You are on page 1of 11

Instituto Nacional de Investigación y Tecnología Agraria y Alimentaria (INIA

Available online at

Forest Systems 2013 22(1), 71-81
ISSN: 2171-5068
eISSN: 2171-9845

The relationship between landscape patterns and human-caused
fire occurrence in Spain
S. Costafreda-Aumedes1*, A. Garcia-Martin2 and C. Vega-Garcia1

Departamento de Ingeniería Agroforestal. Universidad de Lleida. Alcalde Rovira Roure, 191. 25198 Lleida, Spain
Centro Universitario de la Defensa. Academia General Militar.
Ctra. de Huesca, s/n. 50090 Zaragoza, Spain

Aim of study: Human settlements and activities have completely modified landscape structure in the Mediterranean
region. Vegetation patterns show the interactions between human activities and natural processes on the territory, and
allow understanding historical ecological processes and socioeconomic factors. The arrangement of land uses in the
rural landscape can be perceived as a proxy for human activities that often lead to the use, and escape, of fire, the most
important disturbance in our forest landscapes. In this context, we tried to predict human-caused fire occurrence in a
5-year period by quantifying landscape patterns.
Area of study: This study analyses the Spanish territory included in the Iberian Peninsula and Balearic Islands
(497,166 km2).
Material and methods: We evaluated spatial pattern applying a set of commonly used landscape ecology metrics to
landscape windows of 10 × 10 sq km (4,751 units in the UTM grid) overlaid on the Forest Map of Spain, MFE200.
Main results: The best logistic regression model obtained included Shannon’s Diversity Index, Mean Patch Edge
and Mean Shape Index as explicative variables and the global percentage of correct predictions was 66.3%.
Research highlights: Our results suggested that the highest probability of fire occurrence at that time was associated
with areas with a greater diversity of land uses and with more compact patches with fewer edges.
Key words: human-caused wildfires; landscape ecology; logistic regression.

Landscape structure is the result of past and present
interactions between human activities and natural processes (De Aranzabal et al., 2008; Echeverría et al.,
2007; Löfman and Kouki, 2003; Naveh and Lieberman,
1994; Serra et al., 2008). Variations in frequency, magnitude and extension of disturbances produce complex
patterns in vegetation composition, age structure and
patch size distribution over the landscape (Farina,
2006; Regato et al., 1999; Saura, 2010). Thus, the spatial
pattern of vegetation, usually assessed by different metrics, allows understanding historical ecological processes and socio-economic factors. Landscape composition and configuration metrics have been proved to
be influenced by climate (Pickett and White, 1985),

forest pests and diseases (Hatala et al., 2010; Romero
et al., 2007), land use changes (Ferraz et al., 2009;
Gallant et al., 2003; Serra et al., 2008), human settlements (Fuller, 2001), deforestation (Löfman and
Kouki, 2003; Zhang and Guindon, 2005), the abandonment of traditional agrarian tasks (plowing, grazing
and cutting) (De Aranzabal et al., 2008) and f ires:
burned area and frequency (Chang et al., 2007; Moreno, 2007; Naveh and Lieberman, 1994; Pickett and
White, 1985).
In the Mediterranean environment, the landscape
has long been modified by human influence (Pausas,
2006), becoming what we call a cultural landscape
(Farina, 2006). Landscape patterns are created by
direct human action through the design of boundaries
between crops and natural vegetation, wildland-urban

* Corresponding author:
Received: 24-01-12. Accepted: 12-01-13.
Abbreviations used: Landscape Metrics (Table 2), ANND (Average Nearest Neighbor Distance Index).

