Articles | Volume 25, issue 1
https://doi.org/10.5194/nhess-25-383-2025
https://doi.org/10.5194/nhess-25-383-2025
Research article
 | Highlight paper
 | 
27 Jan 2025
Research article | Highlight paper |  | 27 Jan 2025

Modelling current and future forest fire susceptibility in north-eastern Germany

Katharina H. Horn, Stenka Vulova, Hanyu Li, and Birgit Kleinschmit
Abstract

Preventing and fighting forest fires has been a challenge worldwide in recent decades. Forest fires alter forest structure and composition; threaten people's livelihoods; and lead to economic losses, as well as soil erosion and desertification. Climate change and related drought events, paired with anthropogenic activities, have magnified the intensity and frequency of forest fires. Consequently, we analysed forest fire susceptibility (FFS), which can be understood as the likelihood of fire occurrence in a certain area. We applied a random forest (RF) machine learning (ML) algorithm to model current and future FFS in the federal state of Brandenburg (Germany) using topographic, climatic, anthropogenic, soil, and vegetation predictors. FFS was modelled at a spatial resolution of 50 m for current (2014–2022) and future scenarios (2081–2100). Model accuracy ranged between 69 % (RFtest) and 71 % (leave one year out, LOYO), showing a moderately high model reliability for predicting FFS. The model results underscore the importance of anthropogenic parameters and vegetation parameters in modelling FFS on a regional level. This study will allow forest managers and environmental planners to identify areas which are most susceptible to forest fires, enhancing warning systems and prevention measures.

1 Introduction

Over the past decades, climate change has led to a higher intensity and frequency in extreme weather events all over the planet (Kemter et al.2021; Silva et al.2018; Wu et al.2021). In Germany, very low precipitation has occurred more frequently in the last 6 years, leading to an increased number of forest fires (Gnilke and Sanders2021). Long periods of drought have been causing soils and vegetation to dry out substantially. Especially in forests, the drying out of trees, underground vegetation, litter, and soils is making forests highly flammable (Littell et al.2016). Consequently, it is crucial to understand the conditions that cause the emergence and spread of forest fires as well as to detect the areas that are most prone to forest fires (Ambadan et al.2020). This way, forest fire prevention and management strategies can be improved, decreasing the subsequent potential threats to forests, the population, and infrastructure located in proximity to forests. In the long run, this may also decrease the financial costs of climate change (Chicas and Østergaard Nielsen2022).

Apart from meteorological conditions, forest fires are influenced by a number of environmental factors, including soil moisture, topography, sun exposition, lightning strikes, and wind (He et al.2022; Saidi et al.2021; Wang et al.2021). Moreover, they are closely linked to human influence, encompassing the expansion of infrastructure in proximity to forests, as well as the utilisation of forests for recreational purposes (Ghorbanzadeh et al.2019). On a European scale, a study by El Garroussi et al. (2024) shows that 96 % of wildfires are triggered by human influence. In a similar vein, Gnilke and Sanders (2021) state that up to 50 % of the area burnt by forest fires in Germany is caused by human action. German forest fire statistics identified human negligence as the most important factor in the occurrence of forest fires (Federal Office for Agriculture and Food2023). Thus, anthropogenic influences should be carefully considered along with other parameters when analysing forest fires (He et al.2022; Ruffault and Mouillot2017).

Forest fires and the assessment of meteorological, climatic, and anthropogenic parameters have been addressed in numerous studies. Some of them analyse the fire risk of certain regions (Ambadan et al.2020; Saidi et al.2021), whereas others focus on the identification of parameters influencing forest fire emergence (He et al.2022; Ruffault and Mouillot2017). For example, Saidi et al. (2021) developed a GIS–remote sensing approach to investigate forest fire risk in Tunisia, whereas He et al. (2022) studied the drivers of bushfires in New South Wales, Australia, over a time period of 40 years. The current state of research on forest fires suggests that topography, climate, land use, and anthropogenic influences are the most influential parameters (Abdollahi and Pradhan2023; Cilli et al.2022; Ghorbanzadeh et al.2019; He et al.2022; Ruffault and Mouillot2017; Saidi et al.2021; Li et al.2024). For example, Ruffault and Mouillot (2017) consider human influence, land cover, and weather conditions for the assessment of influencing factors for wildfires in the French Mediterranean region.

Forest fire susceptibility (FFS) can be analysed with a variety of methodological approaches, including knowledge-based approaches, such as hierarchical weighting (Busico et al.2019), machine learning (ML) and statistical approaches, or hybrid approaches (Chicas and Østergaard Nielsen2022). ML algorithms include random forest (RF) models (Cilli et al.2022; He et al.2022; Milanović et al.2021; Oliveira et al.2012, 2016), boosting models (Ruffault and Mouillot2017; Wang et al.2021), and artificial neural networks (Ghorbanzadeh et al.2019). Previous research on FFS has focused on bigger research areas (Busico et al.2019; He et al.2022; Saidi et al.2021), whereas research on a smaller scale has fallen short. However, geodata and remote sensing data at high spatial resolution allow for detailed analysis to enhance forest fire research on a local scale. Especially regarding climate change and the growing likelihood of weather extremes such as droughts, local FFS modelling is essential for identifying key drivers on a local scale. This way, improved prevention and management strategies of forest fires can be provided. While future climate data now enable the modelling of future forest fire susceptibility (FFS), those types of studies remain scarce (Busico et al.2019), indicating significant untapped potential for enhancing forest fire prevention efforts.

This study focuses on the analysis of forest fires in Brandenburg, Germany. Due to a high percentage of coniferous forest, this federal state has been particularly prone to forest fires in the past. Furthermore, remnants of old munitions at former military training sites caused forest fires in Brandenburg in 2018 and 2019 (Gnilke et al.2022). Although this issue has been addressed by German newspapers, it has received minimal attention in scientific research (Feng et al.2022). Therefore, this study aims to predict FFS in Brandenburg under two current (2016 and 2022) and two future scenarios (2081–2100) using geodata and remote sensing data at high spatial resolution and the random forest (RF) machine learning (ML) algorithm. Following Zhang et al. (2019), FFS in this study represents “the probability estimation of fire occurrence”. In addition to topographic, vegetation, and soil parameters, this study incorporates a comprehensive set of anthropogenic and land use parameters, including new predictors such as the distance to campsites and military training sites, to expand existing research on forest fires. To our knowledge, only a few studies have analysed FFS at a high spatial resolution so far (Ghorbanzadeh et al.2019; Suryabhagavan et al.2016; Razavi-Termeh et al.2020; Pourtaghi et al.2015), and we do not know of any studies that modelled future FFS at a high spatial resolution. Within the scope of this investigation, the following research questions will be answered:

  • a.

    Which variables are most significant in terms of forest fire spread in north-eastern Germany?

  • b.

    Which areas in Brandenburg are most susceptible to forest fires now? How will these areas change considering future climate conditions?

2 Materials and methods

2.1 Study area

The federal state of Brandenburg (Fig. 1) was selected as the study area for modelling FFS under current and future scenarios. Brandenburg is located in the north-east of Germany. With sandy or sandy–loamy soils and a high number of rivers and lakes, the federal state is characterised by a periglacial landscape. Agriculture and managed forests are the main land uses. The forests are dominated by pine trees (Pinus sylvestris L.) (Matos et al.2010), and the climate is characterised by rather dry summer months. The combination of these conditions is linked to a medium to high forest fire risk (Holsten et al.2009; Matos et al.2010; Reyer et al.2012; Thonicke and Cramer2006). Comparing all German federal states, Brandenburg has been most affected by forest fires (Gnilke and Sanders2021), which is why it was selected for this study.

https://nhess.copernicus.org/articles/25/383/2025/nhess-25-383-2025-f01

Figure 1The federal state of Brandenburg in north-eastern Germany. Basemap © 2024 TerraMetrics, Google, GeoBasis-DE/BKG (© 2009). Border layers © BKG (2024) dl-de/by-2-0 (data not changed).

