Protected areas, drought, and grazing regimes influence fire occurrence in a fire-prone Mediterranean region

Abstract

Background

Extreme fire seasons in the Mediterranean basin have received international attention due to the damage caused to people, livelihoods, and vulnerable ecosystems. There is a body of literature linking increasingly intense, large fires to a build-up of fuel from rural land abandonment exacerbated by climate change. However, a better understanding of the complex factors driving fires in fire-prone landscapes is needed. We use a global database based on the MODIS Fire CCI51 product, and the Greater Côa Valley, a 340,000-ha area in Portugal, as a case study, to investigate the environmental drivers of fire and potential tools for managing fires in a landscape that has undergone changing agricultural and grazing management.

Results

Between 2001 and 2020, fires burned 32% (1881.45 km²) of the study area. Scrublands proportionally burnt the most, but agricultural land and forests were also greatly impacted. The risk of large fires (> 1 km²) was highest in these land cover types under dry conditions in late summer. Areas with higher sheep densities were more likely to burn, while cattle density had no apparent relationship with fire occurrence. There was also a 15% lower probability of a fire occurring in protected areas.

Conclusion

Future climatic changes that increase drought conditions will likely elevate the risk of large fires in the Mediterranean basin, and abandoned farmland undergoing natural succession towards scrubland will be at particularly high risk. Our results indicate that livestock grazing does not provide a simple solution to reducing fire risk, but that a more holistic management approach addressing social causes and nature-based solutions could be effective in reducing fire occurrence.

Resumen

Antecedentes

Estaciones de fuegos extremas en la cuenca del Mediterráneo han recibido la atención internacional debido a los daños causados a personas, hogares, y ecosistemas vulnerables. Hay abundante literatura que liga fuegos incrementalmente intensos y extensos a la gran acumulación de combustibles vegetales debido al abandono rural, exacerbado por el cambio climático. Sin embargo, es necesario un mayor entendimiento de los complejos factores que conducen a estos incendios en paisajes proclives al fuego. Usamos una base de datos global del producto MODIS FIRE CCI51, y el gran Valle Coa, de 340 mil ha en Portugal, como un estudio de caso, para investigar los conducentes ambientales de incendios y las herramientas potenciales para manejar esos incendios en un paisaje que ha atravesado por cambios en su manejo agrícola y ganadero.

Resultados

Entre 2000 y 2020, fuegos de vegetación quemaron el 32% del área de estudios. Los arbustales fueron los que proporcionalmente más se quemaron, aunque áreas agrícolas y bosques fueron también altamente impactadas. El riesgo de grandes incendios (> 1 km) fue mayor en aquellos tipos de cobertura bajo condiciones secas en el verano tardío. Las áreas que contenían grandes densidades de ovejas fueron proclives a quemarse, mientras que la densidad de vacas aparentemente no se relacionó con la ocurrencia de fuegos. Hubo también un 15% de probabilidad más baja de ocurrencia de incendios en áreas protegidas.

Conclusiones

Los cambios en el clima a futuro, que incrementen las condiciones de sequía, probablemente eleven el riesgo de grandes incendios en la Cuenca del Mediterráneo, y las áreas agrícolas abandonadas, cuya sucesión las conduzca hacia arbustales, van a estar particularmente en alto riesgo de incendio. Nuestros resultados indican que el pastoreo por el ganado no provee de una simple solución para reducir el riesgo de incendio, pero una aproximación de manejo más holístico en el que se incluyan las causas sociales y las soluciones basadas en la naturaleza pueden ser efectivas para reducir la ocurrencia de incendios.

Background

Extreme fire seasons, characterized by large, severe fires, have been reported across the Mediterranean basin (Moreira et al. 2020). Although fires are a natural occurrence (Pausas et al. 2008), abnormally large fires can have severe negative environmental and socio-economic consequences, including loss of human life (Linley et al. 2022). Widespread abandonment of rural areas and traditional agricultural practices has been implicated as a major cause of increased fire risk globally (Viedma et al. 2015). Pastoral burning and livestock grazing are considered important mechanisms in maintaining open, mosaic landscapes and reducing the overall risk of large fires (Rouet-Leduc et al. 2021). The loss of such practices, and the absence or low density of large native herbivores, has allowed the encroachment of scrublands and forests into abandoned areas and an accumulation of combustible fuels (Fernandes et al. 2014; Pinto et al. 2020; Vilar et al. (2016)). Afforestation and reforestation efforts, promoted in the Mediterranean for timber production and restoration purposes (Pausas et al. 2004), have created homogeneous stands of flammable species such as pine (Pinus spp.) and eucalyptus (Eucalyptus spp.). Alongside land use change, increasing temperatures and droughts are providing fires with drier, more flammable fuel, and were a key driver in the destructive fires that affected Southern Europe in 2017 (Turco et al. 2019).