(2012) have included independent variables measuring landscape pattern. finding that certain landscape configurations were more vulnerable (fireprone) than others. and only the Northern third has an Atlantic climate. 1999) . many have included geographic or spatial variables (i. 2001. 2009).000). Torre Antón. 1 displays the vegetation classes used.166 km2). 71-81 interfaces.. fire was the main tool used in cleaning and removal of forest residues. 2005. All of them were indices commonly used in the scientif ic literature of fire landscape ecology (Forman. Costafreda-Aumedes et al. The main reference for the study of vegetation cover in Spain is the Forest Map of Spain by Ruiz de la Torre (1990) at 1:200. 2010). Independent variables: landscape metrics To characterize the vegetal landscape pattern we selected 13 metrics related to the area. along with grazing and firewood extraction (Pausas. (2012) did analyze landscape structural factors (11 metrics) related to increased wildfire incidence in forest-agriculture interfaces within the SISPARES monitoring network (observation size 16 km2). Among the studies that have dealt with fire occurrence in the literature. for instance. defined in Table 1. collecting information about other land uses. RomeroCalcerrada and Perry. park-like open forest structures (dehesas) with undergrowth and high forests of variable densities (EEA. 1995. 2009.2% are caused by people (MAGRAMA. Padilla and Vega-García. extensive shrublands. presence of infrastructures. shape. 2007). 2010). (2009) considered area. areas with scarce natural vegetation cover (grasses. About 75% of human-caused forest fires in Spain are related to the rural use of fire for vegetation management (MAGRAMA. Material and methods Study area This study analyzes the Spanish territory included in the Iberian Peninsula and Balearic Islands (497. Hence. Lloret et al. the quantitative analysis of landscape structure becomes a relevant tool to make inferences on future fire occurrence. the group they belong to and a brief description about the information they convey (McGarigal et al. 2002).500 species of trees. Rothermel. Martínez et al. fires are still linked to the persistence of traditional agrarian activities (Martínez et al. density and fragmentation indices (landscape and cropland fragmentation) with socio-economical and geographical variables to predict human-caused fire occurrence at the municipal scale in Spain.e.72 S. encompassing the wide range of compositions and configurations that can be found in Spain. rangelands). 2012. Table 2 shows the selected indices. 1998. Most of the study area is dominated by a Mediterranean climate. Martínez et al. In the past. and 96. 1999. Frohn. we propose that some metrics may be more appropriate than others to characterize and identify fire-prone landscape traits at the national level. 2002. These climatic zones and the complex topography combined with human socio-economical development over millennia have given way to a very uneven spatial distribution of the vegetation. or indirectly by allowing the spread of disturbances. landscape metrics may be proposed as surrogate variables for human activities in our Mediterranean environment. Martínez et al.. 2011) but only Henry and Yool (2004). fragmentation and diversity of the vegetation patches. A recent study by Ortega et al.. Hernández-Stefanoni.. It locates more than 5. Ortega et al.000 scale (digitized 1:50. / Forest Systems (2013) 22(1). Fires are a human artefact emanating from the rural activities that shape the Mediterranean landscapes. The classification was designed according to the fuel models of Rothermel (1972) and species response to f ire (Riaño et al. Consequently. Sturtevant and Cleland. Thus. These 13 metrics were computed for the landscape units in Spain using Patch Analyst 4 (Elkie et al. shrubs and grasses. Fig. McGarigal et al. 2002. shape and diversity indices) in remote sensing images (Landsat TM and SIR-C data) to relate landscape pattern with historical fire occurrence in National Saguaro Park (Arizona). the aim of this paper is to evaluate specifically the relationship between landscape patterns and human-caused forest fire occurrence with a comprehensive array of landscape metrics. In order to fulfill the goals of this study and work at the considered scale (Peninsular Spain and Balearic Islands) it was necessary to reclassify the different plant species and land uses in manageable categories meaningful for risk analysis..600 fires occur per year in Spain. 2010. Approximately 18. Building on these findings. Henry and Yool (2004) calculated landscape metrics (area. Henry and Yool. 2010).. 2004. 2007).. 2004). In current times. 1972. combining the presence of medium-scale farming areas. (2009) and Ortega et al.. Torre Antón.

but some irregular cells on the coastline and in the boundaries between UTM zones 29. and ArcGis 9. Because these landscape units were not constant in area. 2009). Figure 2. The original grid consisted of 5.3 (ESRI Inc. it was not possible in principle to compare the values for each grid. The landscape units corresponded to 10 × 10 sq km UTM grid cells used by the Ministry of Environment in Spain to record locations of fires in the reports (Fig.. 2002). as examples of two forest landscapes in Atlantic (S1) and Mediterranean Spain (S2).278 cells. 30 and 31 were excluded to obtain comparable landscape units (100 ± 25 km2). 19891993. 2). Forest Map of Spain with the classes used in the study grouped in six land uses.Fire occurrence and landscape in Spain 73 Figure 1. The resulting grid of 4. Zoom windows of 900 sq km. since some metrics are sensitive to the size of the landscape unit (Saura. . Human-caused forest fire occurrence in Spain.751 cells was set as the spatial base for calculation of the explanatory variables and for the analyses of the present study.

