Fuel-aware forest fire danger rating system RISICO: a comparative study for Italy

RISICO system

The model’s development started in the late 1990s with research funded by the Liguria Region (Italy) to provide a regional FFDR system,¹ subsequently extended at a national level and adopted by the Italian Civil Protection Department (Dipartimento di Protezione Civile - DPC) to support its activities.² This research led to the development of the RISICO system (RISICO stands for RISchio Incendi e COordinamento, i.e. wildfire risk and coordination) (Fiorucci et al. 2008; Fiorucci et al. 2018), made available for the whole Italian territory in 2003, and for all the Italian regional Civil Protection (CP) authorities from 2011 through the MyDewetra platform (Italian Civil Protection Department and CIMA Research Foundation 2014). The system has been developed and maintained by CIMA Research Foundation,³ and updated over the years based on operational use feedback and technological advancements. This paper presents the latest version of the model.

The model’s objectives were identified thanks to interaction with the wildfire risk managers of the Liguria Region first, and at the national level later. The model aims to determine potential wildfire behavior through the whole territory and during the day, independently from possible ignition causes. From an operational perspective, it is of paramount importance to identify the areas and times of increased fire danger where eventual ignitions could lead to extreme wildfire behavior. This information can trigger early prevention actions such as patrolling and alerts to the population, and support the preparedness phase early, preparing firefighting teams for action, ready to start in case of emergency. Given the importance of the temporal component in scaling up wildfire behavior, it is a priority to reduce the time to intervention and make firefighting actions effective (Taylor et al. 2013). For the same reason, RISICO needed to provide a forecast of wildfire danger for the next 24–48 h, supporting the issuance of wildfire danger bulletins. During RISICO’s development, wildfire research focused on assessing fire danger via the FWI (Van Wagner and Pickett 1985; Van Wagner 1987) spatializing the FWI values computed at each weather station across the territory (Di Giuseppe et al. 2016). Italy’s early adoption of weather forecasts for flood prediction in the 1990s enabled RISICO to integrate from the beginning forecasts from Limited Area Models, allowing fire danger forecasts, now a standard practice.

Although born from independent research with respect to the FWI, the RISICO model has been modified over the years by adopting some concepts from it. However, some important particularities persist, such as the hourly resolution of the information, and the use of both topography and fuel cover. Any model is deeply bounded by the objectives for which it was developed, and so by the context in which it was conceptualized. These choices were driven by the Liguria Region particularities in terms of rapid changes in weather conditions, high vegetational variability and complex topography (Tonini et al. 2020), but similar conditions are found in many Italian and Mediterranean regions, making RISICO an interesting example.

Sub-daily resolution also enables better adaptability to diverse operational needs across the various phases of the wildfire risk management cycle, allowing the subsequent aggregation of information based on specific operational requirements. For instance, the hourly data provided by RISICO are aggregated into daily indices to avoid information overload, using statistical analysis of the hourly values. These daily indices are then used to produce fire danger bulletins. In particular, averages calculated over different percentiles are used to detect the persistence of fire danger patterns throughout the day – typically, the 75th percentile is used to capture persistence over a quarter of the day. However, daily extreme values are also provided for further assessment.

The RISICO system is composed of indices (see Fig. 1) related to different aspects of potential wildfire behavior: Dead Fuel Moisture Content (DFMC, %), Rate of Spread (ROS, m h⁻¹) and Fire-Line Intensity (FLI, kW m⁻¹). Neither the ROS nor FLI index aims to measure wildfire spread or intensity precisely but instead they offer tools to differentiate among potential wildfire behaviors. The indices are computed pixel by pixel by dividing the territory into cells, in which topography and fuel are considered homogeneous.

Fig. 1.

The RISICO schema, highlighting the static (fuel, topography) and dynamic (weather conditions) inputs and their role in the model’s modules: DFMC (Dead Fuel Moisture Content) module; ROS (Rate of Spread) module; FLI (Fire-Line Intensity) module. LHV, Lower Heating Value.

