Intensified burn severity in California's northern coastal mountains by drier climatic condition

The study focused on the coastal mountain region in northern California, mostly foothills and low mountains up to 2265 m, surrounded by low and flat Central Valley ecoregion to the east and Klamath Mountain ecoregion to the north (figure 1). The Berryessa Snow Mountain National Monument is located in the southeast of this region. Herbaceous plants (25%), shrubs (19%), and hardwood woodlands (15%) dominated the lowland of the landscape, while smaller areas of hardwood forests (16%), and conifer forests (6%) occur at higher elevations in the study area (according to FRAP 2015 vegetation type dataset, http://frap.fire.ca.gov/projects/frap_veg/classification).

Zoom In Zoom Out Reset image size

Located within the North America Mediterranean climate zone, the region experiences hot and dry summers and cool and wet winters, with a mean annual precipitation of 851 mm yr⁻¹. Precipitation varies significantly from year to year, ranging from 472 mm yr⁻¹ in 2014 to 1483 mm yr⁻¹ in 1998 (figure S1 (available online at stacks.iop.org/ERL/15/104033/mmedia)). Two extended droughts occurred since 1984; the 1987–1992 drought with a mean annual precipitation (MAP) of 618 mm and the 2012–2016 drought with a MAP of 651 mm (figure 2(a), figure S1). There were also a few years with unusually heavy precipitation such as 1998 and 1995. The long-term mean temperature during the fire season (May–November according to CALFIRE fire season summary) is 22.5 °C during 1980–2017 (figure 2(b), figure S1).

Zoom In Zoom Out Reset image size

We used fire perimeter geographic information system data compiled by the California Fire Resource and Assessment Program (FRAP) (http://frap.fire.ca.gov/), which documented wildfire perimeter records back to 1950. All fires larger than 40.5 ha (100 acres) since 1984 were selected. Smaller fires were omitted because they account for less than one percent of the total areas burned in the region and the records for them are less consistent and accurate. Burn severity was derived from 30 m multispectral data from Landsat satellite with a 16 day repeat cycle. We preprocessed the whole time series of the surface reflectance products of Landsat 5, 7 and 8 on Google Earth Engine. The corresponding quality assurance layers were used to filter out cloud contaminated pixels. Due to the high percentage of herbaceous plants and shrubs in the study area, we derived burn severity map based on initial assessment, which investigate short-term loss of vegetation (De Santis and Chuvieco 2009, Miller and Quayle 2015), because of the rapid recovery of local plant species (Lippitt et al 2013). Specifically, clear sky images within one month before and after fire were used to quantify pre- and post-fire normalized burn ratio (NBR), and then relative differenced NBR (RdNBR) was derived as (NBR_pre-NBR_post)/sqrt(abs(NBR_pre)) (Miller et al 2009). In occasions when the pre-fire NBR was equal to zero, it was replaced with 0.001, to avoid the infinity of RdNBR (Parks et al 2014a).

We further stratified areas burned into four burn severity categories (high, moderate, low, and unchanged) based on the RdNBR thresholds, as defined by Miller and Quayle classification (2015) for initial assessment (table S1). As outlaiers might exist in the RdNBR distributions due to those very small pre-fire NBR values, which cannot be interpreted as the fire-induced changes, we applied two unsupervised anomaly detection algorithms to remove these extreme RdNBR values (Park et al 2014). For each fuel type and each year, the RdNBR distribution was first truncated with anomaly scores calculated by the isolation forest algorithm and then validated by one-class support vector machine to remove those ecologically meaningless observations (Li et al 2003, Ding and Fei 2013).

Two independent variables for fire history were also explored. We created a raster layer at 30 m resolution to quantify the number of years since last fire (YSLF), for each individual year, based on the FRAP fire perimeters. Fire return interval (FRI) was from the California FRI Departure database (Safford and Van de Water 2014; www.fs.usda.gov/detail/r5/landmanagement/gis/).

We calculated time series of two vegetation indices, Normalized Difference Vegetation Index (NDVI) and Enhanced Vegetation Index (EVI), from the Landsat surface reflectance as described above (Reed et al 1994, Huete et al 1997). For each fire year, annual maximum NDVI, maximum EVI, and average NDVI a month before each fire were calculated for each pixel within the fire perimeter (table 1). We also calculated Normalized Difference Moisture Index (NDMI) a month before each fire as an indicator of fuel moisture (Gao 1996). For land cover and fuel type, we used the FRAP vegetation type data, based on the classification scheme of California Wildlife Habitat Relationship (CWHR) system (http://frap.fire.ca.gov/projects/frap_veg/classification), as the vegetation base layer.