2.2 Current and future forest fire susceptibility scenarios

The aim of this research is to compare FFS under different temporal scenarios. To do so, current and future FFS in the federal state of Brandenburg was modelled. To represent the current state, the years of 2016 and 2022 were selected after carefully analysing the monthly precipitation sums and mean monthly air temperature of Brandenburg between 2014 and 2022 (see Figs. S1 and S2 in the Supplement). Based on this analysis, 2016 was characterised by average climatic conditions, whereas 2022 was characterised by conditions of drought (low precipitation rates). Consequently, the 2016 scenario was considered a baseline scenario with average climatic conditions. In contrast to 2016, the 2022 scenario represents a very dry year, which can be expected to occur more frequently due to the expected increase in extreme weather events in the future (Silva et al.2018; Wu et al.2021).

The future scenarios of FFS cover the period of 2081 to 2100 using SSP5-8.5 (Shared Socioeconomic Pathway). SSPs are different projections of future greenhouse gas emissions under distinct potential political and socioeconomic developments. The SSPs range from SSP1-1.9 to SSP5-8.5, covering CO2 concentrations ranging from 393 to 1135 ppm until 2100. SSP5-8.5 represents “a high fossil-fuel development world throughout the 21st century” (Meinshausen et al.2020). We decided to use SSP5-8.5 from the global climate model (GCM) MPI-ESM-1-2-HR. Xu et al. (2023) state that this GCM reflects future drought conditions rather well, which is why it was selected for this study. The climate data (monthly average minimum temperature (°C), monthly average maximum temperature (°C), and monthly total precipitation (mm)) were downloaded from WorldClim (https://www.worldclim.org, last access: 22 December 2023). This website provides gridded multi-annual datasets based on different GCMs for different Shared Socioeconomic Pathways (SSPs) and different time periods between 2021 to 2100 up to 30 arcsec (∼1 km) spatial resolution. In order to include future land cover changes into the future predictions, future FFS was predicted twice: (a) once including only projected meteorological data for 2081–2100 and (b) once including projected meteorological data for 2081–2100 and projected land cover data. Within Figs. 2, 4, 5, 6, and 7, as well as in Table 3, the latter will be labelled with an asterisk (*). Additionally, a third future scenario based on the SSP3-7.0 was predicted. The results can be found in Figs. S10–S13. After analysing the monthly frequency of forest fires in the federal state of Brandenburg, the month of June was selected for the prediction of the four scenarios, since forest fire data showed the highest number of forest fires in this month between 2014 and 2022 (Lower Forestry Authority of the State of Brandenburg2023). For model training, we used all available forest fire events of all months between 2014 and 2022 and pre-processed climatic datasets in accordance with the available forest fire data.

2.3 Data

2.3.1 Forest fire data

To model FFS in Brandenburg under different scenarios, forest fire data as well as a set of predictor datasets were acquired and pre-processed. Data including statistical and geospatial information on forest fires in Brandenburg were provided by the Lower Forestry Authority of the State of Brandenburg (2023), an institution that focuses on analysing the vitality of forests in the federal state (Lower Forestry Authority of the State of Brandenburg2023; Ministry for Rural Development, Environment and Agriculture in Brandenburg2023). The Lower Forestry Authority of the State of Brandenburg (2023) provided data containing the following information: forest district number, section, date and hour, cause of fire, burnt area (ha), and xy coordinates of the fire ignition point.

2.3.2 Predictor variables

To model FFS in Brandenburg, a set of 20 predictors were selected for the analysis. The predictor variables are shown in Table 1 (also see Fig. S4). They cover meteorology, vegetation, topography, soil, anthropogenic influences, and land use and land cover (LULC) and were identified as most relevant to modelling FFS based on an extensive literature review. In the following sections, the predictor variables will be presented in more detail.

DWD Climate Data Center (2023a)Fick and Hijmans (2022)DWD Climate Data Center (2023b)Fick and Hijmans (2022)Lang et al. (2023)Poggio et al. (2021)EEA (2020b)OpenStreetMap Contributors (2023)Esri Environment (2021)

Table 1Predictor variables for modelling forest fire susceptibility in Brandenburg.

The topographic predictors – slope, aspect, and TWI – were computed using the digital elevation model provided by the LGB State Office for Land Surveying and Geoinformation Brandenburg (2023).

Download Print Version | Download XLSX

(a) Meteorology

To assess climatic conditions for both the current and future scenarios, air temperature and precipitation were selected. Since climate change and the consequent increase in extreme weather events such as meteorological droughts around the world may increase the frequency and intensity of forest fires in the future (Abdollahi and Pradhan2023; Silva et al.2018), air temperature and precipitation patterns are crucial for the analysis of FFS. Further climatic parameters such as wind speed, solar radiation, or lightning strikes may impact the emergence of forest fires as well (Abdollahi and Pradhan2023; Busico et al.2019). However, for the scope of this work the focus remained on air temperature and precipitation, since both current and projected data were only available for those climatic parameters. Following the suggestions by He et al. (2022), we used monthly climate data between 2013 and 2022, which were aggregated to 3 months to incorporate precipitation and air temperature prior to the occurrence of a forest fire. Several forest-fire-related studies have used a monthly aggregation of meteorological datasets to model forest fires (Busico et al.2019; Wang et al.2021; He et al.2022). He et al. (2022) further argue that future studies should consider a monthly or quarterly aggregation of meteorological data when investigating forest fires. In particular, in order to identify conditions of meteorological droughts prior to the emergence of a forest fire, we followed the methodology of other authors that used a 3-month aggregation of the broadly used SPEI (standardised precipitation evapotranspiration index) drought index to identify meteorological droughts (Zhou et al.2023; Wen et al.2020; Guo et al.2018).

(b) Vegetation

The type and condition of vegetation is a crucial factor in the emergence of forest fires (Abdollahi and Pradhan2023). Several studies have shown that monocultural forests are more likely to be affected by forest fires not only in number but also in extent (Afreen et al.2011; Bauhus et al.2017). For example, Bauhus et al. (2017) state that coniferous species such as pine trees tend to be highly flammable, which is mainly caused by their resins and oils. Furthermore, the distance to the forest edge can impact tree vitality and the consequent vulnerability to droughts (Buras et al.2018). Buras et al. (2018) analysed the tree mortality of Scots pine forests by comparing trees on the forest edge and trees in the interior of the forests. Their results show an increase in vulnerability to drought of trees located at forest edges, resulting in higher mortality and decreased vitality. Consequently, the selected vegetation-related predictors were the percentage of broadleaf forest, canopy height, tree cover density, and the distance to forest edges.

(c) Topography

Numerous studies have shown the influence of topography on the emergence of forest fires, which is why topographic parameters are commonly used for studying forest fires (Abdollahi and Pradhan2023; Busico et al.2019; Ghorbanzadeh et al.2019; He et al.2022; Maingi and Henry2007; Saidi et al.2021; Wang et al.2021). For example, Preston et al. (2009) have pointed out that bushfires spread with a higher velocity and intensity on upward slopes. Furthermore, they discuss how aspect impacts sun and wind regimes, which may influence forest fires as well. In this regard, Busico et al. (2019) conclude that northern aspects decrease the likelihood of forest fire ignition. Besides slope and aspect, elevation has been pointed out as a significant parameter for forest fires (He et al.2022; Maingi and Henry2007). Chicas and Østergaard Nielsen (2022) performed an extensive analysis of existing studies on mapping FFS, confirming that slope, elevation, aspect, and the topographic wetness index (TWI) are the most commonly used topographic parameters. Following their assessment, those four parameters were selected for the scope of this study.