Outbreaks of large wildfires in recent decades have flagged the need to improve fire prevention strategies (Rouet-Leduc et al. 2021). Direct interventions such as controlled burning and fire-fighting can be costly (Hope et al. 2016; Wang et al. 2021), and fire suppression can have negative ecological effects, especially in fire-adapted ecosystems (Bond & Keeley 2005). By allowing a build-up of fuel, fire suppression can also increase fire risk in the long term (Arévalo & Naranjo-Cigala 2018). Recognizing the value of traditional land use systems, agri-environmental policies aim to reduce rural abandonment and promote livestock grazing (Jones et al. 2016). However, the current understanding of how herbivore behaviors influence fire regimes, for example by reducing fuel loads or changing vegetation structure and moisture (Johnson et al. 2018), is limited. How herbivore effects vary among species and habitats, particularly in tall shrublands and forest ecosystems, where vegetation may be less palatable, remains unclear (Rouet-Leduc et al. 2021).

An alternative use for abandoned lands is biodiversity conservation. Despite being perceived negatively, abandoned lands offer opportunities for conservation and restoration, including restoration of fire-regulating and other ecosystem services. Proponents of rewilding, defined as restoration of self-sustaining and complex ecosystems with minimal ongoing management (Pettorelli et al. 2018), argue that the reintroduction of missing large herbivores represents a cost-effective approach to maintain a mosaic landscape and restore natural fire regimes (Rewilding Europe 2022a). European wild herbivores, in many areas, have been replaced by livestock. Broader rewilding concepts acknowledge the potential importance of domesticated livestock, but it is not clear to what extent they can replicate the spectrum of effects of wild species on fire regimes.

Portugal has recorded the highest fire density in Southern Europe and the largest percentage of burned area in the last few decades (Moreira et al. 2023). The 2017 fire season drew considerable international attention due to record-breaking wildfires, which burned 539,920 ha of forest, scrubland, and agricultural land, and caused 117 deaths (San-Miguel-Ayanz et al. 2020). Previous studies have reported an increase in the frequency of large fires in Portugal and identified several factors driving Portugal’s fires, including climate, weather, land cover, and human activities (Moreira et al. 2011; Viedma et al. 2015; San-Miguel-Ayanz et al. 2019). These studies focus on the link between increased fire risk and rural abandonment. Calls have been made for local-scale studies to fill in the current gap in knowledge regarding the potential for abandoned lands to provide balanced benefits for biodiversity conservation, ecosystem services, and people (Daskalova & Kamp 2023). The Greater Côa Valley, in northeastern Portugal, is an arid, open-forested landscape that has experienced high levels of rural abandonment. There are large-scale efforts to restore and rewild the area, including through the implementation of (semi-)natural grazing practices, intended to improve ecosystem functioning and resilience (Rewilding Europe 2020).

We use the Greater Côa Valley as a case study to explore the influence of management and environmental factors on fires, with the aim of informing management of the land for biodiversity and people. We use the global FRY database (Laurent et al. 2018), based on the Moderate Resolution Imaging Spectroradiometer (MODIS) Fire Disturbance Climate Change Initiative burned area product (FireCCI51) version 5.1, and cloud-based processing, to compile a 20-year dataset of fires and environmental and human characteristics of the Greater Côa Valley. We present a comprehensive study of the contemporary fire regime. Our research had three main objectives: to explore (i) the susceptibility of land cover types to fires; (ii) environmental and human drivers of fire occurrence; and (iii) the relationship between fire occurrence and management, including protected areas (PAs) and livestock type and density.

Methods

We used the methodology adopted by Kirkland et al. (2023), published at https://github.com/btomairekirkland/landscape-fire-analysis, to compile, process, and analyze fire data and data on potential drivers. Full details are provided below.