We used monthly weather and climate data at 270 m spatial resolution from the Basin Characterization Model (BCM) during 1951–2017 (Flint et al 2013, Thorne et al 2015). For each fire year, short-term weather information was derived for the month of a wildfire and was also averaged over the fire year of that fire, for the following variables: average daily maximum, minimum, and mean temperature, and climatic water deficit (CWD) (table 1). CWD was calculated as the difference between potential evapotranspiration and the actual evapotranspiration (Stephenson 1998). We also estimated the short-term cumulative precipitation during 1 to 3 years before the fire. To represent the spatial variation in climate, long-term mean annual values were estimated for temperature (min, mean, and max), precipitation, and CWD from 1951–1980 and 1981–2010. Detrended standard deviations of temperature and precipitation time series were also estimated by first subtracting the trend line from the original time series and then calculating the standard deviation from 1951–1980 and 1981–2010 (Flint et al 2013). Detrended standard deviation were derived to represent long-term climatic variation. We also included mean wind speed and direction during the fire month from the GRIDMET dataset at 4 km (Abatzoglou 2013).

We used the NextMap Digital Terrain Model (DTM) at 5 m resolution (http://www.intermap.com) for elevation, and further derived other topographic variables, including aspect, slope, and hillshade. We also calculated vegetation height by subtracting DTM from the Digital Surface Model (DSM). To characterize the potential human impacts on burn severity, we obtained the wildland-urban interface (WUI) and residential density layers from WUI Fire Threat datasets (www.fire.ca.gov/). All environmental variables were then resampled to 30 m consistent with Landsat remote sensing products using a bilinear interpolation.

We investigated the effects of a suite of environmental factors on burn severity (table 1). Univariate analysis showed only moderate to low correlations between RdNBR and environmental variables (table 1), possibly due to the complex interactions among weather, fuels, and fire behaviors. Therefore, we chose a machine learning approach, random forest (RF), due to its accuracy and performance on non-linear predictions (Lydersen et al 2017). RF generates multiple decision trees based on randomly permuted subset variables in the out-of-bag (OOB) training samples, which can avoid the problem of multicollinearity; it can be used for both classification and regression with high accuracy and efficiency (Breiman 2001, Cutler et al 2007).

We built five sets of RF models to examine what controlled burn severity during the whole time period, and during each of the four climatic conditions: dry years, wet years, warm drought, and cool drought, respectively. Dry and wet periods were selected if the precipitation was one standard deviation away from the 60-year mean and were further validated according to the California's Fourth Climate Change Assessment (Bedsworth et al 2018; figures 2(a), (b)). Dry periods since 1984 include 1985, 1987–1992, 1994, 2001, 2007–2009, and 2012–2016. Two extended drought periods were further identified, as mentioned in section 2.1. The latest 2012–2016 drought also coincided with the historical warm records, i.e. 0.63 °C higher than the 60 year average (figure 2(b)), and is often referred as 'warm drought'. The earlier 1987–1992 drought, was only 0.28 °C higher than the long-term mean, and is therefore referred as 'cool drought' here (figures 2(a), (b)). Unusual wet years included 1986, 1993, 1995–1996, 1998, 2006, 2017.

Classification models were developed using four categories of burn severity (Table S1) as a dependent variable, i.e. high-, moderate-, and low-severity, and unchanged (Miller et al 2009). We also built the regression RF models to simulate the continuous values of RdNBR. For each climate scenario, models were trained using 70% of the randomly selected samples and tested with the remaining 30% of data. The analysis was conducted through a 5-fold cross validation. We evaluated RF classification models based on confusion matrix and quantified overall accuracy with Cohen's Kappa, F1 score, and balanced accuracy, as shown in Table S2 (Berry and Mielke 1988). For the regression model, the predicted RdNBR was evaluated with the mean absolute error, root mean square error (RMSE), and pseudo R-squared (Table S2).