(d) Soil

The spread of forest fires is greatly influenced by the characteristics of the soil and its moisture content (He et al.2022). Therefore, it was considered important to include different soil characteristics as predictor variables. The soil depth chosen for the soil predictors was 0–5 cm, since fires are usually initiated on the soil surface (Badía-Villas et al.2014; Mallik et al.1984). The water retention capacity of soils is significantly influenced by their structure, such as the relative proportions of sand and silt. Soil types characterised by larger pore sizes, such as sandy soils, typically exhibit low water retention capabilities, leading to arid conditions and a diminished field capacity. Conversely, soils with intermediate pore sizes or silty soils have higher moisture levels and more water available for plants (Amelung et al.2018). Therefore, the proportion of sand particles (>0.05 mm) in the fine earth fraction (sand) and the proportion of silt particles (≥0.002 mm and ≤0.05 mm) in the fine earth fraction (silt) were selected for the analysis. Similarly, both bulk density of the fine earth fraction (bdod) and organic carbon density (ocs) can serve as proxies for water retention and therefore for the flammability of the soil (Oyonarte et al.1998). For example, Oyonarte et al. (1998) have shown a high correlation between water retention and organic carbon, as well as bulk density, which underlines their potential influence on FFS. Thus, bulk density of the fine earth fraction and organic carbon density were used as predictor variables as well.

(e) Anthropogenic influences and land use and land cover (LULC)

Finally, anthropogenic factors as well as LULC have been shown to influence the emergence of past forest fires in Brandenburg (Gnilke and Sanders2021). The data provided by the Lower Forestry Authority of the State of Brandenburg (2023) on causes of forest fire ignitions in Brandenburg between 2014 to 2022 (see Table S2) confirm this statement. In a similar vein, He et al. (2022) argue that human activities such as the construction of transportation networks and other types of infrastructure influence forest fire emergence on a local scale. Therefore, they highly recommend including anthropogenic factors into the analysis of forest fires. Likewise, Ghorbanzadeh et al. (2019) relate the increase in forest fires not only to the changing climate but also to anthropogenic aspects such as human activities or demographic expansion. Thus, to predict FFS in northern Iran, they included the proximity to villages, streets, and recreational areas, as well as aspects of land use, as predictor variables. The latter has been emphasised by Busico et al. (2019) as well, who stated that anthropogenic land use significantly contributes to forest fire emergence. Consequently, to include anthropogenic influences as well as aspects of LULC, the distances to urban settlements, streets, railways, campsites, waterbodies, and military sites were selected as predictor variables. According to the respective dataset, we understand the “distance to urban settlements” as the distance to any type of constructed above-ground building (EEA2020b). We assume that this predictor can show (ir)regular human presence at these places that may be related to an increased FFS. Furthermore, to address future land cover changes, we included a dataset on projected land cover change in 2050 provided by Esri Environment (2021). To our knowledge, this was the only available dataset with a high spatial resolution to show future land cover changes, which is why it was selected for this study. Table 1 provides an overview of the predictors as well as their characteristics and origin.

2.4 Data processing

RStudio version 2023.12.0.369 with R version 4.3.1 (2023-06-16 ucrt) was used for data pre-processing; analysis; RF modelling; and the computation of statistics, graphs, and maps. Geospatial packages such as terra, sf, maptools, and ggplot2 were used for data pre-processing and analysis. The caret package was used for modelling and the computation of performance metrics. The dplyr and readxl packages were used for the analysis and formatting of the forest fire data. The open-source software QGIS 3.28.10 Firenze was used for processing, analysis, and visualisation of the geodata. Figure 2 provides an overview of the main data processing steps that will be explained in the following sections.

https://nhess.copernicus.org/articles/25/383/2025/nhess-25-383-2025-f02

Figure 2Methodological approach for modelling forest fire susceptibility under different scenarios.

Download

(a) Pre-processing of predictor layers

Prior to modelling FFS under current and future scenarios, the necessary datasets were downloaded and pre-processed. Pre-processing steps involved projecting the data to the same coordinate reference system (EPSG:25833), cropping to the geographic extent of Brandenburg, masking the forest areas in Brandenburg, and resampling to a spatial resolution of 50 m using bilinear interpolation for numeric variables and nearest-neighbour interpolation for factor variables. Furthermore, several predictor datasets such as the distance to campsites or military areas were created based on available data from OpenStreetMap Contributors (2023) or the LGB State Office for Land Surveying and Geoinformation Brandenburg (2023). The topographic predictors – slope, aspect, and TWI – were computed using the digital elevation model derived from the LGB State Office for Land Surveying and Geoinformation Brandenburg (2023). A forest mask was generated by filtering all pixels with tree cover density greater than or equal to 50 % from the tree cover density dataset. Proximity rasters were computed for various features, including urban settlements, roads, railways, military sites, campsites, waterbodies, and forest edges, by applying the “Proximity (raster distance)” tool in QGIS derived from the GDAL (Geospatial Data Abstraction Library) toolbox.

(b) Processing of training points

The forest fire data table provided by the Lower Forestry Authority of the State of Brandenburg (2023) served as the baseline for the creation of the training points for the RF models. Rows containing NA (not available) values were removed, and the fire data points were converted to the shapefile format for further processing. Looking at the statistics of the burnt area (ha) of each of the fires in Brandenburg between 2014 to 2022, the maximum burnt area of a forest fire was 422 ha. In contrast, the median burnt area was only 0.05 ha, indicating a high number of small fires and a relatively low number of big fires (see Table S1 in the Supplement). Since the spread extent of the fires was not included in the data provided by the Lower Forestry Authority of the State of Brandenburg (2023), a circular fire spread was assumed. The diameter of a circular burnt forest fire based on the median burnt area (0.05 ha or 500 m2) is ∼25 m. Considering that the direction of the fire spread was unknown as well, the doubled diameter of a median-sized forest fire in Brandenburg (50 m) was assumed as a baseline for converting the forest fire points into a raster dataset (see Fig. S3). Consequently, the fire points were resampled to a raster grid with 50 m spatial resolution considering the potential fire spread in different directions. Accordingly, all the predictor variables were resampled to the same spatial resolution.

In addition to the provided set of fire points, a set of non-fire points was created that included the identical number of points per year as the pre-processed fire points from the data table provided by the Lower Forestry Authority of the State of Brandenburg (2023). To create those non-fire points, the maximum extent of each forest fire for each year was computed to identify areas where no fires occurred for each year. To do so, the fire point data table was first subsetted by year and then burnt area was estimated based on the previously described approach. The results were nine raster layers for each year between 2014 and 2022 that contained the maximum extent that was potentially burnt in that respective year. For each year, potentially burnt areas were then removed from the forest mask layer to derive areas where no fires occurred. Based on the forest masks that excluded potentially burnt areas, random non-fire points were created for each year, matching the number of fires that occurred in the respective year. To do so, the randomPoints() function from the R package raptr was used.

Finally, the resulting non-fire points were merged with the fire points to complement the training points. To do so, the training points were assigned to the classes of “fire” and “non-fire”, respectively. Each fire registered by the Lower Forestry Authority of the State of Brandenburg (2023) was paired with a non-fire point with the same date. To prepare the data frame for the RF models, the training points were used to extract the geospatial information of the predictor variables using the extract() function from the terra R package. The resulting data table included the spatial coordinates of all non-fire and fire points and the information of all the predictor variables at those locations. This data frame served as the basis for training RF models to predict FFS under current and future scenarios.