Study area

Situated in northeastern Portugal bordering Spain, the Greater Côa Valley is a 340,000-ha region that connects the Malcata mountain range in the south with the Douro Valley in the north. Rock engravings in the UNESCO Côa Valley Archaeological Park show that large herbivores such as wild horses Equus ferus, aurochs (Bos primigenius, the wild ancestors of domestic cattle), Iberian ibex (Capra pyrenaica), and deer were historically present in the landscape. Over the years, a mosaic landscape of forests, rocky scrubland, and scattered fields has been maintained by traditional farming practices such as sheep grazing and burning to renew or create new pastures by clearing vegetation and fertilizing the land. However, with the loss of traditional grazing systems due to one of the highest land abandonment rates in Europe, scrublands are expanding (Rewilding Europe 2022a). Soils are being eroded and seedbanks depleted by large fires, impeding the development of mature woodland in many areas (Rewilding Europe 2022a). Alongside rural depopulation, and encouraged by European agri-environment subsidies, in some areas traditional sheep husbandry is being replaced by extensive cattle farming (Costa Freitas et al. 2020; Faria and Morales 2020). Unlike with sheep, modern cattle grazing is not associated with the use of fire, but instead with fertilization with manure, rotational grazing, and seeding as ways to maintain and improve grasslands. Strict regulations and prohibitions have decoupled today’s pastural burning practices from traditional ones (de Oliveria et al. 2023).

Over 100,000 ha of land have been set aside for conservation in Natura 2000 Special Protection Areas and Sites of Community Importance, and 34,000 ha have been formally protected through Portugal’s network of natural parks and nature reserves (Fig. 1a). Four sites covering more than 1000 ha are under rewilding-based management by Rewilding Portugal, working closely with Rewilding Europe (Fig. 1a). Their main aim is to restore the landscape to one comprised of Mediterranean woodlands and open areas, using grazers and other keystone species and “engineers” of landscape heterogeneity. Sorraia horses and Tauros (the closest descendants of the auroch) have already been released, with further introductions planned for the future (Rewilding Europe 2023). The intention of such introductions is to reduce shrub biomass, height, and cover and facilitate the expansion of native forest through seed dispersal and creating open spaces. The region is now home to a growing population of wild herbivores and other keystone species (Rewilding Portugal 2021). Additional rewilding measures to mitigate fire risk include restoring riparian and wetland ecosystems.

Fig. 1

Map of the Greater Côa Valley, showing a cumulative area burned by fires, indicated by a black (single burning event) to yellow gradient (eight fire events), between 2001 and 2020, and b details of the study area. Polygons outlining fires larger than ~ 0.24 km², based on the FireCCI51 burned area product, are shown. Voxels of fire and non-fire observations used in the analysis were sampled from the area outlined by the red dashed line. Blue and green regions are Special Protection Areas (SPZ) and Sites of Community Importance (SCI) of the Natura 2000 network. The border between Spain and Portugal is shown by the thick dashed line going from north to south. The boundary of the Greater Côa Valley (GCV) is shown by the solid gray line. Portugal’s National Network of Protected Areas (RNAP) is indicated by the diagonal stripes. The four rewilding sites under Rewilding Portugal management are indicated by green crosses. Locations of roads and cities were obtained from www.openstreetmap.org. Vector data for the Douro River (in the North) were obtained from https://www.naturalearthdata.com/

Fire data

We used the FRY dataset, a global spatial database of fire patches, defined as “groups of adjacent burned pixels with temporal coincidence”, reconstructed from the 250 m pixel-based MODIS FireCCI51 burned area product (Laurent et al. 2018, p. 2). These patches are generated using a “flood-fill” algorithm, parameterized by a cut-off value corresponding to the maximum time difference between the burn dates of neighboring pixels belonging to the same fire patch. We present analyses using a temporal cut-off value of 6 days, resulting in more, but smaller fire patches than larger cut-off values. We selected fire patches that burned within the Greater Côa Valley for the period spanning 2001–2020, inclusive (Fig. 1b). “Functional traits”, including the coordinates of the ignition point and fire patch size, are provided (Laurent et al. 2018).