Dead Fuel Moisture Content (DFMC) index

The dead fuel moisture represents a crucial component in many FFDR systems (Nelson 2000), affecting both ignition probability and fire spread (Silva et al. 2010). It constitutes the dynamic component of RISICO, and is updated at each step of the model given previous fuel moisture status and weather.

The original empirical model was first proposed in Fiorucci et al. (2008). In the case of no rain, the fuel moisture tends toward an Equilibrium Moisture Content (EMC) with changing rate depending on fuel type through a standard response time T₀ (h), which varies based on weather conditions. In the case of rain, the fuel moisture increases until it reaches a fuel-dependent saturation level sat (%), with a changing rate depending on the amount of rain.

The model proposed was able to differentiate between slow and fast fuel moisture dynamics in a single index, discriminating between faster (e.g. grasslands) and slower (e.g. broadleaf litter) responses to weather changes. Also, it adapted to different temporal steps given its explicit dependence on time in the equations.

The inclusion of different fuel-dependent moisture dynamics into a single index and the possibility to adapt to different time steps has been retained in the last release of the model (Perello et al. 2025), which has been modified and calibrated using fuel stick data. The model showed good performance, especially for low fuel moisture values. See Perello et al. (2025) for more details. The model’s implementation in RISICO adopts different T₀ and sat parameters according to fuel maps, as presented in the next section (see Table 1). Also, snow cover can be considered by setting the fuel moisture values to the saturation level in the case of snow, letting DFMC evolve once it melts.

Rate of Spread (ROS) index

The ROS index provides information on the fire’s potential to spread rapidly and, therefore, directly influences the timing and effectiveness of the response to the event. The ROS model modifies the fuel-specific nominal rate of spread v₀ (m h⁻¹) (see Table 1) according to the fuel moisture content, topography and wind:

where W and M are the wind–topography and fuel moisture effects, respectively.

The wind–topography interaction W was developed from the wildfire propagation model PROPAGATOR (Trucchia et al. 2020; Perello et al. 2024), testing the propagation process in any potential directions, and by selecting the maximum contribution as the worst-case scenario:

where V_s and V_w respectively are the adimensional slope and wind contributions, while θ is the propagation direction. See Trucchia et al. (2020) and the Supplementary material for more details – a graphic description of the W contribution can be found in Fig. 2a.

Fig. 2.

The wind–topography effect (a), and dead fuel moisture content effect (b) on the Rate of Spread index (ROS). In particular, in (a), the combined effect of wind direction and aspect is reported by fixing an aspect value (i.e. aspect 0°, black arrow in the central plot) and varying wind direction (blue arrows in the central plot) – the different plots represent different wind directions; in each, the combined effect of wind speed and slope contributions is highlighted.

The fuel moisture contribution is given by the formula:

where x is the DFMC expressed as a fraction between 0 and 1. The M index formulation has been tested over different settings, decreasing toward zero the ROS index for DFMC values greater than 40% (as tested in previous RISICO releases) and by increasing ν₀ for very dry conditions (i.e. less than 5%) – see Fig. 2b.

Fire-Line Intensity (FLI) index

The FLI index highlights areas and times where fire intensity may hinder suppression efforts. It is computed based on the ROS and the fuel moisture content, as per Byram (1959):

where d (kg m⁻²) represents the fuel load and lhv the lower heating value (kJ kg⁻¹). The model distinguishes two main fuel levels: the surface dead fuel component (identified by the index i = 0) and the live fuel and canopy component (identified by the index i = 1). This choice was driven by the need for simplification in the original development of the RISICO model and has been maintained; however, further subdivisions may be considered in future developments of the model. The lhv is computed according to the equation:

The higher heating value hhv (kJ kg⁻¹) is defined for each fuel type, while u_i refers respectively to the dead (i.e. the DFMC index) and the live fuel moisture content – the latter is considered fixed as the worst-case scenario for each fuel type. Despite this strong assumption, RISICO was designed to minimize data requirements, which were limited at the time. However, satellite-derived vegetation indices now offer proxies for live fuel moisture and will be integrated into future model updates. See Table 1 for parameter values.