2.5 Correlation analysis and random forest modelling

To assess FFS in Brandenburg under different temporal scenarios, an RF classification ML algorithm was used. In particular, a total of 10 RF models were run using binary classes (fire and non-fire) for predicting current and future FFS. RF is a well-known and often-used ML algorithm in forestry and remote sensing applications (Gislason et al.2006). In the field of forest fire research, RF has been frequently applied, achieving high levels of accuracy (Eslami et al.2021; He et al.2022; Lizundia-Loiola et al.2020; Milanović et al.2021; Oliveira et al.2016). The RF algorithm is based on the bagging approach, developed by Breiman (1999). It involves the growth of a set of random decision trees to form what is known as a “random forest” (Breiman2001; Kuhn and Johnson2013). As mentioned before, FFS is defined in this study as the estimated likelihood of a forest fire event (Zhang et al.2019). The probability score of a pixel being predicted as a fire pixel represents its susceptibility to a forest fire.

2.5.1 Model for future scenarios

First, a model (RFfuture) containing data from all the available years (2014 to 2022) was set up for the prediction of future FFS scenarios. Following Nguyen et al. (2021), the input data for modelling FFS were split into 70 % for model training (RFtrain) and 30 % for testing the model performance (RFtest). We refer to the 30 % of left-out data as the testing dataset. Before running a RF model, a set of tuning parameters can be set. After initially running the model, the results showed the best model performance at mtry=2. Consequently, the model was run with mtry set to 2.

2.5.2 Models for current scenarios

For current FFS scenarios, a so-called “leave-one-year-out” (LOYO) approach was implemented in order to evaluate the models' capacity for temporal extrapolation. Leaving one year out from training and using the excluded year for testing can be used to assess how models will perform on an unseen (or future) year. In this case, the approach was used for modelling current FFS for the scenarios of 2016 and 2022. LOYO models were computed for all nine available years (2014 to 2022). For instance, LOYO2016 refers to a model trained on all years except 2016, which was used to predict FFS in 2016. As mentioned before, mtry was set to 2 to be consistent with the model for the future FFS scenarios.

2.5.3 Performance metrics

After training the RF models, performance metrics were calculated using the caret and rPROC packages. The confusionMatrix() function provides information on the different performance metrics such as accuracy, kappa, sensitivity, or specificity. Additionally, the F1 score and AUC (area under the curve) were computed using the rPROC package in RStudio. The AUC was calculated by first computing the receiver operator characteristic (ROC) curve using the roc() function. The formulas for calculating the different performance metrics can be found in Table S3. They typically range between 0 and 1, with values close to 1 implying high model performance.

3 Results

3.1 Model accuracy

To assess the reliability of the RFfuture model in predicting FFS in Brandenburg, performance metrics and a confusion matrix (see Table S4) were computed. The training (RFtrain) and testing set (RFtest) for the RFfuture models consisted of 3243 and 1388 points, respectively. A total of 487 out of 681 fire points and 520 out of 707 non-fire points were correctly classified. The performance metrics (Table 2) for both RFtest and the LOYO cross validation all range between 0.654 and 0.718 (excluding the kappa values), showing a moderately high model reliability of predicting FFS in Brandenburg. RFtest had an accuracy of 0.718, reflecting the number of samples that were correctly classified as fire points. The LOYO cross validation indicates a marginally lower mean accuracy of 0.695. The precision values of LOYO cross validation (0.702) and RFtest (0.712) illustrate the proportion of correctly assigned fire points out of all samples that were classified as fire. To further assess the performance of the RF FFS classification, the ROC curve was computed. The area under the ROC curve (AUC) refers to the likelihood that a fire point was correctly classified (Bradley1997). Here, the AUC is 0.694 for the LOYO cross validation and at 0.718 for RFtest. Finally, recall and F1 score metrics show similar values, indicating moderately high model reliability. A detailed overview of all the performance metrics for every LOYO model can be found in Table S5.

Table 2Overview of the validation metrics.

Download Print Version | Download XLSX

3.2 Importance of predictor variables

Overall, the distance to urban settlements, the percentage of broadleaf forest, and the distance to railways were the three most significant predictors for the RFfuture model. The importance of these predictors, as well as others, is shown in Fig. 3. Land use and anthropogenic predictors exhibited moderate to high influence for the model, such as the distance to urban settlements (100 %), the distance to railways (84.3 %), or the distance to campsites (50.9 %). Similarly, vegetation predictors showed varying degrees of influence, ranging from moderate (e.g. distance to forest edge) to high parameter importance, notably the percentage of broadleaf forest (87.8 %). Soil predictors demonstrated medium importance, ranging from 39.9 % for organic carbon density to 53.4 % for silt content. Topographic predictors displayed varied importance, with elevation at 49.1 % and the TWI at 11.6 %. In contrast, climatic variables had a relatively minor influence on model performance, with air temperature contributing only 14.4 % and precipitation accounting for a mere 3.1 %. The value distributions of the three most significant predictors are depicted in Fig. S5. A Wilcoxon test was conducted to test significance. The notably low p values of the Wilcoxon tests, for example p=5.70×10-20 for the percentage of broadleaf, confirm that the value distributions of all three predictors significantly differ between fire and non-fire points. A comprehensive overview of the p values for all predictor variables is provided in Table S6.

https://nhess.copernicus.org/articles/25/383/2025/nhess-25-383-2025-f03

Figure 3Variable importance based on the RFfuture model.

Download

The value distributions of the three most significant predictors (Fig. S5) lead to several conclusions. First, fire points tend to be closer (mean of ∼578 m) to urban settlements than non-fire points (mean of ∼813 m). Second, the distribution in the percentage of broadleaf forest mainly ranges from 0 % to almost 40 % for non-fire points, whereas the percentage of broadleaf forest for fire points is close to 0 % (excluding some outliers). Third, similarly to the distance to urban settlements, non-fire points tend to be further away from railways than fire points. To more deeply explore the relationship between key variables and FFS, partial dependence plots were created (see Figs. S7–S9).

3.3 Forest fire susceptibility under current and future scenarios

Figure 4 shows FFS in Brandenburg for the two current scenarios, June 2016 and June 2022, as well as for the two future scenarios, June 2081–2100 under SSP5-8.5 and June 2081–2100 under SSP5-8.5 including projected land cover data. For comparison, FFS for June 2081–2100 under SSP3-7.0 can be found in Fig. S10. The values range from 0 % to 100 %, reflecting the likelihood of fire ignition at each pixel (FFS). In all four scenarios, FFS is higher in the southern part of Brandenburg. Especially in the south of Berlin, several patches with a FFS of ≥75 % can be identified. In the north and north-east of Brandenburg however, FFS is rather low in all the scenarios, ranging between 0 % and 20 %.

https://nhess.copernicus.org/articles/25/383/2025/nhess-25-383-2025-f04

Figure 4Forest fire susceptibility in Brandenburg under different scenarios. The scenarios in (c) and (d) both show predicted FFS in June 2081–2100 under SSP5-8.5. The scenario in (d) includes projected land cover data, whereas the scenario in (c) does not. Border layer © 2018–2022 GADM.