000-260. 2006). perennial oak. 9 and 10. Land use classes and fuel description. cupressaceous. other broadleaves) Mixed forest-FM Shrub-driven fuels This group is composed by formations with woody plants shorter than 7 m. but conditions that could develop again in the future (Vega-García and Chuvieco. dunes.492 in 2004-2008).468 and 437. 2010). MAGRAMA. and in between. Classes for landscape metrics calculation Forest soil layer-driven fuels Formations with woody plants taller than 7 m and corresponding to fuel models 7.000) hectares than are burned nowadays (50. or Y = 0 if otherwise.000-190. Continental Water-W This study period of 5 years was carefully chosen so that it chronologically followed the time span between the acquisitions of the ortophotos (1982-1986). 71-81 Table 1. 5 and 6 Xerophylous shrubland-XS Mesophylous shrubland-MS Grass-driven fuels This group includes vegetal formations with natural herbaceous habitats (fuel models 1. Viegas et al. According to Chuvieco (1996). 2 and 3) and ferns for their shade-tolerance preference and low height Xerophylous grassland-XG Mesophylous grassland-MG Little/No-vegetation Composed by open spaces with little or no vegetation such as beaches. 2) was summed up for each 10 × 10 sq km UTM cell or landscape unit and coded as Y = 1 if at least one fire took place in the period and cell. industrial units. f ires burned slightly more (90. was very similar (12. The dependent variable was the probability that at least one fire happened in the 5-year period between 1989 and 1993.811 in 1989-1993) to present numbers (10. alone or under tree cover less than 30% fraction cover that correspond to Rothermel fuel models 4. sandy areas. only anthropogenic fires were selected for this study. Conifers-FC (pinaceous. This information could be easily summarized in number of f ires per year for each 10 × 10 sq km UTM grid used by the Ministry to locate fires. poplar stand. It also includes those herbaceous or woody plants that grow in water Dependent variable: fire occurrence The f ire history registry from 1983 to 2008 was provided by the Ministry of the Environment and Rural and Marine Affairs (MAGRAMA) in Spain.932-25. More importantly. (1999) and VegaGarcía and Chuvieco (2006). mines dumps and construction sites from herbaceous and woody agricultural plants Artificial surface-AS Agricultural areas-AA Continental water Contains any surface covered by water bodies or stream courses.000 ha. . We separate urban fabric. bare rocks and sparsely vegetated and burned areas Little/No-vegetation-NV Anthrophic surface High human influence areas. 8. the years 1989 and 1994 were severe fire-years. According to our stated goal. mix of conifers and other conifers) Broadleaves-FB (oak woodland. reflecting worse conditions than at present. nude land. the reasonable period for updating dynamics in vegetation maps is around 4 or 5 years. dividing these into natural (lightning) and human-caused fires. The number of occurrences. a number of large fires occurred in those years (burning 426. mix of broadleaves. though. The fire reports routinely included information about the causes of the fires. beech forest. Fire occurrence data in the historical reports (Fig. 1990).635 ha respectively). the field work (up to 1989) and the date of creation of the Forest Map of Spain (MFE200. eucalypt forest. riparian forest. / Forest Systems (2013) 22(1).913-20. Costafreda-Aumedes et al.74 S.

102 1.624 0.876 cells) of the 4.231 0. being the main that independent variables are uncorrelated with each other.188 7. 2009*. Ortega et al. Romero-Calcerrada and Perry.. out of 772 cells with > 90% forest cover. 2004*.. 2004*. Hernandez-Stefanoni. 2012*) (Henry and Yool.199 (Henry and Yool. 1995...314 (Henry and Yool.993 (Ortega et al.. Ortega et al. 2004*.276 0. Vilar del Hoyo et al. Hernandez-Stefanoni. respectively. Out of 667 cells with > 90% no-forest classes. 2004*) Shape MPAR Diversity PR SHDI Patch Richness Shannon’s Diversity Index 7. and Xi and Bi are the independent variables (the metrics computed in the landscape units) and the estimated coefficients of the model. 2002. Martell et al. and no fire took place in 39. Padilla and Vega-Garcia. 2012*) (Ortega et al. 2002.. Martínez et al.5 (the midpoint of the distribution).. Vega-García et al.. During this study period (1989 to 1993). Henry and Yool. .036 10.565 0. this value is arbitrary and depends on the model goals or the user interests (Jamnick and Beckett. Martínez et al.846 0.751 observations in Spain. Lloret et al.Fire occurrence and landscape in Spain 75 Table 2. Composition seemed influential but not determinant: for instance. 2004*. 2008.534 0. Stolle and Lambin. and it was also chosen for this work. 2004. 2003. at least one human-caused forest fire occurred in 60. 2002. 2005.217 0. 544 had fires (71%).. 2012*) (Henry and Yool. 2012*.875) of the landscape units. Logistic regression requires fewer statistical assumptions than linear.5% (2. 2009. 2004*. The cut-off point of the logistic function is usually set by default to 0. there were 213 with fires (31%) vs 454 without fires (69%). based on a more or less extensive list of independent variables related to the event studied (Equation [1]). 2011. Ortega et al. References where the variable or a similar factor were used Density PD Patch Density 0.839 0... dev. 1987.335 0. Table of all landscape metrics considered in the study Group Abrev. 228 had not (29%).. However.258 SIDI SIEI * References which have used landscape metrics to predict forest fire occurrence. Description Mean Std.682 1.341 6..114 2. The only two landscapes > 90% urban (Madrid) had fires in the study period. 2012*) Area MPS Mean Patch Size 5. 2009. [1] where P is the probability of an event happening (wildfire). 2004) (Henry and Yool. 2012*. Logistic regression models can estimate or predict the probability P that a dichotomous or binomial variable occurs or not.162 1. Statistical analysis: logistic regression Logistic regression has been frequently used to predict fire occurrence (Chuvieco et al. Ortega et al. Romero-Calcerrada and Perry. Ortega et al.721 MedPS Median Patch Size 1...967 2. 140 cells > 90% agriculture had fires..5% (1. 2008) (Ortega et al. 2009*) ED MPE MSI AWMSI Edge density Mean Patch Edge Mean Shape Index Area-Weighted Mean Shape Index Mean Perimeter-Area Ratio 0.327 0. The landscapes for analysis were sufficiently large (100 sq km) and diverse to include all Table 1 classes under different spatial arrangements.502 SHEI Shannon’s Evenness Index Simpson’s Diversity Index Simpson’s Evenness Index 0. Lloret et al.804 (Henry and Yool...040 3. 2008). 2004*) 0. 2004) (Martínez et al. Lloret et al. The only landscape with 40% water had experienced fire. 2012*) (Henry and Yool. 2005.