RISICO fuel map

The complexity and heterogeneity of the fuel, the required spatial extent (e.g. regions or states) and data maintenance make data collection and management difficult, and this remains a major challenge in forest fire research (Abdollahi and Yebra 2023; Aragoneses et al. 2023). Therefore, the development and updates of the fuel map, crucial in RISICO, have been a key challenges. To support CP tasks, RISICO adopts a simplified fuel classification, assuming homogeneous fuel characteristics within types based on worst-case scenarios.

From the first release of the system, fuel classification was based on expert-based classification of land use classes, dividing them according to different expected wildfire behaviors as they have varying propensities for wildfire (see Fig. 3). This simplified and coarse fuel classification enhanced RISICO’s discrimination capabilities, as observed through many years of operational use.

Fig. 3.

The percentage of territory covered by different land use categories with respect to their percentage that experiences wildfires. The analysis covered the period 2007–2022. CORINE Land Cover maps were selected following their availability for the analyzed period (i.e. CORINE 2006; CORINE 2012; CORINE 2018 from SNPA (2024)). The burned areas considered correspond to the official burned areas presented in the paper (see Material and methods section). The land cover categories analyzed are limited to the burnable ones that cover more than 0.5% of the national territory.

The same idea persists in the last release by dividing land cover categories into four general vegetation classes: Grassland and croplands (G), where surface fire behavior is expected; Broadleaves (B), where fire is less likely to turn into a crown fire; Shrubs (S), where bushfire can evolve into high-intensity behavior; and Conifers (C), where very high intensity fires can occur owing to crown fire transition and/or spotting. Class C can be also considered for species that exhibit similar extreme fire behavior to conifers, such as eucalyptus plantations – see the Discussion section (Ganteaume et al. 2021; Toy-Opazo et al. 2024). The map used for the present analysis was defined based on national CORINE 2018 IV level (SNPA 2024) – see Supplementary material for details. To reduce expert-based assessment, as done in previous releases, wildfire susceptibility (Tonini et al. 2020) has been added, intended as the ‘probability that fire occurs in a specific area without considering a temporal scale, assessed on the basis of predisposing factors related to terrain’s intrinsic characteristics’, from Leuenberger et al. (2018). Reflecting a territory’s predisposition to fire occurrence, it further enhances RISICO’s discrimination by dividing the four vegetation classes into low, medium and high susceptibility, resulting in 12 fuel classes (see Table 1).

The fuel types are finally converted into fuel models with specific parameters shown in Table 1. The DFMC and FLI fuel parameters depend mainly on four vegetation classes whereas susceptibility adjusts the v₀ parameter, reducing it by 10% and further 30% for medium and low susceptibility classes, respectively.

The values associated with the different parameters come from years of optimization of the RISICO system, and so they were adopted in the latest RISICO release. The optimization of these parameters lies beyond the scope of this paper.

The susceptibility map was obtained through the methodology proposed by Tonini et al. (2020), Trucchia et al. (2022) and Trucchia et al. (2023), adopting the Machine Learning (ML) Random Forest (RF, Breiman 2001) algorithm. A 100 m resolution pan-European susceptibility map was produced exploiting the EFFIS (European Forest Fire Information System) (San-Miguel-Ayanz et al. 2012) wildfire database from 2008 to 2022 to overcome the limited extent of the time span for susceptibility analysis (https://forest-fire.emergency.copernicus.eu). Different explanatory variables were considered, namely land cover, topography and climatic information listed in Table 2. The procedure for susceptibility analysis is detailed in the Supplementary material.

