value by date and spatial intersect.

Fig. 4.

Histogram showing the frequency of unique days grouped by Burning Index (BI) bins for various wildfire size categories. Unique days refer to distinct days on which wildfires of different sizes occurred. The categories include days with no wildfires, wildfires smaller than 100 ha, wildfires between 100 and 300 ha, between 300 and 500 ha, and wildfires larger than 500 ha. Days with no wildfires include only those days not covered even by the date ranges of wildfires with unknown exact occurrence dates in our complete database. For the no-wildfires category, the BI represents the average BI across the entire study region.

Correlations between wildfire characteristics and BI across spatial, temporal scales and vegetation groups

In general, Spearman correlation coefficients (Fig. 5) between BI and wildfire characteristics (maximum, total and average burned area, as well as wildfire mean duration) were low to moderate and varied across temporal resolutions. The highest statistically significant correlation was observed at the monthly scale, when all wildfires were included, between maximum burned area and BI (Spearman’s ρ = 0.41). Notably, this correlation did not increase for larger fires. Only the weekly correlation for wildfires greater than 500 ha was higher than for other wildfire size groups at the same temporal scale. Furthermore, the correlation between wildfire duration and BI remained generally low across all temporal scales and wildfire sizes. An exception was wildfires larger than 300 ha, which had a Spearman correlation of 0.3 between BI and wildfire duration on a monthly scale. Across all wildfire size groups, daily correlations between BI and wildfire characteristics were consistently the weakest. These results suggest that BI was a better predictor of wildfire behavior over longer time frames.

Fig. 5.

Spearman correlation coefficient matrix between mean national Burning Index and wildfire characteristics: maximum burned area (of a single wildfire within that week/month/day/year), average burned area, total burned area and average wildfire duration. Correlations are analyzed at four temporal scales: yearly, monthly, weekly and daily, as well as across three wildfire groups (all wildfires [>10 ha], those larger than 300 ha, and those larger than 500 ha). Asterisks represent statistically significant correlations (*P < 0.05; **P < 0.01; ***P < 0.0001).

We further examined the correlation between BI and wildfire characteristics across vegetation types at the wildfire event scale. A weak but positive correlation was found between mean BI and wildfire characteristics (burned area and wildfire duration), based on Spearman’s rank correlation coefficient (ρ), calculated separately for each vegetation type. For burned area, ρ values ranged from 0.01 (Woodland) to 0.09 (Agriculture) to 0.2 (Herbaceous) and 0.27 (Shrubland), with statistically significant correlations observed in Herbaceous (P < 0.05) and Shrubland (P < 0.05). In contrast, the relationships were not statistically significant for Woodland (P = 0.87) or Agriculture (P = 0.35). For wildfire duration, correlations were uniformly low and ranged from 0 (Woodland) to 0.09 (Shrubland), and none were statistically significant. When quantile regression was applied without separating vegetation types, the highest R² value for the relationship between BI and burned area (as the response variable) was 0.27 (statistically significant at the 0.75 quantile). In contrast, the variation in wildfire duration (as the response variable) was not well-explained by BI (R² values below 0.01). When the data were separated into four vegetation groups, the highest R² values (based on the quantile regression) for burned area were observed in Shrubland (R² = 0.41) and Herbaceous (R² = 0.42) at the 0.75 quantile, both statistically significant (Fig. 6). In comparison, R² values for Woodland and Agriculture remained very weak across all quantiles. Additional R² values for lower quantiles (0.1, 0.25 and 0.5) are presented in Appendix 4. For wildfire duration, all vegetation groups and quantiles had very low explained variances.

Fig. 6.

Scatter plot of mean Burning Index versus burned area (logarithmic scale) for individual wildfire events, categorized by four vegetation types: Agriculture, Herbaceous, Shrubs and Woodland. The regression line represents quantile regression (τ = 0.75).

Models of wildfire characteristics

The variation in burned area was weakly explained (R² = 0.32) by a GLM that included the following explanatory variables: Military training zones (binary), ΔLFMC 16-8, LFMC week before a wildfire, distance to border, distance to nearest locality, % Woodland cover, % Herbaceous cover, mean BI and % continuous vegetation within 3 km (Fig. 7). BI and the percentage of continuous vegetation within 3 km emerged as the strongest predictors of area burned, and both had significant positive effects (Fig. 7). Increasing the threshold for burned area to include only larger wildfires did not enhance the model’s performance. However, further analysis showed that when wildfires inside military training zones and wildfires caused by arson were excluded, the amount of explained variation in the burned area model improved slightly to 36% (Fig. 7). Furthermore, the highest explained variance in burned area (R² = 0.5) was achieved when a threshold was set to include only wildfires with woodland cover higher than 0. This exclusion removed pure grassland wildfires (typically rangelands and agricultural fields) from the analysis (Fig. 7). In contrast, the GLM for wildfire duration, using the same explanatory variables, had virtually no explanatory power (R² ~ 0).

Fig. 7.

Effect sizes of different variables in three generalized linear models predicting burned area at the level of individual wildfires (2015–2023), using an exponential response. Model 1: all wildfires (n = 1644). Model 2: wildfires outside military training zones and not classified as intentional arson (n = 987). Model 3: wildfires outside military training zones, not classified as intentional arson and with woodland cover greater than 0 (n = 214). All explanatory variables were standardized (z-scores) prior to model fit.

