Modelling wildfire activity in wildland–urban interface (WUI) areas of Sardinia, Italy

Study area

Sardinia (Italy) is located in the western part of the Mediterranean Sea (38°51′N–41°15′N, 8°8′E–9°50′E). With about 24,000 km² of land, Sardinia is the second largest island in the Mediterranean Basin. The island is composed of 377 municipalities distributed in six provinces. The population, about 1.7 million inhabitants, is concentrated around the metropolitan areas of Cagliari and Sassari; the human presence increases from late spring to early fall due to the touristic flows, which are mostly concentrated in the coastal areas.

The physical geography of the island is complex, with a significant presence of hills and low mountains, and quite heterogeneous conditions in both geological and morphological features. The climate is mostly Mediterranean, with the highest temperatures, the lowest cumulated rainfall and an important water deficit observed during the summer months, particularly in the southern portions. Annual precipitation varies from ~400 mm in the coastal areas to ~1100 mm in the mountains. The mean annual temperature ranges between ~8°C in the mountain areas to ~17°C along the coast; during the summer season maximum temperatures often exceed 30°C (Fick and Hijmans 2017; Salis et al. 2023). According to the Sardinian Land Use Map (2008) (UDS_2008), about 41% of the island is covered by grasslands, pastures, and agricultural lands. Mediterranean maquis and garrigue cover about 27% of land and are mainly represented by Pistacia lentiscus L., Arbutus unedo L., Erica arborea L., Myrtus communis L., Olea europea L. var. sylvestris Brot., Phyllirea spp., Juniperus spp., Cistus spp. and Euphorbia spp. Woodlands and broadleaf forests, mostly characterised by Quercus spp. (Quercus ilex L., Quercus suber L., Quercus pubescens Willd., and Quercus congesta Presl), occupy 21% of the island; Castanea sativa Mill., Taxus bacata L. and Ilex aquifolium L. are the main forest species at higher altitudes. Conifer stands cover 2% of the island, and are primarily represented by Pinus pinea L. and Pinus halepensis Mill.

During the period 2005–2019, Sardinia registered approximately 44,300 wildfire ignitions and about 223,300 ha burned (Fig. 1a), with a large interannual variability. On average, 15,000 ha burned every year, although in years with extreme fire-weather conditions annual area burned doubled or tripled that average value (e.g. in 2009, about 44,800 ha; in 2007, about 34,400 ha). Wildfire impacts were driven by few large wildfires: in the study period, 264 events (about 0.6% of the total wildfires) exhibited an area burned larger than 100 ha, and accounted for 56% of the total area burned. The years that experienced the highest number of wildfires were 2014 and 2017 (about 4,800 and 3,700 fires, respectively). Most wildfires occurred in the May-September timeframe, corresponding to about 90% of the total area burned. Typically, near 95% of wildfires exhibit anthropogenic causes. Previous studies on historical fire regimes in Sardinia indicate that small fires are mainly generated from ignitions near coastal or in flat areas, particularly in the southern Campidano plain, while large wildfires are more often associated with high fuel load, complex topography and hilly conditions (Salis et al. 2021; see Supplementary Figs S1a, S2a).

Fig. 1.

(a) Wildfire ignitions as a function of their final size observed in Sardinia in the study period 2005–2019; (b) municipal percent area occupied by WUI, and anthropic area polygons (in black colour).

