This study considered a set of four fires that occurred in the southern Caspian forests of northern Iran, specifically in the Siahkal forest area and in Golestan National Park (GNP; Fig. 1). The south Caspian forests (16 481.95 km2) cover about 1.2 % of the whole of Iran (Marvi Mohadjer, 2005) and range from sea level up to 2500 m a.s.l. (Siadati et al., 2010). The area presents contrasting bioclimatic differences in comparison to the central and southern parts of the country, which are characterized by xeric weather conditions.

The Siahkal forest area is located in northern Iran, occupies 1050 km2 and presents a very high altitudinal range from the lowest areas at 10 to the highest mountains at 2500 m a.s.l. (Fig. 1). The annual precipitation ranges from 600 mm in the southern part to 2000 mm in the northern and highest mountains, and most of the annual rainfall occurs in autumn. Air relative humidity exceeding 80 % is responsible for frequent fogs at the highest altitudes. The average annual temperature is 16 and 25 ∘C in summer. Average minimum temperatures of the coldest month are usually over 0 ∘C (Akhani et al., 2010). The Siahkal forest area is located in the long, narrow vegetation belt on the northward slopes of the Alborz Mountains and constitutes the majority of the Eurosiberian flora in Iran (Djamali et al., 2009). The highest proportion of the Siahkal area is covered by forests (∼46 %) that are dominated by temperate broad-leaved deciduous trees such as various thermophilous Tertiary relict species (e.g., Zelkova carpinifolia, Parrotia persica, Pterocarya fraxinifolia, Quercus castaneifolia) and Asian subtropical trees in some cases (e.g., Diospyros lotus, Gleditsia caspica, Danae racemosa, Albizia julibrissin) (Table 1; Akhani, 1998; Akhani and Ziegler, 2002; Leestmans, 2005; Leroy and Arpe, 2007).

The Golestan National Park is situated in northeast Iran and covers about 920 km2 of land (Fig. 1). The National Park is located in a transitional position between the sub-humid south Caspian region and the semi-arid zones of the central and east-central Iranian Plateau. The GNP ranges from 450 to 2400 m a.s.l. The wet air masses from the Caspian Sea are blocked by the high mountain ranges, which create particular microclimatic conditions, with annual precipitation ranging from 150 mm in the southeast to more than 1000 mm in some central parts of the GNP (Akhani, 1998). The mean annual temperature is 11.5–17.5 ∘C and the mean summer temperature is 28 ∘C (Akhani, 1998). The park exhibits a diverse mosaic of vegetation units, including the Hyrcanian low- to high-altitude mesophytic forests, shrublands, open and closed scrub sometimes mixed with C4 grasslands, Juniperus woodlands, mountain steppes and meadows, Artemisia and Artemisia–Stipa steppes and different transitional and halophilous communities (Table 1; Akhani, 1998; Akhani and Ziegler, 2002).

In the period 2000–2011, northern Iran experienced annually on average about ∼400 fires that burned around ∼2000 ha. Large and extreme fires in the study areas are commonly linked to drought conditions, heat waves, strong winds and fine dead fuel accumulation (Mirdeylami et al., 2014). As many as 90 % of the fires in the northern Iran and study areas are caused by people (Sarkargar Ardakani, 2007; Zarekar et al., 2013; Mirdeylami et al., 2014). Fires in northern Iran commonly occur during the short drought season in autumn, characterized by hot and dry winds that desiccate the forest understory. These conditions mostly lead to low-intensity surface fires that rarely exceed 0.3 m flame height (Adel et al., 2012).