Effectiveness of NFDRS in Israel’s Mediterranean ecosystems

A retrospective analysis of the BI using the wildfire database revealed moderate correlations with burned area, particularly at coarser temporal scales such as monthly and yearly, while correlations with wildfire duration remained relatively weak across all timescales. These findings support our first hypothesis that BI is not effective for predicting wildfire characteristics at finer temporal resolutions (daily and event-based) and for smaller wildfires. This result also aligns with previous findings from the US, where the NFDRS performed best on seasonal timescales but showed limited correlation with fire occurrence and size at finer temporal scales (Roads et al. 2005; Walding et al. 2018). However, we caution that these statistical analyses do not account for firefighting efforts, which directly influence the size and duration of the wildfire (Plucinski 2019). The unaccounted effect of suppression is a fundamental limitation in any attempt to find a relationship between wildfire characteristics and risk indices, as the absence of a variable for firefighting efforts in the models leaves a crucial explanatory factor unaddressed. It is possible that heightened alertness and rapid response by fire brigades prevented fires from spreading, thereby limiting their growth. However, data on firefighting efforts were not available to us.

In Israel, the strongest observed correlation between BI and a fire characteristic was between the monthly BI and maximum burned area (ρ = 0.41 for burned area >10 ha). However, this correlation did not improve significantly for larger wildfires, contrary to our second hypothesis. Although the correlation between BI and fire characteristics improved at longer temporal scales, relying on BI for long-term assessments remains problematic from a practical standpoint. If the index lacks a meaningful relationship with wildfire occurrence on a daily scale, its relevance for real world applications at monthly or yearly scales is also questionable. The limited explanatory power of the BI at finer scales may be attributed to the unique characteristics of Israel’s wildfire regime, where small, short-duration fires dominate (Levin et al. 2016). Additionally, human activities, including military training and intentional arson, play a significant role in ignition patterns, with 738 out of 1788 wildfires in our database attributed to these causes: 130 classified as intentional arson and 608 occurring within military training areas.

Furthermore, the classification of Israel’s diverse vegetation types as herbaceous by the IMS likely contributed to reduced model precision. When we explored the correlation between BI and burned area and duration of wildfires across different vegetation classes, no group demonstrated significantly better performance. However, quantile regression revealed that the predictive power of BI varies by quantile and vegetation type. For burned area, the highest correlations were observed at the 0.75 quantile in Shrubland (R² = 0.41) and Herbaceous vegetation (R² = 0.42), indicating moderate predictive power for larger fires in these classes. In contrast, Woodland and Agriculture consistently showed weak correlations across all quantiles. For wildfire duration, quantile regression confirmed universally weak correlations across all vegetation groups and quantiles, reinforcing the conclusion that BI is not a reliable predictor of wildfire duration, regardless of fire size or vegetation type.

Improving predictive models with ancillary data

Our results showed that incorporating vegetation and anthropogenic variables alongside BI enhanced the retrospective predictive ability of burned area models. Vegetation continuity within a 3 km radius and BI emerged as the strongest predictors of burned area, supporting our third hypothesis. However, the overall explained variance of these models (32%) remained relatively low. Excluding wildfires in military training areas and intentional arson fires (Besenyő 2019; Salvoldi et al. 2020; Perry and Amram 2024) slightly improved model performance (R² = 0.36), underscoring the importance of filtering data to account for context-specific factors. The highest explained variance (50%) was achieved when examining only wildfires that occurred outside pure grasslands. This suggests that using context-specific adjustments to fire risk models, which in Israel include the incorporation of proximity to military zones or historical patterns of human-ignited fires, is essential for ensuring satisfactory model performance. The importance of local characteristics like the ones we found here, underscores the need to account for regionally-important characteristics of fire regimes when trying to adapt a fire risk system which was developed for another region with considerably different driving forces of wildfire activity.

In contrast to the moderate performance of the burned area models, the predictive ability of models of fire duration was consistently weaker, with near-zero explained variance. Possibly, this weak performance was due to the higher complexity of this wildfire characteristic compared to the burned area. Fire duration depends on various interacting factors, including suppression efforts (which in turn depend on accessibility to the fire perimeter and on the amount and capabilities of firefighting forces), fuel type and weather conditions (which together affect fire behavior, affecting the ease of suppression), and is often less accurately recorded due to manual field reporting. Generally, modifications to incorporate locally calibrated thresholds and region-specific parameters, such as precipitation trends and vegetation phenology, may improve daily forecasts of fire risk.

While this study focused on the BI due to its operational use by the IMS, future research could benefit from exploring the individual components of the NFDRS, such as the SC or ERC. These components may offer more direct insight into the physical drivers of fire size. However, such an analysis was not feasible in our case, as only BI and IC are currently calculated by the IMS.

Beyond these data-related limitations, it is also important to acknowledge the conceptual constraints of the NFDRS itself. While our analysis examined relationships between the BI and characteristics of fires that had already occurred (e.g. burned area and duration), the NFDRS was not originally designed to predict specific fire outcomes. Rather, like its Canadian and Australian counterparts, the NFDRS was developed to estimate potential fire behavior assuming ignition (Deeming 1972; Noble et al. 1980; Stocks et al. 1989). Applying the NFDRS beyond this purpose may therefore lead to misinterpretation. Instead, we suggest a practical adaptation: local calibration of danger class thresholds to better reflect observed fire behavior in specific regions. Although we did not conduct such calibration in this study, our findings indicate its potential to improve the relevance of NFDRS outputs in Israel’s Mediterranean context. Indeed, as our findings show, the BI correlates moderately with burned area at coarse scales, especially for certain vegetation types and quantiles, suggesting that calibrated thresholds may allow managers to interpret fire danger outputs in ways that better align with local realities.