Identification and delineation of WUI

WUI can be defined and mapped in different ways that usually depend on the scale of analysis. The two main WUI mapping methods are: (1) the point-based approach (Lampin-Maillet et al. 2010; Bar-Massada et al. 2013) and (2) the zonal approach (Radeloff et al. 2005). These methods differ in the geospatial data that are used to determine the locations of buildings throughout the landscape and to calculate the building density (Bar-Massada 2021). Overall, the interface area can be identified as the boundary strip between urban and wildland use or as the intermix region encompassing buildings and residential areas scattered within forest and rural lands. In this work, we defined WUI as the 100-m buffer zone around anthropic values; a 100-m buffer was adopted in previous works carried out in Sardinia and in other Mediterranean areas (Galiana-Martin et al. 2011; Bouillon et al. 2014; Sirca et al. 2017) and falls within the WUI boundary ranges (50–200 m) used in Italy (Modugno et al. 2016). For this purpose, we intersected the 2008 Sardinian Land Use and the Corine Land Cover shapefiles (EEA_2012 - Corine Land Cover 2012) and aggregated all anthropic land covers of ‘class 1’ (urban, industrial and commercial); starting from this layer, we produced a buffer of 100 m. With this analysis, we obtained a WUI area that included the rural, natural and forest lands around anthropic values. Sardinian WUI resulted in about 226,800 ha, and ranged from 22 to 10,700 ha. The biggest towns and the main villages located in the coastal areas presented more than 20% of the municipal territory characterised by WUI (Fig. 1b). The municipalities with the highest percentage of municipal area covered by WUI were in the southern part of Sardinia.

Socio-economic, vegetation, climatic and zootechnical dataset

We collected data on socio-economic, vegetation, climatic, and zootechnical variables, to create a comprehensive dataset at municipal and WUI levels: for each municipality and WUI, we created a dataset of 311 independent variables and two dependent variables (the area burned and the number of ignitions). Within this dataset, as pointed out in next lines, we removed 288 redundant variables and we focused our analysis considering 23 independent variables (Table S1).

Historical climate data for the period 1970–2000 were gathered from the WorldClim global climate database (Fick and Hijmans 2017) and aggregated as mean or total value at the municipal level. Vegetation layers were derived from the Corine Land Cover Map 2012 (EEA_2012 - Corine Land Cover 2012). The socio-economic information was retrieved from the 2011 data from ISTAT census (ISTAT; http://dati-censimentopopolazione.istat.it/Index.aspx#) (Table S1). For the zootechnical variables, we collected information about the total livestock farms and the number of livestock in each municipality from the National Veterinary Registry (Vet_info, Sistema Informativo Veterinario, Statistiche; https://www.vetinfo.it/j6_statistiche/#/new-list).

Due to effect of terrain heterogeneity and topography on wildfire occurrence and size, and the potential inherent effects on wildfires at WUI, we carried out our modelling analysis at the whole regional scale as well as splitting Sardinian municipalities into two groups based on the average altitude and on the elevation range: (1) the ‘Flat’ (F) group, which composed of 181 municipalities with mean altitude <350 m a.s.l. and elevation range <600 m; (2) the ‘Hilly and mountainous’ (H) group, which composed of 196 municipalities, with mean altitude ≥350 m a.s.l. and elevation range ≥600 m (Fig. S3).

Wildfire dataset

We used the 2005–2019 datasets of wildfire ignitions and area burned perimeters provided by the Sardinia Forest Service. The ignition data includes information about the location, date, municipality of origin, and wildfire size of each event. Wildfire perimeters contain spatial information on the area burned by wildfires larger than 0.1 ha.

To analyse wildfire activity inside Sardinian WUI in the study period, we intersected WUI areas with the abovementioned wildfire data. This allowed to create a dataset of wildfire ignitions (IP_WUI) and area burned (AB_WUI) within each WUI for all municipalities.

Data processing for machine learning modelling

Input data were first normalised to remove the impact of dimensions on models’ predictions, then randomly divided into training set (70%) and testing set (30%). Prediction models were produced by using the training sample, while the accuracy of predictions was evaluated using the testing set.

We performed multicollinearity diagnosis by Pearson’s coefficients, and data cleaning by identifying and removing outliers. Variables with a Pearson’s coefficient higher than 0.90 were excluded from the analysis. The Pearson’s coefficients were calculated at the regional level, and separately for the F and H municipal groups. Within the dataset, 288 redundant variables were removed due to collinearity; a set of 23 independent variables (Table S1) were used to develop the prediction models (correlation matrixes are in Figs S4, S5, and S6).

We used the caret package (Kuhn 2008) in R to build the predictive models. To assess and predict wildfire activity at WUI we used a set of ML models (Table S2). The performances of these models were analysed according to the Pearson’s correlation coefficient (r), the coefficient of determination (R²), and the error indicators Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE).