. plus shape characteristics split into three groups. patch density (fragmentation) and vegetation diversity. Expected R values for randomness should be close to one. the Nagelkerke R 2. 1999. Thus. / Forest Systems (2013) 22(1). Median Patch Size (MedPS) had a moderate correlation to MPE. 1999). Mean Patch Size (MPS). and selected only one at a time for model building trials. but not to MPS and PD.. within an interval ranging from 0 to 2. Regarding shape. The adjustment method for the logistic regression model was the forward stepwise approach. 1995). which proceeds by adding variables with statistical significance (p-value less than 0. The decision on the level of maximum likelihood involves usually predicting correctly both (Stolle and Lambin. SHDI. we tested clustered or dispersed conditions of the over and underestimated errors (false alarms and missed f ires) with the Average Nearest Neighbor Distance Index (ANND) (Martínez et al.05) one by one (Hair. although MSI is more influenced by the area of the observation unit. In order to evaluate the spatial distribution of errors. SIDI and SIEI) in the same group. most not-normally distributed. 2009). The low correlations of MedPS and Mean PerimeterArea Ratio (MPAR) with the other metrics discouraged their grouping with any other landscape metrics. diversity) and their Spearman correlation. their individual capability to predict human-caused f ires occurrence was tested by building one-variable models. and compares their distance mean with the expected mean distance that would occur for a random distribution. more demanding than the backward stepwise approach. shape. Correlation between PR and SHEI and SIEI was low. 2006). We grouped the variables according to their landscape feature typology (size. Patch Density (PD) and Edge Density (ED) were strongly correlated. SHEI.05) (Hair.. an excessive amount of variables reduces the residual errors but makes fitting the equation more difficult (Martínez et al. Silva and Barroso. As should be expected. Both have a similar behaviour. The validation results were evaluated using the classification table and the Kappa statistic (Congalton and Green. the Hosmer-Lemeshow test and the percentage correctly predicted in the classification table. Silva and Barroso. We included the five variables (PR. The number of variables is important when dealing with logistic regression. the correlation between the Mean Shape Index (MSI) and the Area-Weighted Mean Shape Index (AWMSI) was moderate. density. 2003. The overall f it of the model was evaluated by the -2LL value. On the contrary. but these indices showed good correlation with other diversity metrics.939 between Shannon and Simpson’s Indices). Also. Vega-García et al. All diversity metrics were highly correlated (r-values over 0. the significance of the dependent variables was assessed using the Wald statistic and its statistical significance (p-value less than 0. Uncorrelated groups of correlated metrics Size Density Shape Diversity Group 1 Group 2 Group 3 Group 4 Group 5 Group 6 MPS MedPS MPE PD AWMSI MSI ED MPAR SIDI SIEI SHEI SHDI PR . all metrics considered were classified into six groups (Table 3) depicting landscape patch size. Table 3. A small number of variables introduced in any model make it simpler. based on a Spearman’s correlation analysis between all independent variables. Only uncorrelated variables from every metric group and with signif icant relationship to fire occurrence entered the model building process. 71-81 1987).76 S.14. Results Results of the variable selection The Spearman correlation values between variables are presented in the Appendix. A variable selection process was carried out before modeling the relationship between fires and landscape metrics. This index examines the distances between the centroid points of the closest misclassified quadrants. MPS was inversely correlated to ED and PD. Model fit and validation The database for analysis was divided randomly in two groups: 60% of cases were used to adjust the logistic regression function and the remaining 40% were reserved for validation.427 with Patch Richness and over 0. calculated using SPSS 15 (SPSS Inc. Costafreda-Aumedes et al. 1999. and the appearance of high errors in the formulation or nonsignificant values is more likely. In addition. 2004). 2009). 2004).