Figure 5 illustrates the anomalies in FFS relative to the June 2016 reference scenario. In the June 2022 scenario (scenario a), FFS exhibits notable positive anomalies across various regions of the federal state, with anomalies ranging from +5 % to +15 % compared to June 2016. Many areas across Brandenburg maintain FFS levels similar to the 2022 scenario. Only a few selected small regions in the south-east and south-west exhibit negative FFS anomalies compared to June 2016. Regarding future FFS anomalies relative to June 2016, the future scenarios differ rather substantially from one another. Whereas the scenario neglecting land cover changes (scenario b) shows positive FFS anomalies up to 15 % and more in southern, eastern, and western parts of Berlin, one area in the south shows negative FFS anomalies up to −20 %. In comparison to the scenario based on only climatological projections, the scenario incorporating land cover changes (scenario c) shows mostly negative FFS anomalies ranging from 0 % to −20 %, especially in the southern part of Brandenburg. The northern part of Brandenburg however is characterised by an increase in FFS in many areas, reaching anomalies up to +20 %. Additionally, some areas in the south and west also show positive FFS anomalies. For comparison, the FFS anomalies for 2081–2100 under SSP3-7.0 can be found in Figs. S11–S13.

https://nhess.copernicus.org/articles/25/383/2025/nhess-25-383-2025-f05

Figure 5Forest fire anomalies compared to 2016. The scenarios in (b) and (c) both show predicted FFS anomalies in June 2081–2100 under SSP5-8.5. The scenario in (c) includes projected land cover data, whereas the scenario in (b) does not. Border layer © 2018–2022 GADM.

Table 3 presents summary statistics of FFS for the four scenarios. Upon comparing the values across all scenarios, it is evident that the 2016 scenario exhibits the lowest minimum value among the four. Conversely, the 2022 scenario demonstrates higher maximum and mean FFS values, suggesting a greater susceptibility compared to 2016. Notably, the mean susceptibility value for 2022 (0.419) is the highest among the four scenarios, indicating the highest mean FFS. The future scenario excluding projected land cover data shows the highest maximum value and only a slightly lower mean value (0.414) than the June 2022 scenario. Finally the future scenario including land cover data (*) shows the lowest maximum, mean, and standard deviation FFS values compared to the other scenarios.

To assess variabilities in FFS on a local scale, a detailed zoom to an area in the west of Brandenburg is shown in Fig. 6. The four maps show the municipality of Medewitz in the west of Brandenburg. The 2016 scenario shows a fairly low FFS (Fig. 6a). The three other maps show FFS anomalies compared to 2016 (Fig. 6b–d). Whereas the 2022 scenario shows positive anomaly values of 10 % to 15 %, anomaly values are even higher in the future scenario excluding projected land cover data, reaching +20 %. In contrast, the scenario including land cover changes (scenario d) shows negative anomalies up to −15 %. However, pixels in the east and south of the map show positive FFS anomalies as well.

https://nhess.copernicus.org/articles/25/383/2025/nhess-25-383-2025-f06

Figure 6Detailed maps of FFS anomalies in the municipality of Medewitz (Brandenburg). The scenarios in (c) and (d) both show predicted FFS in June 2081–2100 under SSP5-8.5. The scenario in (d) includes projected land cover data, whereas the scenario in (c) does not. Base map © OpenStreetMap contributors 2024. Distributed under the Open Data Commons Open Database License (ODbL) v1.0. Border layer © 2018–2022 GADM.

The four zoomed-in maps in Fig. 7 depict the municipality of Crinitz located in the south of Brandenburg. Whereas the June 2022 scenario (scenario b) mainly shows anomalies close to 0, except for some pixels reaching up to +16 %, the future scenario relying only on climatic projections (scenario c) shows substantial negative anomalies reaching up to −20 %. Similarly, the scenario including projected land cover data (scenario d) shows a substantial proportion of pixels with negative FFS anomalies. However, some areas in the north and south-west of the city show positive FFS anomalies.

https://nhess.copernicus.org/articles/25/383/2025/nhess-25-383-2025-f07

Figure 7Detailed maps of FFS anomalies in the municipality of Crinitz (Brandenburg). The scenarios in (c) and (d) both show predicted FFS in June 2081–2100 under SSP5-8.5. The scenario in (d) includes projected land cover data, whereas the scenario in (c) does not. Base map © OpenStreetMap contributors 2024. Distributed under the Open Data Commons Open Database License (ODbL) v1.0. Border layer © 2018–2022 GADM.

Figures 6 and 7 show that despite the trend of overall increase in FFS between 2016 and the 2081–2100 future scenario excluding projected land cover data (Figs. 4 and 5), FFS differs significantly across the federal state. Furthermore, the future scenario incorporating land cover changes shows substantial differences to the scenario only relying on climatic projections.

Table 3Statistical overview of the four forest fire susceptibility scenarios. The 2081–2100 and 2081–2100* scenarios both show predicted FFS in June 2081–2100 under SSP5-8.5. The 2081–2100* scenario includes projected land cover data, whereas the 2081–2100 scenario does not.

Download Print Version | Download XLSX

4 Discussion

4.1 The drivers of forest fire susceptibility

Overall, the climatic variables did not have a significant influence on the model performance. In contrast, the anthropogenic, LULC, and vegetation predictors showed higher importance. The results reflect the fact that climatic parameters do not appear to play a pivotal role regarding FFS (see Fig. S6). The reason for this finding may be the extent of the study area, as meteorological conditions do not show high spatial variation within Brandenburg. Meteorological conditions may be more important when analysing FFS on a national or international scale (Busico et al.2019; He et al.2022; Li et al.2024). According to the Lower Forestry Authority of the State of Brandenburg (2023), a high number of fires were caused by intentional arson and other anthropogenic actions such as open fires or smoking (see Table S2). Therefore, climatic conditions may not have contributed to the emergence of those fires in a significant way. Furthermore, meteorological projections assume that air temperatures will increase overall. However, the input data used for this study show increased precipitation patterns in Brandenburg in the future scenarios compared to the periods of June 2016 and June 2022 as well (see Figs. S1 and S2). Consequently, this change in precipitation patterns shown by the input data may have lowered future FFS in the study region, thus outweighing the effect of higher air temperatures and contributing to the lower mean FFS in future scenarios compared to the extremely hot and dry year of 2022. The Deutscher Wetterdienst (DWD) (DWD2019) predicts changes between −4 % to +13 % in the annual precipitation sums until the end of the 21st century, illustrating the uncertainty in future precipitation predictions. As a result, in the case of a decrease in precipitation before the end of the 21st century, this will strongly affect the flammability of Brandenburg's forests and thus the future FFS.

Extreme weather events may be a better indicator of future FFS rather than averaged long-term meteorological trends. Extreme weather conditions such as the dry conditions in 2022 were efficiently captured by the current meteorological data, whereas the multi-annually aggregated monthly projected meteorological data (WorldClim) did not reflect these extreme weather events. For instance, the monthly average precipitation sum in Brandenburg shows flatter curves for the future precipitation, whereas more intense changes in mean precipitation values can be seen in 2016 and 2022 (see Fig. S2). For example, the precipitation curve for 2022 shows a substantial drop in March, reflecting a very dry month with low precipitation that may have driven the higher FFS mean value in 2022 compared to other scenarios. Hence, future FFS might turn out to be higher in reality, given the expected increase in extreme weather events that will enhance the likelihood of drought conditions (Rad et al.2021; Silva et al.2018; Wu et al.2021). To assess the future development of FFS on a local scale, climatic data with a higher temporal resolution are needed to reflect weather extremes more adequately than multi-annually aggregated climate data.