In the model training phase, control parameters were set, including cross-validation with 10 repetitions to evaluate the model stability. The specific hyperparameters of each model were not objectively set, but a tuneLength parameter was set to five. The tuneLength parameter instructs the algorithm to try different predefined values for the main parameters and finally compares and uses those that attained the highest performances during the cross-validation.

Assessment of the variable importance

We then quantified the variable importance for the best ML models, for both area burned and wildfire ignitions.

The importance of each single variable was assessed through the ‘varImp’ function implemented in the caret package, which calculates and returns the relative importance of each variable with respect to that with the highest raw importance. This is calculated by keeping track of the combinations of variable weights that allow obtaining the highest prediction accuracy through the minimisation of a cost function, obtained from a modification of the Garson’s algorithm to deduce the influences of variables used in algorithms with neural networks (Garson 1991). In this algorithm, the weights associated with each variable are evaluated based on the variance explained by each one.

Mapping modelled wildfires at WUI

The estimated area burned and wildfire occurrence were displayed as maps that show the average annual wildfire patterns at WUIs. Predictive maps were created separately for both AB_WUI and IP_WUI at the regional scale and for the F and H municipal groups. Finally, we mapped the differences between estimated and observed data to support the spatial evaluation of the model predictions.

Observed wildfire activity at WUI

During the study period, wildfires at WUI represented about 6.7% of the total area burned, and 27% of the total ignitions occurred in Sardinia. On average, about 818 wildfires occurred every year inside WUI, and approximately 995 ha of WUI annually burned. The highest AB_WUI values were observed in 2009 and 2007 (about 2,500 and 1,900 ha burned, respectively), whereas the peaks of IP_WUI were recorded in 2014 and 2017 (about 1,500 and 1,200, respectively). The annual average AB_WUI at the municipal level ranged from 0 to 32.4 ha year⁻¹ observed in north-eastern Sardinia (Fig. 2a). Overall, 214 Sardinian municipalities exhibited an annual AB_WUI lower than 1 ha year⁻¹, whereas only 18 municipalities, mainly located in the southern part of the island and in coastal areas, showed values higher than 12 ha year⁻¹. In a few municipalities, particularly in southern Sardinia, more than 40% of the area burned at the municipal level was located inside WUI (Fig. S1b). The peaks of annual AB_WUI were in 2009 and in 2007 in central Sardinia (388 and 276 ha burned, respectively). On average, the H group exhibited a slightly bigger annual area burned than the F group (513 vs 482 ha year⁻¹).

Fig. 2.

(a) Observed annual area burned and (b) observed annual wildfire ignitions at WUI in Sardinia, Italy for the period 2005–2019.

The average annual IP_WUI was 2.2 # year⁻¹, and ranged from 0 to 45.3 # year⁻¹ in southern Sardinia (Fig. 2b). Annual IP_WUI values higher than 15 were observed in southern Sardinia and in Olbia (north-eastern Sardinia), while most Sardinian municipalities exhibited IP_WUI <2 # year⁻¹. In 38 municipalities, mainly in southern and north-eastern Sardinia, more than 40% of wildfire ignitions at the municipal level occurred in WUI areas, with peaks higher than 70% observed in three municipalities of southern Sardinia (Fig. S2b). The F group municipalities showed annual IP_WUI slightly higher than the H group (419 vs 399 # year⁻¹). The highest annual IP_WUI were observed during the years 2017 and 2014 in the southern municipalities.

Modelled area burned at WUI

Table 1 summarises the performance metrics provided by each ML model on the test dataset. At the regional level, the Random Forest (RF) model showed the best performances among all models. In fact, RF presented the highest Pearson’s correlation coefficient (r) (0.76), the second highest R² value (0.58), the lowest RMSE (0.10), and the lowest MAE (0.06). The lowest r was showed by SVM-L, LASSO, and ENR models, whereas the highest RMSE values (0.25) were observed in OLS and BRR (for acronyms, see Table S2).