We follow the methodology of Kirkland et al. (2023) in which fire patches are converted to a 250 × 250 m grid for each day of the study period, between 24/01/2001 and 22/09/2020. This generated a set of daily voxels representing fire and non-fire observations from which to explore the probability of a fire having occurred during the study period. We assigned voxels with a dichotomous response variable, where 1 indicates that a fire occurred and 0 indicates no fire. We sampled non-fire (n = 54,657, < 1%) observations at a ~ 1:2 ratio of fire to non-fire observations, capturing the spatial heterogeneity of the region and producing a dataset of a computationally feasible size. We sampled voxels from a subset of parishes that cover areas of high and low fire activity and varying livestock densities (Fig. 1a). The Malcata Site of Community Importance in the south represents a site of low fire activity, while the Vale do Côa Special Protection Area in the north represents high fire activity. The area surrounding Malcata was also included as an unprotected site experiencing high fire activity. We ensured an even distribution of voxel centroids across land cover types for non-fire observations. We excluded non-fire observations if a fire occurred in the same grid cell within a year prior to the voxel date, due to fuel limitations after a burn inhibiting reburn. We excluded fires smaller than 1 km², which are likely to include intentional fires used for sheep grazing, to focus on larger, and likely more severe that contribute most to the total burned area. This does not remove any large wildfires that started out as intentional fires.

Input variables

We gathered data on land cover type; Normalized Difference Vegetation Index (NDVI); temperature; rainfall; soil moisture content; elevation; travel speed; human population density; distance to nearest road; livestock density and type; and four fire danger metrics of the European Forest Fire Information System, specifically the Drought Code (DC); Keetch-Byram Drought Index (KBDI); Fire Weather Index (FWI); and Fine Fuel Moisture Code (FFMC) (see Supplementary Table S1). These (or comparable) variables have proved to be important in explaining fires throughout the Mediterranean and elsewhere (see Supplementary Table S1). We extracted data from remote-sensing and other ready-to-use publicly available data. We used “rgee” 1.1.3, a binding package for calling Google Earth Engine (Gorelick et al. 2017; Tamiminia et al. 2020), “raster” 3.5–15 (Hijmans et al. 2020), and “sf’ 1.0–7 (Pebesma et al. 2020) packages in R version 4.0.5 (R Development Core Team 2018) to extract data.

We obtained land cover information for 2007, 2010, 2015, and 2018 from the Portuguese Carta de Uso e Ocupação do Solo dataset. We used the level 1 classification (Supplementary Table S2). We extracted the proportion of each land cover type encompassed within voxels. We did the same for fire ignition points, selecting 39% (n = 13,994) of ignitions detailed in the FRY dataset that intersected fire patches and converting ignitions to a 250-m resolution grid.

We extracted NDVI, an indicator, from − 1 to 1, of vegetation vigor or “greenness,” to use as a proxy for live fuel availability. We combined the MODIS 250 m NDVI datasets MOD13Q1 v061 and MYD13Q1 v061 (Didan 2021a; c), available at 8-day intervals when combined. We extracted NDVI from the data product closest in time that pre-dated (i.e., were chronologically before) the voxel, thereby ensuring pre-fire NDVI values were obtained, since fires modify vegetation dynamics and structure (Pompa-García et al. 2022). NDVI data were, on average, produced 13 days before fires started, and 95% was produced within 17 days of a fire. Five fires (1% of large fires) did not have higher resolution NDVI data, so the 500-m resolution MOD13A1 and MYD13A1 v061 products were used to extract NDVI (Didan 2021b; d).

We collated maximum daily temperature and accumulated daily rainfall from the European Centre for Medium-Range Weather Forecasts Reanalysis v5 (ERA5) (Hersbach et al. 2020). Weather data were unavailable for one fire that occurred in September 2020. This fire was omitted from analyses. We extracted the mean monthly soil moisture content from the TerraClimate dataset (Abatzoglou et al. 2018). We used data from the previous month for fires dated to the beginning of the month (before the 15th), otherwise data from the same month were used. We extracted mean elevation from the Copernicus Digital Surface Model (European Space Agency 2022).

We obtained daily values of meteorological fire danger metrics, specifically the FFMC, DC, and FWI from the Canadian FWI System (van Wagner 1987) and the KBDI (Keetch & Byram 1968), from the Copernicus Climate Data Store (CMES 2019). DC and KBDI provide good measures of prolonged drying conditions and are indicators of the moisture content of deep compact organic layers and large dead woody fuels. FFMC is an indicator of the moisture content of dead fine fuels in the litter layer (e.g., needles, leaves, and small twigs) (de Groot et al. 2005; de Jong et al. 2016). FWI indicates potential fire intensity if a fire were to occur (van Wagner 1987).

We extracted mean travel speeds from the global friction surface of Weiss et al. (2018, 2020), which represents the rate at which humans can move between land-based pixels using the optimal path. We extracted population densities (people/km²) from the (CIESIN, 2017). Vector datasets from the Global Roads Inventory Project were used to calculate the shortest straight-line distance from voxels to the nearest road (Meijer et al. 2018).