3a).67) and MSI (Wald = 5.3%.141).007).092 280.900 . R2MSI = 0. Both general spatial trends agreed with historical forest fire records from the Ministry of Environment in Spain (MAGRAMA). if any. MPAR and AWMSI showed low human-caused fire occurrence predictability in the univariate logistic regression analysis (Nagelkerke R2 less than 0. Mean Patch Edge (MPE) and Mean Shape Index (MSI).001).7% and the percentage of correctly predicted and observed fires was 68. which balanced the percentages of correct matches of the landscape units with fire (Y = 1) and no fire (Y = 0) occurrences.052. We also regarded previous use in the literature. The overall percentage of correct predictions was 66.61) Model building data Observed Validation data Predicted Observed Predicted No-fire Fire No-fire 742 565 Fire 397 1.073 1.147 Total 1. Results in the classification table (Table 5) for the validation data (40% of the initial data) were similar to those obtained with the model building dataset. The cut-off point applied was 0. 2004.085 0.431 –0. Table 4. 2004) and we wanted the results of this study to be comparable. R 2 PD = 0.6% (with an overall percentage of correct predictions of 66.121. Romero-Calcerrada and Perry.97).181 0. Patch Density (PD). 65.111. Guadalquivir). The analysis of signs of the β coefficients indicated that the highest probability of human-caused forest fire occurrence occurs with high values of the SHDI and with low values of MPE or MSI. Results of the logistic regression The best model included three variables: Shannon’s Diversity Index (SHDI).544 2. all significant with p-value less than 0. we selected the most significant variable in terms of individual prediction of the humancaused forest fires occurrence. Lloret et al.221 0... The selected metrics were four: Mean Patch Edge (MPE). comparing predicted and observed fire occurrence. Shannon’s Diversity Index predictive capability in the one-variable models was the highest among all variables (Nagelkerke R2SHDI = 0. Table 4 lists the estimated coefficients and variables of this model. R2SHDI = 0.001) Variables β ET Wald Exp(β) SHDI MPE MSI 1. Ortega et al. the f itted equation was used to map the correct human-caused forest fire occurrence predictions for the 10 × 10 sq km landscape units in the period 1989 to 1993 (Fig.065 –0. followed by MPE (Wald = 24. Classification table of logistic regression (cut-off point = 0.80).61.139 1.163 Total 1. The spatial representation Table 5.7% and only 80. The Shannon’s Diversity Index had been the most widely used metric in studies of landscape diversity (Henry and Yool.Fire occurrence and landscape in Spain Within each group. These uncorrelated metrics best explained f ire occurrence within their groups in the one-variable models built (Nagelkerke R 2 MPE = 0. and the Nagelkerke R2 was 0.1% the probability of forest fire occurrence. Lastly.016.966 24.013 0. MPS. 2012.1% for no-fire and 67% for fire observations. The percentage of correctly predicted no-fire observations was 62. Estimated coefficients and significance values of the best logistic regression model (p-values for the three variables < 0.665 5.023. MedPS. In general. the model identified landscape units with higher fire occurrence probability in Northwest areas and in the Mediterranean coast. Mean Shape Index (MSI) and Shannon’s Diversity Index (SHDI).796 4. Also.801 77 Interpretation of the Wald statistic indicated that SHDI was the variable with greater weight in the adjusted model (Wald = 280. 2002.154).224.851 Total 827 1. The analysis of exp(β) conf irmed this since a unit increase of the Shannon Diversity Index increased by 418. R2ED = 0.307 1.3%). while the unit chance of the MPE meant a decrease of 93.937 0.1% for the MSI. The p-value of the Hosmer-Lemeshow test was significant (p-value < 0.712 No-fire Fire No-fire 462 365 Fire 275 798 Total 737 1. A classification table (Table 5) was used for evaluating the predictability of the model. Agricultural inland valleys with scarcer presence of natural vegetation presented a lesser likelihood of fire (Ebro.

Serra et al. (2012). 2012.3%. RuizMirazo et al.. SHDI and MPS were found to have significant effects on wildfire occurrence in the period 1985-1998 by Ortega et al. such as MPAR and AWMSI. 2007. Atlantic/Mediterranean). We knew other environmental or socioeconomic factors affected humancaused f ire occurrence (Díaz-Delgado et al. 3b) of misclassif ied predictions did not show a clear pattern indicative of a specif ic geographic trend (North/South.570.50) and underestimation errors (missing fires) were aggregated in areas with greater diversity (0. 2001. Padilla and VegaGarcía. Costafreda-Aumedes et al. This value was almost identical to that obtained with the validation sample (66.647 and the ANND COMMISSION z-score value was –14. There were three significant variables in the model: Shannon’s Diversity Index (SHDI). were significant in the analysis with Landsat TM images. Mean Patch Edge (MPE) and Mean Shape Index (MSI).3%). the low value of Nagelkerke R2 (0.58). Correctly (a) and incorrectly (b) classified landscape units of human-caused forest fire occurrence.. in line with the statement by Forman (1995) that two or three well-selected landscape metrics should be sufficient to answer specific questions on landscape processes..224) pointed at the fact that a large portion of the dependent variable variance was not explained by the fitted model. The reduction in the number of variables to include in the fitting of the logistic regression allowed to respect the non-collinearity assumption and made the model more parsimonious. Fuller. Sturtevant and Cleland.47).78 S. 2007). Löfman and Kouki. but it was not our purpose to evaluate those factors in this study. Discussion The probabilistic relationship between landscape metrics and human-caused fire occurrence could be modelled and was found to be significant in Spain. However. which indicated the model robustness. (2009) tested only three landscape metrics (Fragmentation using a 7 × 7 kernel on the Corine Land Cover 1990 grid reclassified into four classes.76) than in the identified as no-fire-prone (0. These results agreed with previous studies that made use of landscape metrics as proxies for the impact of human activities on the territory (Echeverría et al. both significant (p-value < 0. 2003.. 2011. 2004. Fine-grained forest-agriculture mixtures and road density had significant effects in all periods (1974-2008) in their forest-agriculture interface landscapes. but the ANND OMISSION z-score value was –9.001). This should be expected.. 2008. with overall percentage correctly predicted of 66. 2008). Henry and Yool (2004) determined that SHDI and MSI explained some of the variability of f ire occurrence in Arizona from remote sensing images in a fusion of SIR-C and Landsat TM images. / Forest Systems (2013) 22(1). These selected variables were also found significant in other studies. Patch Density and . Other variables. The study by Martínez et al. (Fig. Romero-Calcerrada et al. but lower than in the fireprone areas correctly classified (1. 71-81 a) b) Figure 3. Overestimation errors (false alarms) were aggregated in locations with high diversity (mean SHDI 1. The results of the classification table suggested a moderate predictive capability of the best model.