Focusing on flat municipalities, the best performances were observed for the SVM-R model which exhibited the highest r (0.6) and R² (0.36) values, a low MAE (0.08) and the lowest RMSE (0.14). The worst performances were observed for the KNN model, which showed the lowest r and R² (0.40 and 0.16, respectively) and high RMSE and MAE values (0.16 and 0.09, respectively).

As far as hilly WUI are concerned, RF was the best model: in fact, RF presented the highest r (0.82) and R² (0.67), combined with the lowest RMSE and MAE (0.11 and 0.06, respectively). Again, KNN showed the lowest r (0.69) and R² (0.47), and was the less performant model (Table 1).

Considering the above points, we selected RF for modelling AB_WUI at the regional scale and the H group, and SVM-R for the F group.

According to RF at the regional scale, the most important variable that affected AB_WUI was the WUI area covered by herbaceous vegetation (HERB) (100%) (Fig. 3a). On average, HERB cover 44% of the WUI area, varying from 39% in the H group to 50% in the F group; in 54 municipalities, it represents more than 60% of the WUI area, particularly in the western side of the island (Fig. S7a). Other important variables that affected AB_WUI at the regional scale included some socio-economic variables, namely the net taxable income in the municipality (POPINC) (43%), the number of commuters (COMM) (35%), and the resident population living inside urban areas (POPURB) (34%). Within climatic variables, the most important one was the annual average wind speed (WS) (12%). The other variables showed values lower than 9% (Fig. 3a).

Fig. 3.

Variable importance for the best machine learning (ML) models used to assess AB_WUI and IP_WUI at the regional scale (a, d, respectively), and in F (b, e, respectively) and H (c, f, respectively) municipal groups. The best ML models for AB_WUI were RF (a), SVM-R (b), and RF (c), while for IP_WUI were RF (d), RF (e), and SVM-L (f). Variable groups: socio-economic (grey); vegetation (green); climatic (blue); zootechnical (red). The variable acronyms are in Table S1.

For F and H municipal groups, the most important variables influencing AB_WUI were the socio-economic variables, particularly COMM (94% and 100%, respectively), POPINC (100 and 91% respectively), and POPURB (89% and 100%, respectively) (Fig. 3b, c). HERB was also important for both groups (73% and 86%, respectively). The WUI area covered by shrublands (SHR) showed similar importance values for F and H groups (30 and 32%, respectively), whereas the WUI area covered by forests (FOR) presented high value only for the F group (58%). Within climatic variables, the total annual municipal precipitation (PREC) exhibited a high importance value for the F group (60%). The annual average municipal maximum temperature (TMAX), the annual average municipal solar radiation (SRAD) and the number of illiterate people in the municipality (ILLIT) variables exhibited relatively high importance values only for the F group (25, 23, and 42%, respectively).

The observed annual AB_WUI maps (Fig. 2a) were compared with the estimated annual AB_WUI maps (Fig. 4a, b). The estimated annual values at regional scale ranged between 0.1 and 23.9 ha year⁻¹: the highest AB_WUI values were observed in some big municipalities in northern Sardinia and nearby Cagliari. The lowest AB_WUI values were concentrated in the central part of Sardinia.

Fig. 4.

(a) Modelled annual Area Burned at WUI (AB_WUI) (ha year⁻¹) considering the analysis at the regional scale and (b) the separate analyses for H and F municipal groups in Sardinia, Italy.

Overall, the map of the differences between modelledand observed data obtained by the analysis at the regional scale (Fig. 5a) exhibited good predictions for the entire island; the main underestimations of AB_WUI were observed in a south-western municipality (−15 ha year⁻¹) and in some other municipalities (−7 ÷ −9 ha year⁻¹). The highest overestimations of AB_WUI were found in the north-western municipality of Alghero (15 ha year⁻¹).

Fig. 5.

(a) Differences in annual Area Burned at WUI (AB_WUI) (ha year⁻¹) between modelled and observed data considering the analysis at the regional scale, and (b) the separate analyses for the H and F municipal groups in Sardinia, Italy.