We intersected voxels and land cover shapefiles with the boundaries of national and private PAs and Natura 2000 sites. The Instituto Nacional de Estadistica (2023) provides data on livestock numbers at 10-year intervals. Manual collection of livestock data for individual parishes was time-consuming. We therefore focussed on a subset of the Greater Côa Valley, as mentioned, covering areas of high and low fire activity, encompassing varying livestock densities (Supplementary Fig. S1a and S1b). We extracted livestock densities within voxels using livestock data from the year closest in time to the voxel date. This resulted in three data points for each parish (1999, 2009, and 2019). We focused on sheep and cattle to explore the implications of replacing traditional sheep husbandry with cattle farming.

Statistical analyses

We explored the relationship between explanatory variables and fire occurrence, through binomial generalized additive models (GAMs), allowing non-linear relationships between the response and predictor variables. GAMs are increasingly used to model spatio-temporal fire processes (McWethy et al. 2018; Pinto et al. 2020). We increased the “penalty” per degree of freedom to reduce overfitting, by setting gamma to 1.4 in the model formulas (Wood 2017). We selected among covariates using the double penalty approach (Marra and Wood 2011), which has the effect of shrinking effects of non-influential variables to zero. We focussed on the main fire season from June to November (n = 215, 96.0% of large fires). We included year as a random effect to explain any variance in annual fire probability not associated with the fixed effects.

Pearson’s correlation revealed a latitudinal pattern in elevation (r =  − 0.68). We omitted elevation from the model in favor of a spatial term (a function of latitude and longitude), capturing variation in elevation and other spatially explicit drivers of fire size not included in the model. There were correlations between temperature, soil moisture, rainfall, and the fire danger indices (r > 0.60). We ran competing models omitting collinear variables, and selected the top-performing model based on the Akaike information criterion (Supplementary Table S3), whereby the model with the lowest AIC has the “best” relative fit, given the number of parameters (Akaike 1998). We tested for an interaction between NDVI and land cover type. Inclusion of this interaction reduced model fit (∆AIC = 6695) and visual inspection of plots showed no difference in the relationship with fire probability within land cover types. The final model of the logit of the probability of a fire p_i occurring in voxel i (i.e., the probability of a fire p_xyt occurring in a grid cell located at xy on day t) is detailed in Eq. 1. β₀ is the model intercept; g₁() to g₃() are the proportion of each of the land cover types within voxel i; Population_i, Bovine_i, and Sheep_i are log-transformed densities; (x_i,y_i) are the coordinates of the voxel; u is the random effect of year; and β₁ describes the likelihood of a fire occurring in a PA or not. The functions g_i() are the non-parametric smooth terms describing non-linear effects.

$$\text{logit}(\text{p}_{\text{i}}) = \beta_{0}+{\text{g}}_1({\text{Forest}}_{\text{i}})+{\text{g}}_{2}({\text{Scrubland}}_{\text{i}})+{\text{g}}_3({\text{Agriculture}}_{\text{i}})+{\text{g}}_4({\text{NDVI}}_{\text{i}})+{\text{g}}_5({\text{Travel}}_{\text{i}})+ \,{\text{g}}_6({\text{Population}}_{\text{i}})+{\text{g}}_7({\text{Roads}}_{\text{i}}){+}{\text{g}}_8 ({\text{Bovine}}_{\text{i}}) + {\text{g}}_{10}({\text{Sheep}}_{\text{i}})+ {\text{g}}_{11} ({\text{KBDI}}_{\text{i}} )+ {\text{g}}_{12} ({\text{FFMC}}_{\text{i}})+ {\text{g}}_{13} ({\text{Year}}_{\text{i}}) +{\text{g}}_{14} ({\text{x}}_{\text{i}}{,}{\text{y}}_{\text{i}}) +\beta_11\text{(Protected) +u}$$

(1)

To account for the disproportional sampling of non-fire observations across land cover types, data points were weighted according to Eq. 2. W_l refers to the weight assigned to all voxels whose centroids overlay land cover type l, S_l is the total number of voxels associated with land cover type l included in the model, and T_l is the total number of voxels associated with land cover type l across the study region, so that undersampled land cover types are represented proportionally in the model. GAMs were fitted using the R package “MGCV” 1.8–34 (Wood 2019) in R 4.0.5.

$${\text{W}}_{\text{l}}\text{=}\frac{{\text{T}}_{\text{l}}}{{\textrm{S}}_{\textrm{l}}}$$

(2)

We performed model validation by comparing observed fire occurrences with the predicted number of fire occurrences given the raw data, using the area under the receiver operator curve to identify the threshold that maximized sensitivity plus specificity and convert probabilities into binary values (Robin et al. 2011). This was done using the “pROC” 1.18.5 package (Robin et al. 2023).