Wildfires in the GNP, as well as in the Siahkal forests, are distributed from June to December, with two peaks in the number of fires and area burned in June–July and November–December (Fig. 2). Although observed annual fires and burned area in the GNP and the Siahkal forests present high interannual variability during the period 2000–2011, the hardest wildfire campaigns correspond to the latest years and especially to 2010 (Fig. 3). During the period 2000–2011, the Siahkal area experienced on average about ∼13 fires per year and about ∼60 ha burned (Department of Forestry, Natural Resources Office, Guilan, Iran; Fig. 3). Approximately 85 % of the fires in Siahkal burned less than 10 ha; a small number of fires (about 15 %) is responsible for half of the area burned (Fig. 4) and no fires larger than 100 ha were observed in the studied period. In the GNP during the period 2000–2011, ∼12 fires per year have been recorded on average, with ∼200 ha burned (Fig. 2). In this area, the largest fires (> 100 ha) accounted for about 15 % of the fires and were responsible for almost 75 % of the total area burned (Department of Environment, Golestan, Iran; Fig. 4). The largest wildfire in the GNP (Cheshme-Sardar fire event) was observed on 15 November 2010 and burned an area of about 900 ha.

Four wildfires that affected the study areas during the 2010 and 2011 fire seasons were selected as case studies: Toshi and Malekroud fires in Siahkal forest and YekeBermagh and Gharangi fires in the Golestan National Park (Fig. 1). The exact location, main types and dominant species of vegetation together with fire data for the different case studies are summarized in the Table 1. For all case studies, ignition location coordinates were determined from fire reports (Department of Forestry, Natural Resources Office, Guilan, and Department of Environment, Golestan, Iran, personal communication, 2011, 2012) and interviews with forest rangers, firefighters and park managers, and burned area perimeters were recorded after the fire events using the Global Positioning System (GPS).

The Toshi wildfire occurred near the village of Toshi (lat 37∘11′ N, long 49∘88′ E) in August 2010, and the 25 h fire event burned 34 ha (Fig. 5; Table 1) of mixed dense woodland (∼16.4 ha), grasslands (∼13.4 ha) and grass–shrublands (∼4.7 ha). The ignition point was located near a steep slope in an agricultural area (Fig. 5). The weather was characterized by a maximum temperature of 35 ∘C, average relative humidity of 50 % and northeast winds (Table 2). The fire spread towards southeast, driven by the wind and the topographic conditions.

The Malekroud wildfire occurred near the town of Malekroud (lat 37∘03′ N, long 49∘84′ E) in December 2010 and burned approximately 24 ha covered by heterogeneous structural characteristic mature forest in a low-elevation area (Fig. 5; Table 1). The fire started near a road along the southern border of the fire perimeter. It was extinguished by the forest firefighters after 17 h near a road along the northern border of the fire perimeter (Fig. 5). The day was characterized by a moderate maximum temperature (∼25 ∘C), average relative humidity of 58 % and southern winds. The fire was driven northward by the mild slope and the wind.

The YekeBermagh wildfire occurred in the southern part of the Golestan National Park (lat 37∘22′ N, long 56∘03′ E) in July 2011 (Fig. 5; Table 1). The northern part of the YekeBermagh area is characterized by a flat topography, while the southern part has a more complex and steep terrain with high spatial and temporal variability in wind speed and direction. Most of the 60 ha burned was covered by grasslands. Juniperus woodlands and grass–shrublands composed by montane Juniperus excelsa in steep slopes and subalpine Juniperus communis on exposed high slopes (Akhani, 1998) were also affected by the fire. The day of the fire the weather was hot (31 ∘C maximum temperature) and dry (21 % relative humidity). Fire spread was driven by the topography and the southwestern winds.

The Gharangi wildfire occurred in March 2011, in the southern part of the Golestan National Park (lat 37∘21′ N, long 56∘02′ E), and burned about 10 ha (Fig. 5; Table 1) of dense-mixed woodland. The area presents a mountainous orography with an altitude range between 1200 and 2160 m a.s.l. The fire weather was mild, with a maximum air temperature of 17 ∘C and average relative humidity of 49 %. The fire spread towards the north and northeast, driven by southwest winds. The fire intensity was low due to the shielding effect of the dense and closed canopy.