We further examined the relative variable importance based on each RF classification model, using the permutation importance metric (Altmann et al 2010). For each variable, when excluded from the full model, mean decrease in prediction accuracy and percentage of increase in MSE within out-of-bag samples represents its relative importance in predicting the burn severity variation. This metric captured not only the individual impact of each predictor variable but also multivariate interactions with other predictors (Strobl et al 2007). To further clarify the decision rules for what combinations of environmental conditions were associated with each burn severity class, we built a simple representative decision tree (global surrogate for the ensemble RF trees) that had a similar structure (i.e. the number of splits and nodes) to the final random forest model (Rokach and Maimon 2005, Lakkaraju et al 2017).

For RF regression models, we further investigated the relationship between each important factor and burn severity based on the partial dependence plot (PDP). PDP is a graphical representation of the functional association between the continuous dependent variable and each independent variable, when marginalizing the derived RF regression model over mean values of all other variables (Hastie et al 2009). Specifically, the X-axis represents the range of a particular target environmental driver for burn severity, and the Y-axis shows RdNBR predictions from the model at a given value of the target variable, with all other predictors kept at their mean values. A PDP thus depicits how RdNBR partially varies with the important variables. For the PDP of each variable, we removed data points beyond the upper and lower whiskers to avoid outliers.

Finally, we tested the significance of differences in the distribution of important environmental factors under different climatic conditions using the Kruskal Wallis test, a nonparametric equivalent to the one-way analysis of variance (Kruskal and Wallis 1952).

Wildfires have been increasing since 1984 in the study area, with the annual area burned increased by 793.4 ha per year according to a least-square linear regression (figure 2(c)). A total of 206 kha was burned by 104 fires during a total of 17 dry years (12 kha per year). During the most recent extended warm drought, much more areas were burned, 22 kha by 7.2 fires each year on average, compared to 11 kha burned by 6 fires each year during the extended cooler drought (figures 3(a), (b)). This was mostly caused by the larger fire size, 3.2 (±8) kha per fire during the warm drought vs. 1.9 (±4.9) kha during the cool drought. During the 7 wet years, 59% of the areas burned occurred in 2017 (after the extended drought), which may cause the relatively smaller difference of mean annual burned areas between wet and dry years. The mean burned area per year during wet years excluding 2017 is 9.5 kha (18 kha including 2017).

Zoom In Zoom Out Reset image size

Around 36% of the total mapped burned areas of all fires during 1984–2017 had high burn severity, i.e. with RdNBR greater than 0.73 (Miller and Thode 2007, Miller and Quayle 2015). Dry years experienced much higher burn severity, with a mean RdNBR of 0.69 (±0.45), than wet years with a mean of 0.44 (±0.39) (figure 3(c)). The Kruskal Wallis test further confirmed this significant difference, with a p-value lower than 0.0001. The warm drought had the most severe fires with a mean RdNBR of 0.75 (±0.43). The mean RdNBR during the cool drought was 0.49 (±0.44), significantly lower than that during the warm drought (p-value < 0.0001). Violin plots also showed a larger proportion of high RdNBR areas during the dry periods, especially when during the warm drought (figure 3(c)).

When summarized by severity categories, we found a significant increasing trend of high severity fires, according to the regression slope test on the annual time series of the percentage of high severity areas burned over the study area (a slope of 0.8% per year with a p-value of 0.018, figure 2(d)). During dry years, 47% of the total burned areas were high-severity burns (figure 3(d)). In contrast, wet years had only 20% high-severity patches and the majority of the burned areas experienced moderate (38%) or low (27%) severity. The percentage of high-severity patches during the warm drought (52%) was almost twice of that during the cool drought (29%), although the proportions of moderate severity patches were similar.

We also found that the areas burned during dry years were mostly shrubs (34%) and hardwood woodland (27%), while denser vegetation, hardwood forest, was the dominant vegetation type burned (40%) during wet years (figure 3(e)). A higher percentage of hardwood woodland burned during the extended warm drought (27%) than during the cool drought (18%). Although the herbaceous plant is the most common vegetation type in the study area, the proportion of herbaceous burned across the four climatic conditions was relatively constant, fluctuating around 10% of the total burned areas. The percent of high-severity burns increased from wet to dry periods for all vegetation types (figure 3(f)). Shrub vegetation had relatively higher burn severity than other vegetation types, e.g. 41% and 73.7% burned at high severity during the wet and dry periods (figure 3(f)). The most recent extended warm drought intensified the burn severity significantly for forests.