Results

Fires across space and time

There were 680 fires that occurred in the Greater Côa Valley between 2001 and 2020, resulting in a cumulative burned area of 1875 km² (Fig. 1). Fires averaged 2.76 km² in size, with 226 reaching > 1 km². These “large” fires accounted for 98.8% (1859 km²) of the cumulative burned area, burning around a third (1031 km²) of the Greater Côa Valley, and also spreading out to burn adjacent areas. Roughly a third (30.6%) of burned grid cells burned more than once, most (77.8%) burning twice during the study period. Those that burned more than twice experienced large fires, on average, every 6 years. The maximum number of times a grid cell burned was eight. The largest numbers of fires occurred during the summer, from June to August, half of all fires (49.1%, n = 111) occurring in August (Supplementary Fig. S2). The greatest number of fires was recorded in the summer of 2003, followed by 2017 (Supplementary Figs. S2 and S3). Six large fires occurred between 2018 and 2020.

Potential drivers of fire occurrence

The majority of large fires started in and burned scrubland, where 71.1% (n = 464) of ignitions occurred and large fires burned 726 km² of land, equating to 46.9% of the cumulative burned area within the Greater Côa Valley (Table 1). This is one of the dominant land cover types, covering 27.4% of the Greater Côa Valley, after agriculture (29.9%) and forest (29.6%). The high propensity of scrubland to burn (Supplementary Fig. S4) was reflected in a positive linear relationship between the area of scrubland in a voxel and the probability of a fire occurring (Fig. 2a). Substantial areas of forests and agricultural areas also burned, accounting for 28.4% and 14.8% of cumulative burned area, respectively (Supplementary Fig. S4). This was reflected in a positive relationship between fire occurrence and the proportion of forest and agriculture in a voxel, though the strength of the relationships was weaker compared with the effect of scrubland (Fig. 2b and c). Heterogeneous agricultural land was the most heavily burned type of farmland (29.7%), followed by permanent and temporary crops (16.6%). “Artificial surfaces” (“territórios artificializados”), such as urban areas, and agroforestry areas burned in very small proportions (< 10%, see Table 1, Supplementary Fig. S4). Predicting the probability of a fire in grid cells comprised of each of these three land cover types at the upper quartile values of KBDI and FFMC indicated a c. 30% chance of a fire having occurred during the main fire season, over the course of the 20-year study period.

Table 1 Summary of large fires by land cover type in the Greater Côa Valley (GCV) between 2001 and 2020. Protected areas (PAs) include Natura 2000 sites and nationally designated PAs. Cumulative burned area includes repeated burning of the same area. All areas are shown in km²

Fig. 2

Estimated effects (solid line) of covariates on the average probability of a fire occurring in a 250 m grid cell in the Greater Côa Valley, during the main fire season (June-November) over the 20-year study period, showing the effect of: the proportion of a scrubland, b forest, and c agriculture a cell; d NDVI; e population density; f distance to road; h cattle and i sheep density; j the Keetch-Byram Drought Index (KBDI); and k the Fine Fuel Moisture Code (FFMC). The dotted lines represent 95% confidence bands. Vertical lines are the average values for the three major land cover types (measured from grid cells with proportional coverage > 0.8 of each land cover type). The gray shaded area is the relative density of values within sampled voxels, scaled to fit within the plot area. Summary statistics for the GAM are shown in Table S3