The map of the annual average of the modelled AB_WUI conducted separately for the two municipal groups (Fig. 4b) exhibited a range between 0.13 and 23.5 ha year⁻¹. The highest estimated values were in the north-east and the south-west, while the lowest AB_WUI values were located in central Sardinia.

Fig. 5b shows the difference values obtained between the model predictions for the F and H groups and the observed annual AB_WUI. The map exhibited difference values close to 0 for several municipalities. Overestimations were limited to few municipalities and were lower than 6 ha year⁻¹; the highest underestimations were presented by some municipalities located in central (−20.37 ha year⁻¹) and south-western Sardinia (−15 ha year⁻¹).

Overall, both analyses exhibited good predictions of AB_WUI, showing many of the municipalities with limited differences compared to the observed values. The predictions obtained considering the H and F municipal groups exhibited a higher predictive power than that at the regional scale. For both analyses, the highest underestimations of AB_WUI for both the analyses were observed in the municipalities affected by the largest wildfires that occurred during the study period (i.e. Mores, which was impacted by the 10,500-ha Bonorva wildfire occurred in July 2009; Nuoro, which was affected by the 9000-ha Orani wildfire in July 2007). In general terms, the models underpredicted AB_WUI in those municipalities where the observed AB_WUI was high. Conversely, the overestimation errors were mostly slight, apart from the analysis at the regional scale which overestimated AB_WUI in Alghero (north-western Sardinia).

Modelled wildfire ignitions at WUI

Performance metrics for the analysis of wildfire ignitions at regional scale showed the RF as the best model (Table 1). RF presented the highest r coefficient (0.80) and the highest R² value (0.64). The lowest values of R² and r were observed for the KNN model (0.40 and 0.63, respectively). OLS, LASSO, and ENR models presented the highest RMSE value (0.10), whereas the lowest RMSE values were exhibited by SVM-R and RF models. Although the lowest value of RMSE was observed for the SVM-R model, RF showed both the highest R² value and the highest r, which overall made RF the model with the best performances (Table 1).

The analysis of the F group confirmed the skills of the RF model compared to the other models. In fact, RF showed the lowest RMSE (0.06) and MAE (0.03) and the second-highest values of r and R² (0.80 and 0.64, respectively). The ENR model exhibited the highest r and R² values (0.83 and 0.69), but the highest RMSE and MAE values (0.09 and 0.06, respectively).

For H group, the SVM-L model revealed the highest R² and r values (0.96 and 0.98, respectively), and RMSE and MAE values (0.05 and 0.03, respectively) among the lowest observed.

According to the analysis of the model performances, we selected RF for modelling IP_WUI at the regional scale and for the F group, and SVM-L for the H group.

At the regional scale, the socio-economic variables showed the highest influence on IP_WUI (Fig. 3d), particularly POPURB (100%), COMM (86%), POPINC (82%), and ILLITER (33%). HERB was the fifth variable in order of importance (21%).

The separate analysis conducted by RF for the F municipal group and by SVM-L for the H group showed different outcomes (Fig. 3e, f). For both groups, the socio-economic variables represented the most influencing variables, particularly COMM (100% for both), POPINC (36 and 94%, for F and H, respectively), and POPURB (25 and 91%, respectively) (Fig. 3e, f). However, the outcomes between the municipal groups differed when considering other variables, especially for the F group, where some of them exhibited values lower than 5%. Several variables contributed to influencing wildfire occurrence inside WUI of the H group, such as the socio-economic variables represented by ILLITER (87%) and ROADS (21%), vegetation (HERB and SHR (75% and 42%, respectively)), and some climatic variables (WS (41%), TMAX (40%), SRAD (32%), and PREC (21%)) (Fig. 3f).

The map of observed annual IP_WUI (Fig. 2b) was compared with the maps of the estimated values (Fig. 6a, b).

Fig. 6.

(a) Modelled annual wildfire ignitions at WUI (IP_WUI) (# year⁻¹) considering the analysis at the regional scale and (b) the separate analyses for the H and F municipal groups in Sardinia, Italy.