We built five RF classification and regression models, for the whole time period and for four individual climatic conditions, respectively, using 35 environmental factors as independent variables (table 1). The five-fold validation showed that both classification and regression RF models were able to capture the variability of burn severity categories and continuous RdNBR during the entire study period and under each climatic condition. The overall RF classification model, based on training data randomly drawn from all pixels burned by fire events since 1984, had an overall accuracy of 79% and a kappa of 0.71, when compared with the remaining testing data (table 2). Among each individual burn severity category, the models also performed well, especially for predicting the high-severity burns, with a user's accuracy of 85% and a producer's accuracy of 90%.

The prediction also successfully captured the spatial patterns of burn severity categories, as illustrated by the predicted burn severity maps for the 1990 Eagle fire during the cool drought, the 2015 Rocky fire during the warm drought, and a series of fires in the 2017 northern California firestorm during the wet period (figures 4(a)–(f)). For instance, for the Rocky fire, 67% of the burned areas were high-severity and the model predicted 73% of the areas as high-severity. The predictions preserved most large severity patches but still has the tendency of omitting certain small low-severity regions. Separate RF classification models for each of the four climatic conditions also achieved similar or higher accuracy (table S3).

Zoom In Zoom Out Reset image size

When trained with randomly-sampled RdNBR observations from areas burned from 1984 to 2017, the overall RF regression model explained 70% and 67% of the variance in the training and testing data across burned pixels from all fires, respectively. The predicted RdNBR values agreed reasonably well with the observations in the testing data set, as shown by the 2D density scatter plot (figure 5), and had a similar distribution with a similar mean value of 0.58 as observations, according to the paired statisticaltest.

Zoom In Zoom Out Reset image size

Topography, long-term climate, and fuels were among the top predictors for burn severity levels (high, moderate, low, and unchanged), as shown by variable importance from the RF classification models (figure 6(a)). Slope and aspect were the two most influential factors in determining burn severity categories across all fires since 1984. The OOB accuracy would decrease by more than 200% if they were randomly permuted from the full model. Similar relative importance was found for several long-term temperature metrics including minimum, mean, and maximum daily temperature (figure 6(a)). Fuel characteristics, as represented by NDVI and EVI values, both right before fire and the annual maximum, and pre-fire NDMI, also played an important role, as well as vegetation type. Removing any of these variables from the full model would result in more than 150% decrease in OOB accuracy. Hillshade, variations in the mean long-term daily minimum temperature, and the long-term mean CWD were also found significant, followed by annual CWD during fire year, elevation, housing density, and long-term mean annual precipitation (figure 6(a)). Wind and socioeconomic factors were among the least important factors.

Zoom In Zoom Out Reset image size

The surrogate classification tree, built on the 16 most important variables, showed how environmental conditions shaped burn severity categories (figure 7). For the training sets for the overall RF classification model, the node first split shrubs vs other vegetation types (figure 7(a)). For shrubs, high severity patches occurred over areas with fire year max NDVI lower than 0.75 or when both NDVI and fire-year CWD were higher. Due to the complexity of burn severity among non-shrubs, we further built individual classification trees for other dominant vegetation types (figures 7(b), (c)). For hardwood woodland, burn severity was relatively low within areas with lower CWD during the fire year, especially when there was high preceding NDMI over areas with long-term mean precipitation between 540 mm yr⁻¹ and 947 mm yr⁻¹, while the moderate burns occurred with lower preceding fuel moisture (figure 7(b)). In contrast, all high-severity burns of hardwood woodland occurred with higher fire year CWD, and the majority were associated with higher long-term mean daily minimum temperature and pre-fire NDVI (figure 7(b)). Hardwood forests were more likely to experience fire with higher severity burns at an elevation higher than 352 m (figure 7(c)). Furthermore, over warmer but wetter areas, i.e. the long-term max temperature higher than 21.1 °C and long-term mean precipitation greater than 949 mm yr⁻¹, high severity hardwood forest burns occurred when annual CWD was higher than 70 mm yr⁻¹; while over the climatologically cooler areas, higher elevation (>770 m) but with higher long-term CWD were more vulnerable to high fire severity (figure 7(c)).

Zoom In Zoom Out Reset image size

Across the whole landscape, results from the regression model validated the importance of fuel characteristics and climate (figure S4). The partial dependence analysis further revealed that higher NDVI right before fire was found to increase burn severity (figure 8(a)). A consistent negative relationship was also found between burn severity and fuel moisture, i.e. RdNBR dropped rapidly when pre-fire NDMI was higher than 0.2 (figure 8(b)). Higher CWD during the fire year was found to be associated with higher RdNBR (figure 8(c)). Interestingly, long-term mean temperature showed bimodal effects on burn severity, i.e. RdNBR increased with 30-year minimum temperature, but started to decrease when the temperature reached around 6.5 °C, and then peaked around 8.5 °C (figure 8(d)). As for topographic factors, both slope and elevation were positively related to RdNBR (figures 8(e) and (f)). RdNBR increased significantly with slope until 15° and then plateaued at approximately 24°, when other variables were held at the mean effects (figure 8(e)).