NDVI exhibited a unimodal relationship with fire occurrence, peaking at moderate values (~ 40–50%, Fig. 2d). There was a negative sigmoidal relationship between human population density and fire occurrence (Fig. 2e). Visual inspection of plots indicated no effect of distance to road or travel time (Fig. 2f and g). There was an initial positive relationship between fire occurrence and sheep densities, which plateaued at higher densities (Fig. 2i), but no effect of cattle density (Fig. 2h). KBDI showed a positive relationship with fire occurrence (between KBDI values of 25 and 50), though the probability of fire began to decrease at the high end of the scale (Fig. 2j). The sigmoidal relationship between FFMC and fire occurrence (Fig. 2k) indicated a low probability of fires at low FFMC values, but an exponential increase in probability as FFMC increased. Consistent with the fire danger indices documentation, fires were most likely to occur at FFMC values above ~ 70, reflecting lower moisture in the fine fuel layer (CMES 2019). There was spatial and temporal variation in fire occurrence (Supplementary Table S4). The spatial effect indicated a lower probability of a fire in the Malcata Site of Community Importance (Supplementary Fig. S5), particularly where drought indices (KBDI and FFMC) were lower (Supplementary Fig. S1). Over three quarters of the deviance in fire occurrence was explained by the model (Supplementary Table S3).

The predicted number of burnt voxels, based on a threshold of 0.51 to convert probabilities into a binary outcome, was 16,988, compared with the observed 15,941 (i.e., 45.6% and 42.7% of sampled voxels, respectively). Kendall’s correlation coefficient also revealed a strong relationship between observed data and predicted probabilities (τb = 0.704, P < 0.001). This indicated a good model fit.

Fires and protected areas (PAs)

Five Natura 2000 sites cover 36.3% of the Greater Côa Valley (1157 km²), and three national PAs cover 10.8% (344 km²). The dominant land covers in PAs were forests (33.8%), scrubland (30.3%), and agriculture (23.0%). One quarter (292 km²) of protected land within the Greater Côa Valley burned; 25.2% (292 km²) of land within the Natura 2000 sites (Table 1), mostly in the northern Vale do Côa Special Protection Area, and 10.2% (35 km²) of state-owned and private PAs (Fig. 1). This equated to a cumulative burned area of 394 km². Double the amount of unprotected land burned (739 km²), which equated to 36.4% of unprotected land. Taking into account land cover type, fires were, on average, 14.6% less likely to occur inside a PA than outside (Fig. 3, Supplementary Table S4). The most fire-prone land cover type in PAs was scrubland, which accounted for 56.3% of cumulative burned protected land and 43.0% of scrubland within PAs (Table 1).

Fig. 3

Probability of a fire occurring over the 20-year study period in unprotected vs. protected land during the main fire season of the Greater Côa Valley, at different levels of scrubland and forest coverage. Proportion refers to the proportion of each land cover type in a 250-m grid cell. Lines are colored by land cover type. The solid line represents unprotected land and the dashed line represents protected land. The shaded areas represent 95% confidence intervals

Discussion

Our findings indicate the need for effective fire management to account for land use and climatic changes, which together are increasing risks of fire in the Mediterranean. Fires cumulatively burned more than 1800 km² of the Greater Côa Valley from 2001 to 2020, 32% of the total area and a grid cell containing any of the three dominant land cover types in the region—agriculture, forest, and scrubland—had a c. 30% chance of burning during the main fire season. Our work adds to a growing body of evidence of Mediterranean scrubland as a fire-prone land cover type (Casals et al. 2023; Damianidis et al. 2021), but also highlights the risk of fire in agricultural and forested areas. Complete fire suppression is not desirable from an ecological perspective and can lead to long-term accumulation of fuel and, consequently, more frequent, intense fires (Moreira et al. 2020). This is especially the case in Mediterranean ecosystems, where fires are an integral natural process that supports biodiversity, regulates nutrient flows, and maintains habitats (Bond & Keeley 2005). However, as large fires have increased in frequency, so too have the costs to people and the environment (Johnston et al. 2021; Wang et al. 2021). We recorded an increase in the number of large fires in the region in 2017, a year in which large wildfires, attributed to drought and high temperatures, were recorded throughout Portugal (Turco et al. 2017). We found that livestock density was not associated with reduced fire occurrence and that PAs were at lower risk of burning than unprotected land, which has direct implications for fire prevention management.