To retrieve the susceptibility classes needed for fuel mapping in Italy, the 25th and 75th percentiles at the national level were adopted to identify low and high classes, respectively. This type of classification aims to narrow down the areas where resources should be prioritized for operational purposes at the national level, highlighting the differences between wildfire spreading potentials (see Results section).

Different modeling choices can be made for further differentiation within a territory – this aspect, however, lies beyond the scope of this paper.

Model performance assessment

RISICO was compared with FFDR indices from the literature (see Table 3) – a description of the different indices can be found in Swiss Federal Institute of Forest, Snow and Landscape Research WSL (2012), from which the indices were retrieved. The comparison analysis used the CHAPTER weather dataset, derived from ERA5 via Weather Research and Forecasting (WRF) model downscaling, considering ~3 km spatial and 1-h temporal resolution (Tartaglione et al. 2024). A reanalysis was used to reduce reliance of index performance on forecast skills (Di Giuseppe et al. 2016). RISICO incorporates topography and fuel, whereas most indices rely solely on weather data and possibly on some climate information calculated from CHAPTER. The analysis was conducted at a 1 km spatial resolution, consistent with RISICO’s operational implementation, rescaling the fuel map and adopting the most representative value in each cell. Topographic information (i.e. slope and aspect) was obtained from the Multi-Error-Removed Improved Terrain digital elevation model (MERIT DEM; Yamazaki et al. 2017), reported at 1 km resolution. The temporal resolution was daily, based on wildfire finer temporal information, i.e. ignition dates. Hourly indices were aggregated to a daily scale by averaging values above the 75th percentile or below the 25th percentile for low-value fire-related indices such as RISICO-DFMC, Angstrom, Sharples Fuel Moisture Index (FMI), Orieux Water Reserve (WR) – this method follows an established methodology from operational RISICO implementation (see RISICO system section). Although other percentiles could be used, this choice ensures consistency with standard RISICO practice across all hourly indices, leaving refinements for future studies.

All indices were implemented in Rust and made openly available on GitHub (see Data availability section). The indices were run starting from 2006 to allow for a 1-year transition period, conducting the analysis for 2007–2022 using Italian wildfire data.

Since the early 2000s, Italy has been collecting burned area data through ground-retrieved polygons at the regional level.⁴ A consistent dataset of 102,904 wildfires from 2007 to 2022, including ignition dates, was used for the analysis in this paper.

Italy experiences an average of 79,781 ha burned and 6432 wildfires annually, with extreme wildfire years occurring every 4–5 years (Fig. 4d). Wildfires peak in two seasons: winter (November–April) in the Temperate zone (Northern and Central Italy) and summer (May–October) in the Mediterranean zone (coastal and southern Italy), see Fig. 4c. Summer accounts for most fires (Fig. 4b). Although large fires (over 100 ha) represent just 2% of events, they account for 50% of the burned area and are the primary operational concern (Fig. 4e).

Fig. 4.

Summaries of burned area and number of wildfires occurring in Italy from 2007 to 2022. (a) Daily trend of total burned area and number of wildfires for the entire period 2007–2022 (note: the burned area y-scale has been set to a maximum of 10,000 ha for the sake of plotting; however, some years had values out of the range: 23,838 ha in 2007 (23 July 2007); 34,219 ha in 2009 (23 July 2009); 12,981 ha in 2021 (23 July 2021); the reference date corresponds to the wildfire ignition date). (b) Cumulative burned area and number of wildfires through the year for the whole period. (c) Burned areas locations are divided into the main wildfire seasons, namely, summer and winter. (d) Total burned area and number of fires per year, divided into summer and winter seasons. (e) Wildfire contributions to number of fires and burned area when classified into five size categories according to their extent: under 1 ha, 42,719 fires; 1–10 ha, 42,071; 10–100 ha, 16,097; 100–1000 ha, 1921; over 1000 ha, 96.