Xu C. Global Ecol Biogeogr 16: 426-439. Martínez et al. the sharing of edges between roads and agricultural areas (Martínez et al. Congalton RG. where large extensions of croplands exist and natural vegetation is scarce. Bu R. (2009) and Padilla and Vega-García (2011). 2007. Echeverría C. must be considered in fire occurrence models in Spain. this likelihood of occurrence was greater in landscape units with fewer edges and with more compact patches. Int J Wildland Fire 16: 34-44. and the authors obtained better results in predicting overall fire occurrence in Spain (model building data: 85. 2006) because. Bishop I. Misclassified units (errors) were found scattered throughout the Spanish territory (Fig. Hu Y. according to Martínez et al. Ecol Ind 8: 672-685. Boca Raton. Li X. Lara A. 2010). González I. Benayas JMR. Díaz-Delgado R. 2008). due mainly to a fractured topography. Landscape Ecol 19: 731-745. 3a) agreed with that of Martínez et al. 2009. Chuvieco E. 2004. Ministry of Environment and Rural and Marine Affairs (MAGRAMA). Int J Wildland Fire 15: 187-196. Newton AC. and specifically diversity and shape. Lloret F. and validation: 76. Chuvieco E. Chang Y. NE. a scenario of significant urban development linked to tourism and the influx of population in the summer overlaps with dry weather to increase fire risk levels (Vilar del Hoyo et al.2%) by including socio-economic factors in their model. Aguilera P. Our analysis concluded that in the 10 × 10 sq km units with greater landscape diversity the probability of human-caused forest fire occurrence was generally higher. Modelling of landscape changes derived from the dynamics of socio-ecological systems. 2009). Results in previous studies indicate that it is not possible to have good wildfire predictions taking into account only landscape structure. 1. Lewis Publications. These clusters respond to the presence of local conditions that influence the occurrence or absence of fire. high rainfall and humidity (Fig. Pallares-Barbera M. Landscape diversity was the main factor in predicting human-caused forest fires in this study. 2006. Green K. There is also agreement in the Mediterranean coast (Coastal Catalonia and the Baetic Ranges). but their distribution was locally aggregated. Acknowledgements The authors gratefully acknowledge the provision of fire occurrence historical data from the National Forest Fire Statistics database (EGIF). Coomes DA. Assessing the accuracy of remotely sensed data: principles and practices. He HS. SA. there it is greater the consistency in the relationship of the fitted model between landscape structure and fire occurrence.Fire occurrence and landscape in Spain MedPS) but only agricultural land fragmentation was selected for their model. A case of study in a semiarid Mediterranean landscape. De Aranzábal I. Int J Wildland Fire 18: 430-441. The map obtained by applying the fitted equation (Fig. S1). 3b). Is in these areas where most of the human-caused fires occur in Spain. and with more recent fire data. Schmitz MF. Spain. 2008. Also. (2009) found that landscape metrics showed comparatively lower significance compared to socio-economic changes in rural and urban areas and 79 traditional activities associated with f ire.. and consequently. References Badia-Perpinya A. Spatial distribution of ignitions in Mediterranean periurban and rural areas: the case of Catalonia. Ediciones RIALP. once the Forest Map of Spain MFE50 (Andalucía not available yet) and MFE25 (expected 2017) are published. Impacts of forest fragmentation on species composition and forest structure in the temperate landscape of southern Chile. 2007. Long-term forest landscape responses to fire exclusion in the Great Xing’an Mountains. These characteristics are common in humanized environments (Badia-Perpinyà and Pallares-Barbera. Verdú F. The landscape configuration of the Atlantic zone is characterized by small and highly fragmented patches with high diversity of species. Aguado I.. but landscape pattern variables.4%. for example. Yebra M. In the Northwest these landscape characteristics are associated with risk factors such as a traditional use of fire to obtain open areas for increasing pasture land and the low profit from forests by local people (Torre Antón. 1999. The areas with greater agreement between observed and predicted values in the model are given in the Atlantic North of Spain. 1996. Pineda FD. China. Pons X. Most of the landscape units without fire occurrence in the period are plains with fertile deep soils where intensive agriculture is the most profitable economic activity: the Ebro and Guadalquivir river basins and the Meseta Central. It would be highly convenient to test these results at different scales. . Spatial patterns of fire occurrence in Catalonia. Fundamentos de Teledetección Espacial. Prediction of fire occurrence from live fuel moisture content measurements in a Mediterranean ecosystem. (2009) at the municipal level.