Zoom In Zoom Out Reset image size

The Kruskal Wallis test further confirmed the significant differences of important environmental factors, as identified from the previous section, over areas burned under the four climate conditions (figure 9). Areas burned during dry years had much lower fuel amount, as represented by prefire NDVI of the year, than fires during wet years, with a mean of 0.49 ± 0.14 vs. 0.63 ± 0.18 (figure 9(a)). However, fuels were much drier, e.g. as shown by much lower pre-fire NDMI of 0.09 ± 0.15 over those burned during dry years as compared to 0.19 ± 0.19 in wet years (figure 9(b)); The mean fire-year CWD over dry year fires was 48 mm higher than those during wet years (figure 9(c)). Compared to those in wet years, fires during droughts tended to occur at areas with a slightly lower long-term minimum temperature (8.13 ± 0.84 °C vs. 8.23 ± 1.08 °C) but higher temperature during fire month (22.6 ± 2.38 °C vs. 21.7 ± 1.02 °C) and higher mean annual temperature during fire year (16.23 ± 0.96 °C vs. 15.13 ± 0.63 °C) than during wet years (figures 9(d)–(f)). The differences of environmental variables between warm and cool droughts were significant, but much smaller. These evidences suggest that the burn severity was intensified during dry years mostly by the drought conditions.

Zoom In Zoom Out Reset image size

The RF classification model during dry years identified a set of similar top ten important drivers for categorical burn severity, with an accuracy decrease of more than 150% if removing any one of those variables (figure 6(b)). Burn severity was mainly shaped by a mixed effect of aspect, long-term temperature and CWD, NDVI, and variations in long-term temperature during dry years (figure 6(b)). In contrast, during the wet years, besides topographic effects (slope and elevation), both long term mean and fire-year CWD were found dominating burn severity, with a decrease in OOB accuracy ranging from 200%–300 % (figure 6(c)); additionally, long-term minimum temperature and annual maximum NDVI also had relatively strong impacts on burn severity. As for the differences between the warm and the cool drought (figure S3), severity of warm drought fires was shaped by the mixed effects of aspect, long-term temperature, and fuel availability, while cool drought burn severity was mainly driven by slope, followed by fuel type, CWD, and temperature.

Topographic factors, such as slope and aspect, strongly controlled burn severity across the landscape of California's northern coastal Mountains. This is similar with the findings in Sierra Nevada and Klamath mountains of northern California by Kane et al (2015) and Estes et al (2017). In several ecosystems of the western US, however, previous studies indicated more marginal effects of topography on severity (Thompson and Spies 2010, Parks et al 2018). The differences in the relative importance of variables may depend on the study area and the associated topographic variability. For example, Thompson & Spies only examined the 2002 Biscuit fire in southwestern Oregon, which had relatively low variability in elevation. Across the western US, topography was found with negligible influence on burn severity (Parks et al 2018), but for California ecoregions their high-severity probability model based on nine environmental features still showed significant local effects of topography on burn severity.

Climatic and weather extremes have long been argued driving the increase in the intensity of wildfire size and severity (Bessie and Johnson 1995, Thompson and Spies 2010). A study in the western US showed that high temperature and precipitation extremes intensified droughts and accumulated CWD, which further increased burn severity (Crockett and Westerling 2018). Our results showed that both long-term minimum temperature and long-term CWD were key factors affecting the variation of burn severity across all fires since 1984, followed by annual CWD during the fire year, especially over areas burned during the wet periods. Under most climatic scenarios, weather conditions, such as temperature and precipitation, during the fire year/month were not among the top important variables, except during the wet years. This might be because the indirect impacts of other preceding weather have been represented by their influence on fuel characteristics variables such as NDMI in our study, consistent with findings from previous studies (Lutz et al 2010, Parks et al 2018).