The map of the average annual estimated IP_WUI at the regional scale (Fig. 6a) displayed a range between 0.04 and 30.7 # year⁻¹. In general terms, the highest values for both the analyses were predicted for the main towns of the island. The comparison between observed and modelled IP_WUI at the regional scale exhibited a high predictive power, with overestimations in a few municipalities located in central (13.6 # year⁻¹) and northern Sardinia (up to 12.9 # year⁻¹) (Fig. 7a). In contrast, the main underestimations were in southern Sardinia (−14.5 # year⁻¹).

Fig. 7.

(a) Differences in annual wildfire ignitions at WUI (IP_WUI) (# year⁻¹) between modelled and observed data considering the analysis at the regional scale, and (b) the separate analyses for the H and F municipal groups in Sardinia, Italy.

Regarding the separated analysis for H and F municipal groups (Fig. 6b), the map of the annual average estimated IP_WUI exhibited a range from 0 to 30.5 # year⁻¹. The highest values were in the south (30.5 # year⁻¹). IP_WUI values close to 0 were estimated for several municipalities located in central Sardinia. The differences between observed and modelled data for the two municipal groups F and H showed a better agreement than for the analysis at the regional scale, notably for the municipalities of the F group (Fig. 7b).

For both analyses, the underestimations for IP_WUI were concentrated in the municipalities located in southern Sardinia, that during the study period exhibited the highest number of wildfire ignitions inside the WUI, or in the municipalities that presented the highest values of POPINC, POPURB, and COMM (Fig. S7b–d). Conversely, the overpredictions were more significant for the analysis carried out at the regional scale, and were mainly observed in the most populated municipalities of central and northern Sardinia.

Modelling approach

Although several wildfire prediction models are available worldwide (Jain et al. 2020), they cannot be effective enough to be used for large scales or in different geographical conditions without local calibrations to improve the accuracy of estimations (Sharma et al. 2022). In this work, different regression models and machine learning models were developed to identify the drivers of area burned and wildfire ignitions inside WUI of Sardinia, Italy, using as input a comprehensive standardised dataset of socio-economic, vegetation, climatic, and zootechnical explanatory variables at the municipal level. The results show that, among the ML models used in this work, RF, SVM-L, and SVM-R outperformed the others, exhibiting better performances according to the Pearson’s correlation coefficient and the error indicators. This finding is consistent with the results from other studies on wildfire regime modelling which highlighted that ML models were more competitive than traditional methods like regression models in terms of prediction accuracy (Rodrigues et al. 2014; Milanović et al. 2020; Gao et al. 2024).

The high predictive potential of SVM models can be explained by their power to manage dynamic relationships and complex data, showing also to be an advantageous algorithm for binary classification problems (Pham et al. 2020; Sharma et al. 2022). Good SVM prediction performances were evidenced for susceptibility mapping of forest fires in western Serbia, India, and Himalaya (Gigović et al. 2019; Sharma et al. 2022; Rihan et al. 2023) and for predicting wildfire occurrence probability in China (Gao et al. 2024). Nevertheless, some authors considered SVM less adequate for wildfire occurrence predictions because its calibration and optimisation are considered more time-consuming with respect to other models such as RF (Bar-Massada et al. 2012; Rodrigues and De la Riva 2014).

The good performance of RF can be explained by its high tolerance for data outliers and the capability to handle categorical variables, considering the interactions between the variables (Pourtaghi et al. 2016; Tonini et al. 2020; Achu et al. 2021). RF showed good predictive performances in several wildfire modelling studies. For example, exhibiting a better overall predictive power than linear regression models for the identification of the main explanatory variables for wildfire occurrence and probability mapping in Serbia (Milanović et al. 2020), or outperforming other models in elaborating susceptibility maps in Liguria region (Tonini et al. 2020) and western Serbia (Gigović et al. 2019). In other works, RF outperformed Multiple Linear Regression for the analysis of the spatial patterns of wildfire occurrence in Mediterranean Europe, displaying its ability in the assessment of fire-related phenomena at broad scale (Oliveira et al. 2012; Achu et al. 2021).