The FFDR indices were evaluated against past wildfire events by selecting pixels at least 70% burnt to ensure relevance and reduce edge effects. Values were taken on the wildfire ignition date, though this approach may not fully capture the duration of large, multi-day fires. Pixels classified as ‘non-fire’ were unaffected on that date, subsequently used for testing detection capability. Only wildfires larger than 70 ha were included owing to the 1 km resolution and selection criteria.

As noted by Di Giuseppe et al. (2016), FFDR indices should associate extreme values with wildfire activity. This was tested using the rank-percentile method (Eastaugh et al. 2012), which provides a non-parametric measure useful to compare fire indices considering their different distributions. It consists in fitting the ranked percentiles of the fire values with the Theil-Sen technique (Theil 1950; Sen 1968) to reduce outliers impact: a good index should be fitted by a zero-slope line with intercept equal to 100. Note that the percentiles were computed as inverse values from the complementary cumulative distribution function for indices where fire activity was associated with lower values. Along with their discrimination capability, Receiver Operating Characteristic (ROC) curves were generated together with the Area Under the Curve (AUC) to test indices detection capability (Mas et al. 2013; Richardson et al. 2024).

An important aspect of testing FFDR indices for operational use is comparing the percentage of territory classified into fire danger classes (Sirca et al. 2018) with wildfire classifications. A robust index should minimize large high-risk areas while accurately categorizing wildfire events. The analysis evaluated the territorial coverage of each class over time, examining the mean percentage of territory classified on wildfire days (i.e. when wildfires pixels were selected). This helps assess how well the index supports resource management, as too much high-risk territory complicates prioritization. Danger classes were defined using the same percentiles for all the indices, as commonly used (de Jong et al. 2016) (see Table 4) - the percentiles were based on RISICO’s operational implementation to carry out comparisons between FFDR indices.

The proposed analysis differentiates between summer and winter wildfire seasons as defined above, based on Italian wildfire regimes.

Fuel map analysis

The susceptibility model showed good performance at the European level with an AUC of 0.89, mean squared error (MSE) of 0.14 and accuracy of 0.81. The Gini impurity index (see Supplementary material) identified vegetation continuity as the most important variable, followed by temperature and relative humidity, consistent with similar studies (e.g. Trucchia et al. 2022). For Italy (see Fig. 5a), the susceptibility map was analyzed using the Italian wildfire dataset, observing wildfire occurrence (i.e. burned area locations) aligning with higher susceptibility areas as expected from the purpose of the map. The analysis also considered different wildfire size classes (see Fig. 5c), revealing a clear distinction resulting from susceptibility, with larger wildfires found in the highest quartile. This emphasized the discriminative capability of the susceptibility map not only in identifying fire-prone areas, but also in distinguishing between areas with higher or lower potential for fire spread, supporting its use in fire spread assessment using the ROS module.

Fig. 5.

Analysis of susceptibility and fuel maps for Italy. (a) Susceptibility map; (b) fuel map. (c) Distribution of susceptibility values compared with those obtained in the case of wildfire pixels (top plot). Also, the values distribution is shown with respect to wildfire extent classes (bottom plot - A refers to burned area). (d) Percentage of national coverage of each fuel class compared with their percentage affected by past wildfires.

The final fuel map is then presented in Fig. 5b. A comparison between fuel class distribution and percentage burnt (Fig. 5d) highlights the map’s ability to prioritize areas: it identifies classes (i.e. highly susceptible shrublands and conifers) that account for a limited area but are most affected by wildfires.

Model performance

The rank-percentile analysis is shown in Fig. 6, with index distributions and discrimination diagrams shown in the Supplementary material. In the seasonal analysis, the ranking is computed with respect to seasonal indices distributions because indices may show strong seasonal variation. However, FFDR indices should maintain strong discrimination capability across seasons.

Fig. 6.