Fuel model and canopy characteristic maps for the study areas were produced by field sampling on the vegetation complexes existing on the 1:25 000 land-cover maps of 2004 (Department of Forestry, Natural Resources Office, Guilan, and Department of Environment, Golestan, Iran) due to the lack of information on forest and shrubs cover types that could allow standard fuel model assignment. Furthermore, with the georeferenced data derived from field sampling in the study areas we generated fuel model maps and photo guides, improving the initial 1:25 000 land-cover maps and creating finer-scale vegetation layers. The field samplings were conducted following the line-intercept sampling (Marshall et al., 2000, 2003) method, with the objective of measuring the surface fuel model parameters and canopy characteristics.

On the whole, according to the topography in the study areas and the vegetation types, 21 line transects with a distance of 150 m in Siahkal forests and 25 line transects with a distance of 100 m in the GNP were used to georeference 188 and 250 sampling plots, respectively (Table 3). Considering the spatial distribution and the coverage degree for the different species within the different vegetation types, 1 m × 1 m square sampling plots were used for herbaceous fuel types and 10 m × 10 m square sampling plots for shrubby and forested vegetation types. We measured species composition, fuel-bed depth, litter type (conifer or broadleaf), herbaceous cover, shrub cover and canopy cover (Table 3). Visual estimations were used to assign a canopy cover class (< 1, 1–5, 6–10, 11–25, 26–50, 51–75 and 76–100 %) in every plot. We also produced a photo guide for main fuel models types (Fig. 6)

In this study, standard fuel models (Anderson, 1982; Scott and Burgan 2005) were assigned to the existing vegetation and land use land-cover types based on their similarities in structural characteristics (Figs. 5, 6; Table 3). The grass-dominated standard fuel models used were GR3, GR4, GR5, GR6, GR7 and FM3. GS1, GS2, GS3, GS4, FM5 and FM6 fuel models were considered for the vegetation, presenting a mixture of grass and shrub components. SH1, SH2, FM5 and FM6 fuel models were assigned to areas with sparse grassland among shrubby patches covering at least the 50 % of the surface. In forested areas with grass–shrub and litter mixed understory, TU1, TU2, TU3, TU5, FM8, FM9 and FM10 fuel models were used, whereas TL2, TL6, TL8 and TL9 were used for woody fuels beneath forest canopies. FM9 and FM10 fuel models were assigned to the timber litter, hardwood litter and litter and understory areas. Non-burnable (NB) fuel models were assigned for roads, buildings, urban areas, ploughed agricultural lands, water bodies and bare ground; in that case the geospatial information was gathered from the 1:25 000 digital topographic maps (National Cartographic Center of Iran).

Fire spread simulation systems require spatial grids of topography (slope, aspect and elevation), surface fuels (fuel model) and fuels canopy characteristics (i.e., stand height, crown base height, crown bulk density, canopy cover) as basic inputs for the simulations. These data layers were assembled in a landscape file with 10 m resolution. Topography layers were derived from the digital elevation model (10 m resolution; National Cartographic Center of Iran, NCC) for each study area. As previously described, surface fuel layers were prepared based on land-cover maps and field sampling.

Weather data of the day of the fire, corresponding to hourly air temperature, relative humidity, rainfall, wind speed and wind direction, were collected from the weather stations nearest to the wildfire case studies (Fig. 5 and Table 2).

Initial fuel moisture content for the 1, 10 and 100 h dead fuels (Table 3) was determined following the methodology proposed by Rothermel (1983; Table A1). With this method, we estimated the fine dead FMC for each case study, and then we derived 10 and 100 h dead moisture by adding 2 and 4 %, respectively, to the 1 h dead FMC (Hardison, 2003). The live herbaceous and woody FMC values (Table 3) were estimated from literature data (Arca et al., 2007; Sağlam et al., 2008; Chuvieco et al., 2011) and mostly from field observations.