Drivers of wildfire activity at WUI

We analysed the role played by several drivers in determining wildfire activity in Sardinian WUI in the study period. By investigating the spatial relationship among wildfires at WUI and many drivers including socio-economic, vegetation, climatic, and zootechnic factors, we were able to point out that given factors can impact differently local-scale wildfire activity according to specific territorial conditions and heterogeneous characteristics of local communities. In this sense, our study aims at identifying (and ranking the importance of) relevant forces influencing wildfires at WUI in Sardinia, and at evidencing how much given factors play a potential role in wildfire activity at the WUI level.

Regarding the role of socio-economic factors, previous works pointed out that wildfire regime is influenced by the socio-economic context at the local (Mancini et al. 2018a) and broader scales (Jain et al. 2022; Jones et al. 2022; Ochoa et al. 2024).

In our work, the net taxable income in the municipality (POPINC) was identified as one of the main drivers of the area burned and wildfire ignitions at WUI. For AB_WUI, this variable exhibited a strong importance for both municipality groups, and a little lower role at the regional level, while for IP_WUI was a key driver at the regional level and for the H municipal group. Low financial resources and budgets of landowners and local population can affect their possibility to support fuel management activities within WUI and in neighbouring lands. Furthermore, public investments in wildfire prevention and forest management may vary based on the municipal availability of financial resources, which could be limited or even insufficient in remote and low-income areas.

The number of illiterate people in the municipality (ILLITER) exhibited a strong influence on IP_WUI. Many authors reported a correlation between wildfire ignitions and educational level; being better informed, people with high levels of education feel confident on how to perceive and act in wildfire-prone areas or situations (Oliveira et al. 2020a), are more able to comprehend and comply with prevention and safeness regulations (Oliveira et al. 2020b), are likely to adopt wildfire mitigation measures (Gan et al. 2015), and show more awareness of and sensitivity towards wildfires, confirming higher caution and care when dealing with them (e.g. renewal of pastures, burning of stubble, releasing flammable waste) (Mancini et al. 2018a; Elia et al. 2022). Moreover, the educational level might be associated to other factors that indirectly influence IP_WUI, such as dwelling locations and conditions, or budget available for fuel management nearby WUI.

Resident population living inside urban areas (POPURB) also played a considerable role in AB_WUI and IP_WUI in the study area. Over the past decades, the migration from the countryside towards the more populated areas, combined with the increase in the number of workers employed in the tertiary sector rather than in the primary sector, traditionally more involved in the control and management of the territory, has led to a progressive abandonment and changes in the landscape structure (Tonini et al. 2018; Ascoli et al. 2021; Salis et al. 2022). In addition, demographic shifts from rural to urban areas also influenced wildfire occurrence and behaviour (Koutsias et al. 2010; Salis et al. 2019; Spadoni et al. 2023). Furthermore, the presence of residential properties intermingled with forested areas can further increase the probability of wildfire activity nearby WUI with respect to ‘classic interfaces’ (Galiana-Martin et al. 2011; Ganteaume et al. 2013); fragmented distributions of residential properties can also lead to more difficulties in managing fuels and lowering wildfire risk.

Another socio-economic variable that influenced wildfire activity at WUI was the number of commuters (COMM) who travel for work, study, or other purposes. Overall, the more commuters are present, the more area is burned inside WUI and the more frequent are wildfire ignitions at WUI. COMM can therefore likely increase the probability of anthropic ignitions (either accidental or intentional), which in the Mediterranean basin represent most of the total wildfire ignitions. These results support the findings of previous works which selected the ratio of commuting workers in total population as one of the most relevant predictors influencing positively the average fire size in Italy (Mancini et al. 2018a). Our abovementioned findings about the influence of socio-economic variables on AB_WUI agree with previous works that found a strong influence of the human presence and anthropic activities on the area burned, particularly at regional and local scale (Aldersley et al. 2011; Curt et al. 2016; Mansuy et al. 2019), and an increasing influence on wildfire occurrence (Chas-Amil et al. 2015; Elia et al. 2019, 2020; Milanović et al. 2020).