Area of Defense Against Forest Fires. modeling. San Diego etc.. Wildfires and landscape patterns in the Eastern Iberian peninsula. Chuvieco E. 2007. Hair JF. Stocks BJ. On the comparative importance of fire danger rating indices and their integration with spatial and temporal variables for predicting daily human-caused fire occurrences in Spain. Cambridge University Press. 2001. Councilman JS. Barrera I. Vegetation dynamics under fire exclusion and logging in a Rocky Mountain watershed.1995. Fuller DO. Mediterranean vegetation dynamics: modelling problems and functional types. Inv Agr Sist Rec For 1: 383-398. Frohn RC. Relationships between landscape patterns and species richness of trees. 2004. Jiménez JJ. Monte DK. Northwest Science and Technology. Pausas JG. Martell DL. Springer-Verlag. Theobald DM. 2007.3 GIS. CRC. 2002. A mathematical model for predicting fire spread in wildland fuels. Hernández-Stefanoni JL. Virginia. Carr AP. 2010. Salas J. Millington JDA. 2010]. Martínez J. Landscape-scale patterns of forest pest and pathogen damage in the Greater Yellowstone Ecosystem. Landscape Ecol 17: 745-759. Gallant AL. Estudios Geograficos: research/fragstats/fragstats. 2004. The relative dependence of Spanish landscape pattern on environmental and geographical variables over time. 2011. Vega-García C. Jamnick MS. Simulating Mediterranean landscape pattern and vegetation dynamics under different fire regimes. New York. Pausas JG. Agroforest Syst 85: 331-349. Multivariate data analysis. 2008. Kouki J. 2003. Hatala JA. Vettorazzi CA. Vega-García C. Landscape Ecol 16: 627-642. Sánchez JE. Plant Ecol 187: 249-259. Bunce RGH. 2010. Moorcroft PR. Rothermel RC. Categories and types for sustainable forest management reporting and policy. Land mosaics: the ecology of landscapes and regions. Calvo E. / Forest Systems (2013) 22(1). Ogden. Riaño D. Landscape vulnerability to wildfires at the forest-agriculture interface: half century patterns in Spain assessed through the SISPARES monitoring framework. Ortega M. 1999. 1994. 2007. CA. Belbahri L. Lieberman AS. 2002. 1985. Academic Press. Landscape Urban Plan 66: 217-232. Ferraz SFDB. Landscape Ecol 23: 341-354. Beckett DR. Pickett ST. Human-caused wildfire risk rating for prevention planning in Spain. 1856-1996. 1987. 1998. Madrid. Cambridge etc. Padilla M. 1999. Giménez S. Can J For Res 18: 330-336. Can J For Res 17: 394-401. Geocarto Int 19: 25-37. Lefort F et al. Castejón M. . Forest Fires in Spain 2008. ESRI. 2006. González-Ávila S. Gómez-Sanz V. Principles and methods in landscape ecology: towards a science of landscape. A logistic model for predicting daily people-caused forest fire occurrence in Ontario. 1972. Amherst. montes-y-politica-forestal/mapa-forestal/ [6 October. European forest types. 4ª Conferencias Internacionales sobre Incendios Forestales. Novillo CJ. Farina A. PrenticeHall. García del Barrio JM.marm.html. MAGRAMA. 5ª ed. Tella G. Crabtree RL.umass. 2nd ed. Remote sensing for landscape ecology: new metric indicators for monitoring. Forest Ecol Manag 175: 247-252. ElenaRosselló R. Lloret F. Geostatistical Analyst. Cambio global e incendios forestales: una visión desde España. Scale and dynamics of a transforming forest landscape. Moreno JM. White PS. Díaz-Delgado R. Remote Sens Environ 114: 375-384. 2009. USA evaluated with multitemporal Landsat imagery. A logit analysis of private woodlot owners harvesting decisions in New-Brunswick. Dordrecht. Ene E. Using indicators of deforestation and land-use dynamics to support conservation strategies: a case study of central Rondônia. A tool for quantifying landscape structure. ArcGis 9. Romero MA. FRAGSTATS: spatial pattern analysis program for categorical maps. Central Spain 1984-1999. Copenhagen. Patch Analyst User’s Manual. 1999. 2005. Romero-Calcerrada R. Betz DW. Sevilla. 2003. J Phytopathol 155: 289-295. Löfman S. 2008. J Environ Manage 90: 1241-1252. Plant Ecol 179: 53-65. Available at the following web site: http://www. Naveh Z. Ecol Appl 13: 385-403. Ontario. For Ecol Manage 257: 1586-1595. Pons X. Landscape ecology: theory and application. Inv Agr Sist Rec For 17: 114-129. 1987. Denmark. Plant Ecol 140: 27-39. 1999. Regato P. Ortega M. Henry MC. GómezJiménez I. Rempel RS. Fuel type mapping from remote sensing data: contributions to a sustainable environmental development. EEA. Springer. Otukol S. 2001. Mapa Forestal de España [on line]. 13-17 de mayo de 2007. MAGRAMA. USA. Assessing relationships between forest spatial patterns and fire history with fusion of optical and microwave remote sensing. España. Cushman M. Int J Wildland Fire 20: 46-58. Forest fragmentation in Loudoun County. 2011. Computer software program produced by the authors at the University of Massachusetts. Saura S. Redlands. Brazil. GIS analysis of spatial patterns of human-caused wildfire ignition risk in the SW of Madrid (Central Spain). Halligan KQ. 2009. Res Pap INT-115. The role of land abandonment in landscape dynamics in the SPA ‘Encinares del río Alberche y Cofio. Forman RTT. Cambios recientes en los paisajes de los sistemas forestales mediterráneos de España. 2006. EEA (European Environment Agency). ElenaRosselló R. The ecology of natural disturbance and patch dynamics. Available in: http://www. Yool SR. New Pythium taxa causing root rot on Mediterranean Quercus species in south-west Spain and Portugal. Costafreda-Aumedes et al. Trapero A. Hansen AJ. Canada. 71-81 Elkie PC. ESRI (Environmental Systems Research Institute) Inc. Chuvieco E. shrubs and vines in a tropical forest. Perry GLW.80 S. and assessment of ecosystems. McGarigal K. 2009. ElenaRosselló R. Romero-Calcerrada R.