The moderate to low influence of topographic predictors in predicting FFS is most likely due to the rather homogeneous topography in Brandenburg. For vegetation parameters, the percentage of broadleaf forest was most important for the modelling. This result aligns with several studies that have shown monocultural coniferous forests being more sensitive to forest fires (Afreen et al.2011; Bauhus et al.2017; Gnilke et al.2022). Being dominated by pine trees makes Brandenburg particularly susceptible to forest fires. For example, Gnilke et al. (2022) assessed the fire damage in pine forests in Brandenburg, concluding that pure pine stands showed the most burning marks, whereas mixed tree stands were more resilient to forest fires. Furthermore, Buras et al. (2018) have underlined the vulnerability of pine trees located at forest edges, similarly to our results about the influence of the distance to the forest edge (mean distance for fire points of 148.5 m and mean distance for non-fire points of 174.8 m; also see Table S6). Thus, forest edges in Brandenburg may require special protection to avoid future forest fires.

On a regional scale, anthropogenic parameters appear to be more relevant to FFS. In particular, the distance to urban settlements and railways showed a high significance for modelling FFS in Brandenburg. This confirms the statistics of forest fire emergence in Brandenburg provided by the Lower Forestry Authority of the State of Brandenburg (2023) (see Table S2) highlighting that most forest fires in Brandenburg emerge from human negligence or malicious arson. Several other studies have reached the same conclusion (Busico et al.2019; Cilli et al.2022; Ghorbanzadeh et al.2019; Gnilke and Sanders2021; He et al.2022; Ruffault and Mouillot2017). However, the distance to military sites only moderately influenced the RF models (see Fig. 3). Furthermore, the Wilcoxon test (see Table S6) was not significant, underlining that there was no clear difference in the distribution of fire and non-fire points across Brandenburg. Therefore, the data and model results do not show a clear relationship between the distance to military sites and FFS.

4.2 Assessing current and future forest fire susceptibility

Overall, the 2081–2100 future scenario (excluding projected land cover data) revealed a substantial increase in mean FFS compared to 2016. However, in 2022 the mean FFS was higher than in 2016 and the two future scenarios. The comparatively high mean FFS of 2022 can be explained by significantly drier and hotter conditions compared to 2016. Nevertheless, the mean FFS value of the future scenario neglecting land cover changes is only slightly below the mean FFS value of 2022 and higher than the mean FFS value of 2016. Considering exclusively future climatic conditions, this indicates an expected overall increase in FFS in Brandenburg until the end of the 21st century compared to June 2016. However, since the future modelled climate data rely on multi-annual monthly averages of air temperature and precipitation, future FFS is possibly underestimated in this study.

The second future scenario including both projected land cover changes (*) and future climatic conditions paints a different picture. As shown in Table 3, mean FFS was the lowest of all scenarios, indicating an overall decrease in FFS. This result can most likely be explained by two aspects: first, Esri's Land Cover 2050 – Global dataset (Esri Environment2021) used to plot the future distance to urban settlements projects a decrease in urbanised areas in the future compared to the Impervious Built-up dataset (EEA2020b). Shrinking urban areas can be explained by demographic changes, such as the ageing and decline of the German population, especially in the east of Germany (Kroll and Haase2010). Although Kroll and Haase (2010) state that the ageing of the German population has not yet influenced land use changes, they argue that this is likely to change in the future. Second, Esri's Land Cover 2050 – Global dataset (Esri Environment2021) has a lower spatial resolution (300 m) than the Copernicus Impervious dataset (EEA2020b) used to map the distance to “current” urban settlements (10 m). As a result, Esri's dataset may show some inaccuracies due to mixed pixel effects. For instance, some smaller settlements may not appear in the future land cover dataset. Our results underscore how the inclusion of projected land cover data significantly changes the projected FFS in the future, an aspect that can be further explored in future studies with new land cover projections.

Based on our findings, it can be argued that future urban development trends will significantly influence FFS. Hence, population decline and the abandonment of villages and rural areas may decrease FFS in those areas. However, new settlements due to continuous suburbanisation processes may require additional forest fire prevention efforts in the future. Regardless of these trends, the expected increase in drought events in Brandenburg (Gnilke et al.2022) may intensify FFS in Brandenburg in the future. Consequently, effective forest fire management strategies in Brandenburg need to address these aspects. Therefore, the following section provides key strategies for the management of forest fires in the future.

4.3 Strategies for forest fire management in Brandenburg

Forest fire management strategies include the improvement of forest fire prediction, prevention, detection, extinction, the constant monitoring of meteorological conditions, and the assessment of previous forest fires to improve management strategies (Martell2007). An effective forest fire prevention strategy in Brandenburg involves promoting the growth of mixed forests instead of the prevalent monocultural pine forests. In particular, increasing the percentage of broadleaf trees is needed (Ministry for Rural Development, Environment and Agriculture in Brandenburg2024; Gnilke et al.2022). Protection measures should put particular emphasis on forest edges and forests in proximity to any type of anthropogenic infrastructure. The prediction of FFS as implemented here provides a helpful tool to identifying the most susceptible forest areas in Brandenburg, where the implementation of forest fire management strategies is most important. Complementing the constant monitoring of meteorological conditions, it can provide a powerful means to predict FFS and to provide an early warning system for forest fires. In addition to that, constantly updated meteorological data, as well as drought indices and the forest fire danger index provided by the Deutscher Wetterdienst, are essential to predicting FFS in Brandenburg (Fekete and Nehren2023).

The conventional approach to fire detection involves integrating public reports with observation towers and aerial patrols (Martell2007). Increasing the number of observation towers in forest areas with high FFS could speed up fire detection and extinguishment. A valuable forest fire prevention measure is the restriction of human activities in forests or the closure of forests to the public in accordance with meteorological conditions, given the large anthropogenic contribution to FFS. This is recommendable especially in forest areas with high FFS to decrease the number of fires caused by anthropogenic influences. However, the meaning of forests for recreational purposes, as well as the economic factor of touristic forest users, should be considered before implementing such measures. Additionally, implementing public education initiatives regarding forest fires through school programmes and media campaigns is imperative for fostering greater awareness of forest fires and modifying behaviours to reduce ignition risks (Martell2007).

Moreover, the implementation of fire breaks is recommendable to limit the spread of forest fires (Berčák et al.2023). Another strategy can be the thinning of pine forests to reduce fire risk. For example, Crecente-Campo et al. (2009) have concluded that the thinning of Pinus sylvestris can contribute to the growth of a mixed-leaf forest that has shown to be more resilient to forest fires (Afreen et al.2011; Bauhus et al.2017; Gnilke et al.2022). Finally, it is crucial to employ interregional forest fire management strategies, since forest fires, such as the fire in Bohemian Switzerland National Park in 2022, may spread from neighbouring countries to Germany or vice versa (Boháč and Drápela2023). Considering the high FFS in the south-east of the federal state, forest fire management authorities in Brandenburg should consider closer cooperation with the neighbouring country of Poland to develop and implement joint management strategies.

4.4 Shortcomings and future perspectives

Analysing FFS on a local scale ideally requires climatic data at both high spatial and temporal resolution. High-temporal-resolution meteorological data better reflect extreme weather events such as droughts. Consequently, the availability of climatic data at both high spatial and temporal resolution may significantly enhance the quality of future FFS assessments. Ideally, future FFS analysis should incorporate projected climate data with a monthly temporal resolution to reflect future drought events more effectively. Similarly, forest fire products based on remote sensing data with high spatial and temporal resolution would strongly improve forest fire assessments on smaller scales. However, this data type is not available yet, and its development is limited by the fact that current satellites used for meteorological observations are not able to create images both at high spatial and temporal resolution due to technical restrictions (Kussul et al.2023). Forest fire data providers such as the European Forest Fire Information System (EFFIS) supply frequently updated representations of burnt areas in Europe, the Middle East, and North Africa, which is helpful for forest fire analysis on national or international scales. However, the EFFIS burnt-area product is based on the 250 m spatial resolution of the MODIS optical scanner, resulting in smaller forest fires not being included (Achour et al.2022). Thus, this product is not appropriate for the assessment of FFS at smaller scales.