Wind speed and direction at 4 km resolution had low predictive ability according to our model results for the whole time periods and warm droughts, probably due to the spatial resolution of wind datasets. The spatial variations of wind patterns within fire perimeters could not be captured by the coarse resolution GRIDMET wind data (Abatzoglou 2013). When compared among the four climatic conditions, however, wind speed had relatively the highest importance for burn severity during wet years (figure 6). This is consistent with some previous studies showing an increasing wind effects on recent extreme fire events, such as the wet year 2017 with extreme fires (Nauslar et al 2018, Mass and Ovens 2019). Further studies are needed when high-resolution wind data becomes available, to separate fires into fuel- and wind- dominant types for a better quantification of the impacts of wind on burn severity (Keeley and Syphard 2019).

In addition, although we considered the potential lag effect of precipitation, e.g. previous 1- to 3- year cumulative rainfall, the potential compounding effects of the transition from a prolonged multi-year drought to a wet year are worth a comprehensive study. For example, the catastrophic 2017 fire storms in the wine counties coincided with strong Diablo wind events, a relatively wet winter, and preceding four-year drought (Nauslar et al 2018). In California, the increasing frequency of a rapid transition from extreme dry-to-wet precipitation events has been observed in the past 2 decades (Berg and Hall 2015, Swain et al 2018). This climate whiplash is projected to continue to increase through mid-to-late of the century, on top of extreme temperature, further worsening future severe fire risk in the region (Swain et al 2018).

Fuel characteristics, such as fuel loading and moisture, were key variables in determining burn severity across all fires, according to both classification and regression models. This is consistent with previous findings that pre-fire stand conditions affect burn severity significantly both within individual fires and across fires in western US (Bigler et al 2005, Thompson and Spies 2010, Birch et al 2015, Estes et al 2017, Safford and Stevens 2017, Parks et al 2018). Our results showed that fuel levels played an important role in burn severity across the landscape, as indicated by NDVI and annual maximum EVI. Interactions may exist among NDVI, fuel type, and vegetation height. Across the four climatic scenarios, fuel amount was found to have stronger impacts on burn severity when under warm and dry conditions. A possible explanation is that fuels desiccate faster during droughts (Bigler et al 2005). Accumulations of dry fuels might thus increase burn severity during droughts.

Across the landscape, the moisture of fuels was affected by long term climate such as 30 year mean CWD and further regulated by short term weather, e.g. both lack of precipitation and higher temperature drying out fuels. NDMI is a proxy of leaf water content from Landsat remote sensing observations and indicates fuel moisture. It has been found to serve as a robust driver of severity (Estes et al 2017; Parks et al 2014b). In our study, lower NDMI before fire was found significantly associated with higher severity of burns.

Wildfire adaptation, mitigation strategies, and management practices are needed for local communities, to reduce future burn severity, in response to increased warming and drying and associated increasing wildfire activity (Hessburg et al 2015). The predictive capability of our RF models is expected to help us anticipate the probability of high severity burns, based on near real-time fuel information from satellite observations and weather, in addition to the baseline climate. For example, the predictions showed a very heterogeneous pattern of high burn severity across the whole study area and also suggested an increased high severity probability during the dry years than wet years, especially in the northwest and southern part of the region (figure S7). Our model, however, used the explicit near real-time information from Landsat satellite observations, which limits the potential for projecting burn severity under the future climate change scenarios. Future studies are needed to incorporate the dynamic vegetation and fuel models for a more realistic projection. Overall, the RF classification model has the capability to predict the probability of each burn severity category in near real time or forecast the short-term burn severity at a 30 m scale in the study region if a fire would occur, driven by current and preceding fuel information and the corresponding weather data.

Although factors like long-term temperature or temperature variations are externalities to local management, this study can help us to identify areas subject to high fire severity, to design corresponding adaptation strategies to mitigate the impacts of future extreme weather events for both ecological and human systems. Our study highlighted that the topographic and long-term climate conditions are associated with severe burns along with fuels. For example, more mitigation efforts could beneficially be focused on steep slopes and areas with higher long-term minimum temperature and CWD, or areas where the projected increase in minimum temperature and CWD is more pronounced.

Our study also suggested that the degree to which one could manage the landscape varies with climate conditions. For example, fuel reduction and management may be helpful in reducing future risks of high severity burns, especially under dry and warm conditions, but it may be less effective during wet years. Our approaches can also be applied to other Mediterranean coastal mountain ranges, e.g. western Australia or the Mediterranean Basin, which are typically subject to high fire risk and have extensive wildland-urban interface (Rundel 1998, Fernandes 2013).