The majority of ignitions (c. 97%, Meira Castro et al. 2020) in Portugal are human-caused. However, in the Greater Côa Valley, large fires mostly occurred outside of urban areas, with fire occurrence decreasing as population density increased beyond c. 100 people/km², around the size of the smallest town, becoming extremely unlikely in cities. This can be attributed to the high risk of rural fires due to greater abundance and continuity of fuel, as well as frequent use by local people for grazing livestock. Pastoral fires are the source of one in ten fires in Portugal (ICNF 2018). Of the 89% of fires in the Greater Côa Valley region that are caused by humans, 65% are caused by negligence, which include pasture fires, as well as burning crop residues without authorization and working with machinery on hot days. Only 2% of fires in the region have natural ignition sources (Meira Castro et al. 2020; ICNF 2023). Outbreaks of wildfires from intentional pasture fires have been linked to a loss of traditional knowledge and the risk of criminal charges following legal restrictions placed on the practice in Portugal (de Oliveria et al. 2023). Against this backdrop, though pasture burning could be contributing to land management goals at the local scale (Torres-Manso et al. 2014), the analysis presented here suggests that the once traditional practice may be increasing the risk of large fires.

Findings from other European landscapes suggest that PAs are particularly at-risk rural areas due to a lack of fuel management, as well as insufficient fire detection and fire-fighting (Arellano-del-Verbo et al. 2023; Kirkland et al. 2023; Rodrigues et al. 2023). Yet, in the Greater Côa Valley, a much larger proportion of unprotected forest burned (244.50 km², 43.4%), compared with protected forests (72.77 km², 18.4%). PAs may be at lower risk of fires due to reduced public access and restrictions on land use, particularly within nature reserves, as well as surveillance by Portugal’s park rangers (Rewilding Europe 2022b). Education, increasing public awareness, and restricting access to hazardous areas, during the fire season, could therefore be important tools for reducing the risk of fires in rural areas.

The lower risk to PAs could also be attributed to the protection and management of native forests, which include fire-resistant species such as cork and holm oak (see Fig. 3). The degradation of such forests is of major conservation concern (Monteiro-Henriques and Fernandes 2018). Concurrent with other studies from the Mediterranean region (Turco et al. 2017; Urbieta et al. 2015), dry conditions, indicated by KBDI and FFMC, were directly linked to fire risk in the Greater Côa Valley. The decline in fire risk at high KBDI values likely reflects lower fire incidences towards the end of the fire season, as expected following prolonged droughts limiting fuel availability (Andrade & Bugalho 2023). KBDI values within the Malcata Site of Community Importance (see Supplementary Fig. S1) suggest that management of PAs can help to reduce fire risk through maintaining hydrological processes that protect against droughts. Insights from the Greater Côa Valley could therefore elucidate how to reconcile fire management with the conservation of threatened species and habitats and ecosystem services in the Mediterranean.

Integrated within the literature linking fire risk and land abandonment is the argument for livestock grazing to control the accumulation of fuel (Rouet-Leduc et al. 2021). The perceived value of livestock is evident in current agri-environment subsidies: the Portuguese Fundo Florestal Permanente, for example, is subsidizing sheep grazing in fire-sensitive areas (ICNF 2018), whereas the European Common Agricultural Policy payments are more closely coupled with cattle rearing (Faria & Morales 2020; Ribeiro et al. 2014). Any effect of livestock may have been attenuated by the coarse spatial and temporal resolution of the livestock data used in this study, hindering a more nuanced understanding of the interactions between livestock grazing, land cover, and their impact on NDVI, a proxy for live fuel availability. However, the current management of grazing systems in the Greater Côa Valley, e.g., single-herbivore systems grazing, small parcels, and supplementary feeding, does not appear to be significantly reducing fire risk. The probability of burning was highest at moderate values of NDVI, within the range of mean NDVI values for agriculture, forest, and scrubland, whereas higher NDVI values, likely corresponding to dense, green vegetation and possibly indicating the peak growth stages of vegetation, was associated with a lower probability of burning. The NDVI and livestock results could be further evidence that, where herbivores are grazed extensively and feed on green, herbaceous vegetation, grazing, in isolation, may not be an effective technique to reduce the more flammable woody or fine fuels (Calleja et al. 2019; Bashan and Bar-Massada 2017). This may be particularly true under increasingly dry conditions.

Conclusion

The risk of large fires in the Mediterranean region poses a significant threat, particularly in the face of climate change and rural depopulation. Without implementing effective practices to reduce ignitions, this risk is only set to increase, as evidenced by Portugal’s worst wildfire season to date, in 2017. Our findings indicate that alone, increasing grazing pressure is unlikely to provide a simple solution, as is sometimes promoted throughout Europe. Only through integrated management strategies that consider various environmental, climatic, and social factors can we mitigate the growing threat of large fires in the Mediterranean while allowing abandoned lands to naturally regenerate and safeguard both the environment and communities at risk.