Focusing on vegetation drivers, the WUI area covered by herbaceous vegetation (HERB), which includes grasslands, agricultural areas or pastures, was the most influencing driver of AB_WUI at the regional level, and played also an important role in the analyses of the separate municipal groups. This result was confirmed by comparing the map of HERB distribution (Fig. S7a) with the map of observed annual AB_WUI (ha year⁻¹) for the period 2005–2019 (Fig. 2a): for both variables, the highest values were distributed on the western side of the island, particularly in southern Sardinia. Our results evidence that the relation between the presence of herbaceous vegetation and AB_WUI is more evident for the H group. Differently, the role of vegetation on IP_WUI is more limited, exhibiting a strong influence of HERB and of the WUI area covered by shrublands (SHR) only for the H group. The above findings are consistent with previous works carried out in Sardinia and other European countries, which highlighted that herbaceous vegetation is often linked to large events and, especially when its distribution is combined with a complex topography and strong winds, is characterised by high wildfire transmission potential (Ager et al. 2014; Alcasena et al. 2017; Elia et al. 2019; Salis et al. 2021). In this regard, agro-pastoral land abandonment and the inherent increases in dead fine fuel load and continuity should be carefully considered in future years, as the probability of large wildfire occurrence able to spread towards WUI or forest areas could be exacerbated (Pais et al. 2020; Rodrigues et al. 2022; Salis et al. 2022).

At the regional level, climatic variables did not exhibit a strong influence on AB_WUI and IP_WUI. For the separated analyses, the annual average municipal maximum temperature (TMAX), the annual average municipal solar radiation (SRAD) and the total annual municipal precipitation (PREC) played a role in determining AB_WUI, more in particular for the F group. Differently, the annual average municipal wind speed (WS), TMAX, SRAD, and PREC influenced IP_WUI for the H group. According to this result, the Sardinian hilly areas, which are mostly located in the inner zones of the island, are characterised by limited levels of fuel management, particularly in forest and marginal pastures, and high dryness conditions in the summer season, are therefore more exposed to wildfire ignitions and large wildfire spread; the effect of elevation on wildfire occurrence was also observed in eastern Europe (Milanović et al. 2020). Other authors highlighted climatic factors as the most influencing variables of area burned at regional, continental, and global scales (Aldersley et al. 2011; Curt et al. 2015; Jain et al. 2020, Rodriguez-Jimenez et al. 2023). Grünig et al. (2023) found a strong correlation between the increasing aridity and the largest fires in Europe, while Elia et al. 2020 highlighted the relevant role played by climate variables on wildfire regime in insular regions of Italy.

Lastly, any of the variables of the zootechnical group showed significant effects on AB_WUI and IP_WUI.

As far as the spatial model predictions of AB_WUI and IP_WUI are concerned, the results were overall promising for both analyses (regional scale vs municipal groups), with limited differences between observed and estimated values. In general terms, the analysis carried out considering F and H municipal groups performed slightly better than the regional scale analysis. We evidenced underpredictive issues of both models in the municipalities characterised by the highest values of AB_WUI. Regarding IP_WUI, we noticed an underestimation of both models in southern Sardinia, particularly in those zones characterised by high municipal areas occupied by WUI; on the contrary, overpredictions were mainly concentrated in the most populated municipalities of central and northern Sardinia for the analysis carried out at the regional scale.

The proposed approach in this study has a number of intrinsic limits. For instance, we attributed a set of driving factors and variables for each municipal area in Sardinia, but this did not allow to take into account the spatial complexity of WUI conditions, particularly in some touristic coastal areas or main towns and suburbs. Moreover, when determining WUI, we did not take into account isolated houses or groups of houses that were not captured by the input layers used, and this could have influenced our final results. Furthermore, an obstacle for the adoption of similar ML methods to model wildfire activity is the perceived lack of interpretability of such methods, which are often considered to be black box models (Jain et al. 2020).