The rank-percentile points (solid lines) and fitted curves (dashed lines) for the whole dataset (in blue) and divided between summer (S - in orange) and winter (W - in purple) fire seasons. The slope and intercept of each curve are also reported in each plot (the couples of numbers reported in the bottom-right part of each plot).

Among the RISICO indices, DFMC shows the best overall discrimination (see slopes and intercepts in Fig. 6), outperforming others like ROS. However, DFMC’s performance decreases in winter, whereas other indices, like FLI, remain more stable and even improve in winter. All indices, except FLI, show reduced discrimination in winter. The W index, as a sub-model of ROS, shows poor discrimination – it is intended to be useful operationally once a wildfire starts to assess wind impact. The ROS and FLI indices, commonly used in RISICO, perform well, comparable with other benchmarks like FWI and FFDI.

The ROC curves and AUC values are shown in Fig. 7 and summarized in Table 5. Among the RISICO indices, ROS has higher AUC values for the entire period and summer, whereas FLI performs better in winter. As noted, W has poor detection capability. All FFDR indices except FLI show decreased performance in winter. However, ROS outperforms other FFDR indices in both overall and seasonal cases, whereas DFMC and FLI have AUC values comparable with other benchmarks like FWI and FFDI.

Fig. 7.

The ROC curves for the different indices for the whole dataset (in blue), and divided between summer (S - in orange) and winter (W - in purple) fire periods. The AUC values of each curve are also reported in each plot (number in the bottom-right part of the plot). The dotted grey line represents the threshold above which the ROC curve should lie to indicate acceptable model performance.

The percentage area of Italy classified into different danger classes focuses on ROS, FWI and FFDI in the following figures, with additional analysis in the Supplementary material. Fig. 8 highlights the High (H) and Extreme (E) classes daily evolution, focusing on areas above the 99th percentile. ROS shows less seasonality than FWI and FFDI. ROS classifies a smaller area in higher classes (under 5% for High and 0.5% for Extreme) in summer compared with FWI and FFDI, which show peaks above 10% for High and 2% for Extreme classes. FLI shows similar behavior to ROS, whereas DFMC exhibits strong seasonality with a greater area classified in higher danger classes, like FWI. Analyzing the other indices, FFWI and Sharples-F show weak seasonality and behave similarly to ROS and FLI.

Fig. 8.

The daily areas classified into High (H - in red) and Extreme (E - in black) danger classes, resampled as average weekly values, are reported for the ROS, FWI and FFDI indices, highlighting summer (S - in orange) and winter (W - in purple) fire periods.

Fig. 9 summarizes the mean area classified into each danger class during wildfire days, comparing the percentage of Italian territory with how wildfires were classified in each danger class for summer and winter periods. The overall period is reported in the Supplementary material. The sum of High and Extreme classes is also reported in Table 5. ROS and FLI indices classify limited territory, especially in the Extreme class, but effectively categorize fire pixels in higher danger classes. For instance, in summer, ROS classifies 2.32% of the territory in High and Extreme classes, containing 29.24% of fire pixels, whereas FWI classifies 6.01% of the territory with 32.14% of fire pixels. ROS reduces the classified area by over 50% while maintaining similar fire pixel classification. Considering also the Medium-High (MH) class, ROS classified 11.68% of the area with 65.7% of fire pixels, while FWI classified 23.44% with 71.37%. FFWI shows better performance in summer compared with other indices, which instead present similar behavior to the FWI. In winter, ROS still performs well, classifying 1.69% of the territory with 16.28% of fire pixels. Most fire danger indices do not classify areas in higher classes, except for ISI, FDI and FFWI. Among these, ROS classifies more fire pixels in higher classes. DFMC shows weaker performance in winter, whereas FLI performs similarly to ISI.

Fig. 9.

Comparison between mean area classified into danger classes during fire days (bottom row) and the classification of fire pixels (top row) – each matrix reports on a different fire danger index. The difference between summer (a) and winter (b) is reported.