Cleland DT. 2010.541 1 . III.271 –0. Lambin EF. Indonesia.051 SHEI 0. SPSS 15: Command Syntax Reference.985 0. Can J Remote Sens 31: 153-166. Pons X. Viegas DX.007 AWMSI MPAR 1 0. Empleo de técnicas de regresión logística para la obtención de modelos de riesgo humano de incendio forestal a escala regional. Ferreira A.569 0.204 –0. Sol B.476 –0. J Environ Manage 98: 43-50. Landscape Ecol 21: 595-605. Madrid. Memoria general. Int J Wildland Fire 9: 235-246.162 0. La situación actual de los incendios forestales en España.463 –0.540 –0.269 –0.102 0.Fire occurrence and landscape in Spain UT: US Department of Agriculture.017 0. Conferencias y Ponencias del 5º Congreso Forestal Español.220 0. Boletín de la AGE 47: 5-29 Zhang Y.246 –0. 2003. Vilar del Hoyo L.149 0. Int J Wildland Fire 16: 398-413.293 –0.205 –0.490 –0. Appl Geogr 28: 189-209. Saurí D. Chicago.454 1 MedPS –0.245 –0.583 –0. 2005. Guindon B. Ruiz-Mirazo J. Int J Wildland Fire 5: 101-111.276 –0.485 –0.036 SIEI 0. SPSS Inc. Mapa Forestal de España. 81 Sturtevant BR.976 0.175 0.003 0.184 0.966 0.481 –0.495 –0. Appendix.499 –0.991 0.003 –0.179 0.456 0. Silva LC.493 MPAR 0.253 1 AWMSI 0.735 1 0.122 1 0. Bovio G. Titus SJ. Int J Remote Sens 23: 4853-4880. Vega-García C.473 1 0. Stolle F. Pastoral wildfires in the Mediterranean: understanding their linkages to land cover patterns in managed landscapes.114 –0. 1990. Environ Conserv 30: 375-387. Spearman’s correlation matrix of all landscape metrics PD MPS MedPS ED MPE MSI PD 1 MPS –0.112 0. Martín Isabel MP. 2006.258 –0. Serra P. Nosenzo A.415 –0.176 0.553 –0. Interprovincial and interannual differences in the causes of land-use fires in Sumatra. Landscape analysis of human impacts on forest fragmentation in the Great Lakes region.728 0. 2004.010 0. Cuad Soc Esp Cienc For 31: 179-195.037 SIDI 0. Barroso IM. La Muralla.479 –0. 2002. Martínez Vega FJ.173 0.189 0. Lee BS.333 –0. Torre-Antón M. Saura S. 1995. Effects of minimum mapping unit on land cover data spatial configuration and composition. Adamowicz WL. Land-cover and land-use change in a Mediterranean landscape: a spatial analysis of driving forces integrating biophysical and human factors. Applying local measures of spatial heterogeneity to Landsat-TM images for predicting wildfire occurrence in Mediterranean landscapes.931 0. Martínez-Fernández J.005 –0.262 –0.077 1 MPE –0.017 –0. Vega-García C. Vega-García C. 2008. Intermountain Forest and Range Experiment Station. 2012.353 –0.003 0. Comparative study of various methods of f ire danger evaluation in southern Europe.006 SHDI SHEI SIDI SIEI PR 1 0. 2008. Ruiz de la Torre J. Human and biophysical factors influencing modern fire disturbance in northern Wisconsin. 2010.633 1 0. Del rodal al paisaje: un cambio de escala. Conferencias y Ponencias del 5° Congreso Forestal Español.014 0. Cuad Soc Esp Cienc For 31: 213-239. 2006.002 0.014 SHDI 0.190 0. A logit model for predicting the daily occurrence of human caused forest fires.030 PR 0. 1999.162 1 MSI 0. Woodard PM.002 0. Saura S. Chuvieco E.180 0.849 1 ED 0.134 0.947 0. 2007. nuevas perspectivas para la planificación y ordenación forestales.227 –0. Regresión logística.527 –0.