Fire simulations were run at 10 m of resolution using different combinations of standard fuel models (Anderson, 1982; Scott and Burgan, 2005) for the main fuel types (i.e., grasslands, grass–shrublands, shrublands, timber understory and timber litter) affected during wildfire events (Table 4). For all simulations and fuel models, the adjustment factor for the fire spread rate was set at 1.0. Suppression activities were not considered in the simulations due to the lack of information, as well as spot and crown fires, since both were not observed in the case studies presented in this paper. Ignition location and fire spread duration used as inputs for each case study are provided in Table 1. Vector files of the simulated fire perimeters and gridded data of simulated rate of spread (m min-1), fireline intensity (kW m-1) and flame length (m) were exported and analyzed in GIS environment.

The influence of fuel models on the accuracy of simulated fire spread and behavior was assessed for all the case studies. An error matrix between observed and simulated fire perimeters was calculated to define the frequency of each case (presence/absence of burned areas). Sørensen's coefficient (SC; Legendre and Legendre, 1998) and Cohen's kappa coefficient (K; Congalton, 1991) were used as measures of the spatial accuracy of the extent of the simulated fire spread (Arca et al., 2007; Salis, 2008).

Sørensen's coefficient was used as indicator of the exclusive association between observed and simulated burned areas. SC values were calculated as follows: SC=2a2a+b+c, where a is the number of cells coded as burned in both observed and simulated data (burned area agreement), b is the number of cells coded as burned in the simulation and unburned in the observation (modeling overestimation) and c is the number of cells coded as unburned in the simulation and burned in the observation (modeling underestimation; Arca et al., 2007).

Kappa statistics computes the frequency with which the simulated area agrees with the observed area, with an adjustment that takes into account agreement by chance (Filippi et al., 2014). K values were calculated as follows: K=N∑i=1rxii-∑i=1x1+x+1N2-∑i=1rxi+x+i, where r is the number of rows in the matrix, xii is the number of observations in row i and column i, xi+ and x+i are the marginal totals of row i and column i, respectively, and N is the total number of observations. Both K and SC coefficient values typically range between 0 and 1, with values close to 1 indicating very high spatial agreement between simulated and observed fire perimeters (Arca et al., 2007).

Moreover, the “zonal statistics” tool of ArcGIS 10 was used to analyze and summarize the fire behavior data (ROS, FLI and FML) for each fuel model.

For all the case studies, the simulated burned areas were compared with the observed fire perimeters (Fig. 7, Tables 4 and 5). Overall, the statistics showed that FARSITE performances with the highest values for K and SC coefficients, and therefore the scenarios that better replicate the observed fires, were obtained for all the case studies using the standard fuel models of Scott and Burgan (2005) with the exception of simulation II of Malekroud, where the standard fuel model (FM9) of Anderson (1982) showed the best accuracy in replicating the fire perimeter (Table 4).

In the Toshi fire event, the best results were obtained in simulation III (Fig. 7a, Table 4), where about 30.1 ha of the final fire area coincided with the observed fire size, while 4.1 and 5.5 ha were underestimated and overestimated, respectively, by FARSITE. As previously pointed out, the best values of SC and K coefficients were obtained in simulation III (SC = 0.86, K=0.82; Table 4), whereas the other simulations presented lower accuracies, with SC values ranging from 0.48 to 0.83 and K values from 0.45 to 0.81. The best performance for Toshi wildfire, regarding the standard fuel models used, was obtained by the GR6 fuel model (SC = 0.92, K=0.87; Table 5) for grasslands and the worst was observed for the TU3 fuel model (SC = 0.75, K=0.73; Table 5).

Simulation II of the Malekroud wildfire event (Fig. 7b, Table 4) replicated well the observed fire event, with an agreement between the observed and simulated fire area of about 20.6 ha and FARSITE underestimation and overestimation of 3.5 and 5.5 ha, respectively. The statistical analysis showed that the FM9 fuel model in simulation II provided the highest SC and K values (SC = 0.85; K= 0.82; Table 5), while the other simulations using TL6 and FM10 fuel models gave SC values ranging from 0.73 and 0.79 and K values ranging from 0.71 and 0.75 (Table 4). Focusing on single fuel models, the FM9 fuel model in Toshi case study provided the worst accuracy performance (SC = 0.48; K= 0.45; Table 4).