In a similar vein, an analysis of forest fire detection systems by Barmpoutis et al. (2020) underlines the limitations of satellites in providing both high temporal and spatial resolution. Although satellites such as MODIS or Landsat have thermal infrared bands that can serve for active fire detection, those satellites have their limitations. MODIS has a high temporal resolution but a spatial resolution of only 1 km for the thermal infrared bands. Landsat satellites, on the other hand, provide higher-spatial-resolution data (e.g. 100 m for the thermal infrared band for Landsat 8 and 9) but are limited to a temporal resolution of 16 d (Acharya and Yang2015; Chanthiya and Kalaivani2021; Fu et al.2020). However, new developments of real-time detection and live tracking of wildfires based on a set of over 20 satellites such as that provided by OroraTech (OroraTech2021) show the potential of future analysis of forest fires.

Nevertheless, it is crucial that local forest fire management institutions provide data on smaller fires as well. However, in the case of the Lower Forestry Authority of the State of Brandenburg, forest fire data were not provided in the form of polygons of burnt areas but in the form of fire ignition points. Despite the fact that the burnt area (ha) was provided, the exact extent of it could only be assumed. Consequently, model results of FFS prediction might have been more accurate if the actual extent of the forest fires had been available. Nevertheless, with continuous advances in remote sensing, forest fire data may be openly available at higher spatial resolutions in the future, which represents a significant potential for future FFS predictions on a local scale.

Apart from the spatial resolution of forest fire products, the modelling approach to predicting FFS should be carefully selected. As previously discussed, meteorological parameters did not have a significant influence on the model. Therefore, future research may consider applying a long short-term memory (LSTM) model to better incorporate meteorological trends and to improve the understanding of how forests react to droughts and heat waves (Burge et al.2021; Natekar et al.2021).

Furthermore, the future land cover change dataset (Esri Environment2021) had some limitations. First, it only included information on “Artificial Surface or Urban Area”. Consequently, a differentiation of different anthropogenic land uses (e.g. campsites, streets, urban settlements, or railways) for the future scenarios was not possible. Instead, the dataset was only used to project the future distance to urban settlements. Second, the projection of the dataset was only available for 2050. Ideally, a dataset reflecting the land use changes until the end of the 21st century would have led to more accurate results. Third, compared to the other land use and land cover datasets used in this study, the spatial resolution of the future land cover change dataset (Esri Environment2021) was relatively coarse. Therefore, the dataset may contain some inaccuracies, thus potentially decreasing the accuracy of the future FFS projections. Nevertheless, to our knowledge, this dataset had a relatively high spatial resolution compared to other datasets, which is why it was selected for the study. In the end, the expansion of renewable energy (Hilker et al.2024), the settlement of new companies and factories (e.g. Tesla Gigafactory in Grünheide) (Kühn2023), suburbanisation processes around Berlin driven by rising housing prices (Leibert et al.2022), and finally the abandonment of smaller villages due to ageing and population decline are likely to lead to future land cover changes and either heightened or decreased pressures on forests. Consequently, including this dataset in the analysis provides valuable information on potential land cover changes. Future research may consider including higher-spatial-resolution land cover change data to model FFS.

Finally, future FFS research may integrate further predictors, dynamic predictors in particular, into their analysis. Following Rad et al. (2021), key variables shaping drought conditions are precipitation, soil moisture, and streamflow. Thus, it may be beneficial to include soil moisture data in particular in future analyses. However, due to a lack of soil moisture projections, this parameter was not integrated into this study.

5 Conclusions

This study successfully predicted FFS on a regional scale in the federal state of Brandenburg under different scenarios with the RF ML algorithm. The FFS maps show a high FFS in the south and south-east of the federal state. Considering only future meteorological conditions, future FFS is expected to increase compared to the 2016 reference scenario. Extreme events such as droughts can significantly intensify FFS, which was demonstrated by the higher mean FFS value of 2022 compared to the other scenarios. However, including both projected land cover change and future meteorological data into the future projections showed a decrease in FFS. This trend might be driven by demographic changes ultimately leading to future land use changes.

The selection of a 3-month temporal aggregation of the meteorological datasets was appropriate to reflect long-term meteorological trends. Using climate data at a higher temporal resolution would have shown the effect of extreme weather events more adequately. Therefore, future research could aim at integrating climate data at higher temporal resolution (e.g. weekly) to integrate the effect of extreme weather events into the predictions.

Our study emphasised the importance of anthropogenic predictors such as the distance to urban settlements, railways, or campsites. Thus, it is crucial to protect forests from anthropogenic influences to reduce the occurrence of forest fires, especially during drought events. Furthermore, we showed the higher resilience of mixed forests in contrast to monocultural forests, confirming previous literature. Forest managers should therefore prioritise the growth of broadleaf trees. Soil parameters such as the percentage of silt and sand had medium to high importance, suggesting that pore sizes and the consequent capacity of the soil to carry and maintain water restrict the availability of water for trees. Finally, topographic parameters such as slope or TWI had rather low importance for predicting FFS in Brandenburg, which is likely due to the overall rather flat topography of the federal state.

This study and FFS maps can serve local forest managers and firefighters in the prevention of forest fires in the region. Furthermore, the identification of contributing variables can support the development of forest fire management strategies adapted to local circumstances.

Code availability

The code used for this study, as well as the forest fire susceptibility maps (Figs. 4–7), are publicly available on Zenodo at https://doi.org/10.5281/zenodo.14214917 (Horn2024) and at https://doi.org/10.5281/zenodo.14710876 (Horn2025).

Data availability

The forest fire data were provided by the Lower Forestry Authority of the State of Brandenburg and can be acquired upon request at the Eberswalde Forestry Competence Centre (EFCC).

Supplement

The supplement related to this article is available online at: https://doi.org/10.5194/nhess-25-383-2025-supplement.

Author contributions

KHH: writing (original draft), visualisation, validation, software, methodology, investigation, formal analysis, data curation. SV: writing (review and editing), supervision, methodology, conceptualisation. HL: writing (review and editing), methodology, conceptualisation. BK: writing (review and editing), supervision, resources, project administration, funding acquisition, conceptualisation.

Competing interests

The contact author has declared that none of the authors has any competing interests.

Disclaimer

Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors.

Special issue statement

This article is part of the special issue “Current and future water-related risks in the Berlin–Brandenburg region”. It is not associated with a conference.

Acknowledgements

We would like to acknowledge the support of the Einstein Foundation Berlin and Berlin University Alliance. We would also like to thank Alby Duarte Rocha, Chunyan Xu, Christian Schulz, and Pedro Alencar for their support on technical and conceptual matters. For the scope of this research, ChatGPT 3.5 was used to support coding, writing, and editing of an earlier version of this article.

Financial support

This research was funded through the Einstein Research Unit Climate and Water under Change supported by the Einstein Foundation Berlin and the Berlin University Alliance (grant no. ERU-2020-609). Further support came from the Open Access Publication Fund of TU Berlin.

Review statement

This paper was edited by Axel Bronstert and reviewed by two anonymous referees.