In simulation VI of the YekeBermagh case study (Fig. 7c, Table 4), the simulated fire area was characterized by an overestimation of 30.7 ha, mainly in the right back-flank of the fire spread. The agreement between the simulated and observed fire area was about 46.8 ha, while 11 ha of the fire area was underestimated (Table 4). The statistical test showed that in simulation VI, the GR4 fuel model provided the best SC and K values (SC = 0.82, K= 0.81; Table 5), while the worst performances were provided by the FM3 fuel model in simulation VII (SC = 0.13, K= 0.12; Table 4) due to the wide underestimation of the area burned. The large underestimation was also confirmed for the FM5 and FM6 fuel models (Table 4).

In simulation I of the Gharangi wildfire event (Fig. 7d, Table 4), about 7.5 ha of the observed fire area was correctly simulated as burned area by FARSITE. The extent of the underestimation by the simulation was approximately 2.6 ha and the overestimation was 2.2 ha. The best agreement between simulated and observed fire was linked to TL9 fuel model (SC = 0.91; K= 0.91; Table 5), which was characterized by small overestimation and underestimation of the FARSITE perimeter.

Comparing the standard fuel models associated with the best simulations of FARSITE for each case study, the higher SC and K values were obtained using the GR6 grassland model in simulation III of the Toshi fire (SC = 0.92; K= 0.87; Table 5) and the TL9 timber model in simulation I of the Gharangi fire (SC = 0.91, K= 0.91; Table 5). The worst performances were provided by the model TU1 in simulation I of the Gharangi fire event (SC = 0.47; K= 0.45; Table 5). On the whole, GR6, TU2, TU5 and TL9 fuel models replicated well the observed area burned (SC ≥0.90 and K≥0.82; Table 5).

Due to differences in fuel models characteristics, topography and weather conditions, the simulations revealed diverse potential fire behavior. Surface fire rate of spread, fireline intensity and flame length were analyzed for each of the fuel models used in the four case studies (Fig. 8 and Table 5). The fire simulation outputs showed complex patterns that were generally related to the dominant fuel types and the topography.

Overall, the average wind speed conditions ranged between 14 and 23 km h-1 for the case studies presented (Table 2), and this is one of the main reasons why fires did not show high, fast-spreading output values (0.53–2.61 m min-1 average ROS; Table 5). The highest values of simulated ROS were observed with tall and dense grasslands and sparse shrubland vegetation in Toshi and YekeBermagh case studies (Table 5). The grasslands presented the fastest ROS, which varied from 0.05 to 10.84 m min-1 (Table 5) depending on topography; the shrublands showed a ROS ranging from 0.05 to 8.06 m min-1 (Table 5). The lowest ROS (< 1 m min-1; Table 5) was obtained for the areas covered by mixed hardwood forest (TU1) and pure hardwood forest (TL6) in the Gharangi wildfire. In woodlands, modeled fire ROS was very slow due to the high fuel compactness and the relatively high moisture content; this explains the ROS values 2–3 times lower than in grassland fuel types (Table 5).

In addition to the ROS, relevant differences in terms of fire intensity (i.e., FLI and FML) were found between the different vegetation types. The high-fuel-load tall grass (GR6) presented the highest values (average FLI >350 kW m-1 and FML > 1.4 m; Table 5) and the compact litter and understory (TU1, TU2 and TL6) showed the lowest values (average FLI < 100 kW m-1 and FML < 0.5 m; Table 5). Nonetheless, high values were also locally associated with shrubby fuel models (SH1 and SH2; FLI < 250 kW m-1 and FML > 1 m; Table 5) in YekeBermagh wildfire case study. The obtained fire intensity results in the different fuel models resembled observed flame lengths during the analyzed fire events.