References

Abdollahi, A. and Pradhan, B.: Explainable artificial intelligence (XAI) for interpreting the contributing factors feed into the wildfire susceptibility prediction model, Sci. Total Environ., 879, 163004, https://doi.org/10.1016/j.scitotenv.2023.163004, 2023. a, b, c, d, e

Acharya, T. D. and Yang, I.: Exploring Landsat 8, International Journal of IT, Engineering and Applied Sciences Research (IJIEASR), 4, 4–10, 2015. a

Achour, H., Toujani, A., Trabelsi, H., and Jaouadi, W.: Evaluation and comparison of Sentinel-2 MSI, Landsat 8 OLI, and EFFIS data for forest fires mapping. Illustrations from the summer 2017 fires in Tunisia, Geocarto Int., 37, 7021–7040, https://doi.org/10.1080/10106049.2021.1980118, 2022. a

Afreen, S., Sharma, N., Chaturvedi, R. K., Gopalakrishnan, R., and Ravindranath, N. H.: Forest policies and programs affecting vulnerability and adaptation to climate change, Mitig. Adapt. Strat. Gl., 16, 177–197, https://doi.org/10.1007/s11027-010-9259-5, 2011. a, b, c

Ambadan, J. T., Oja, M., Gedalof, Z., and Berg, A. A.: Satellite-Observed Soil Moisture as an Indicator of Wildfire Risk, Remote Sens.-Basel, 12, 1543, https://doi.org/10.3390/rs12101543, 2020. a, b

Amelung, W., Blume, H.-P., Fleige, H., Horn, R., Kandeler, E., Kögel-Knabner, I., Kretzschmar, R., Stahr, K., and Wilke, B.-M.: Scheffer/Schachtschabel Lehrbuch der Bodenkunde, 17th edn., Springer eBook Collection, Springer Spektrum, Berlin, Heidelberg, https://doi.org/10.1007/978-3-662-55871-3, 2018. a

Badía-Villas, D., González-Pérez, J. A., Aznar, J. M., Arjona-Gracia, B., and Martí-Dalmau, C.: Changes in water repellency, aggregation and organic matter of a mollic horizon burned in laboratory: Soil depth affected by fire, Geoderma, 213, 400–407, https://doi.org/10.1016/j.geoderma.2013.08.038, 2014. a

Barmpoutis, P., Papaioannou, P., Dimitropoulos, K., and Grammalidis, N.: A Review on Early Forest Fire Detection Systems Using Optical Remote Sensing, Sensors, 20, 6442, https://doi.org/10.3390/s20226442, 2020. a

Bauhus, J., Forrester, D. I., Gardiner, B., Jactel, H., Vallejo, R., and Pretzsch, H.: Ecological Stability of Mixed-Species Forests, in: MIXED-species forests: Ecology and management, edited by: Pretzsch, H., Forrester, D. I., and Bauhus, J., Springer-Verlag, Berlin, 337–382, https://doi.org/10.1007/978-3-662-54553-9_7, 2017. a, b, c, d

Berčák, R., Holuša, J., Kaczmarowski, J., Tyburski, L., Szczygieł, R., Held, A., Vacik, H., Slivinský, J., and Chromek, I.: Fire Protection Principles and Recommendations in Disturbed Forest Areas in Central Europe: A Review, Fire, 6, 310, https://doi.org/10.3390/fire6080310, 2023. a

Boháč, A. and Drápela, E.: Present Climate Change as a Threat to Geoheritage: The Wildfire in Bohemian Switzerland National Park and Its Use in Place-Based Learning, Geosciences, 13, 383, https://doi.org/10.3390/geosciences13120383, 2023. a

Bradley, A. P.: The use of the area under the ROC curve in the evaluation of machine learning algorithms, Pattern Recogn., 30, 1145–1159, https://doi.org/10.1016/S0031-3203(96)00142-2, 1997. a

Breiman, L.: Using adaptive bagging to debias regressions, Tech. rep., University of California at Berkeley, https://www.stat.berkeley.edu/users/breiman/adaptbag99.pdf (last access: 10 August 2023), 1999.  a

Breiman, L.: Random Forests, Mach. Learn., 45, 5–32, https://doi.org/10.1023/A:1010933404324, 2001. a

Buras, A., Schunk, C., Zeiträg, C., Herrmann, C., Kaiser, L., Lemme, H., Straub, C., Taeger, S., Gößwein, S., Klemmt, H.-J., and Menzel, A.: Are Scots pine forest edges particularly prone to drought-induced mortality?, Environ. Res. Lett., 13, 025001, https://doi.org/10.1088/1748-9326/aaa0b4, 2018. a, b, c

Burge, J., Bonanni, M., Ihme, M., and Hu, L.: Convolutional LSTM Neural Networks for Modeling Wildland Fire Dynamics, arXiv [preprint], https://doi.org/10.48550/arXiv.2012.06679, 2021. a

Busico, G., Giuditta, E., Kazakis, N., and Colombani, N.: A Hybrid GIS and AHP Approach for Modelling Actual and Future Forest Fire Risk Under Climate Change Accounting Water Resources Attenuation Role, Sustainability, 11, 7166, https://doi.org/10.3390/su11247166, 2019. a, b, c, d, e, f, g, h, i, j

Chanthiya, P. and Kalaivani, V.: Forest fire detection on LANDSAT images using support vector machine, Concurr. Comp.-Pract. E., 33, e6280, https://doi.org/10.1002/cpe.6280, 2021. a

Chicas, S. D. and Østergaard Nielsen, J.: Who are the actors and what are the factors that are used in models to map forest fire susceptibility? A systematic review, Nat. Hazards, 114, 2417–2434, https://doi.org/10.1007/s11069-022-05495-5, 2022. a, b, c

Cilli, R., Elia, M., D'Este, M., Giannico, V., Amoroso, N., Lombardi, A., Pantaleo, E., Monaco, A., Sanesi, G., Tangaro, S., Bellotti, R., and Lafortezza, R.: Explainable artificial intelligence (XAI) detects wildfire occurrence in the Mediterranean countries of Southern Europe, Sci. Rep.-UK, 12, 16349, https://doi.org/10.1038/s41598-022-20347-9, 2022. a, b, c

Crecente-Campo, F., Pommerening, A., and Rodríguez-Soalleiro, R.: Impacts of thinning on structure, growth and risk of crown fire in a Pinus sylvestris L. plantation in northern Spain, Forest Ecol. Manag., 257, 1945–1954, https://doi.org/10.1016/j.foreco.2009.02.009, 2009. a

DWD: Klimareport Brandenburg. Fakten bis zur Gegenwart – Erwartungen für die Zukunft, Report 1, Deutscher Wetterdienst, Offenbach am Main, Deutschland, https://www.dwd.de/DE/leistungen/klimareport_bb/klimareport_bb_2019_download.pdf;jsessionid=5BB70292853A51CB031B8D1E338E7F56.live21071?__blob=publicationFile&v=5 (last access: 20 December 2023), 2019. a

DWD Climate Data Center (CDC): Grids of monthly averaged daily air temperature (2 m) over Germany, version v1.0., Tech. rep., Deutscher Wetterdienst, https://opendata.dwd.de/climate_environment/CDC/grids_germany/monthl/air_temperature_mean/ (last access: 14 March 2024), 2023a. a

DWD Climate Data Center (CDC): Grids of monthly total precipitation over Germany, version v1.0., Tech. rep., Deutscher Wetterdienst, https://opendata.dwd.de/climate_environment/CDC/grids_germany/monthly/precipitation/ (last access: 14 March 2024), 2023b. a

EEA – European Environment Agency: Forest Type 2018 (raster 10 m), Europe, 3 yearly, October 2020, Tech. rep., European Environment Agency, https://land.copernicus.eu/en/products/high-resolution-layer-dominant-leaf-type (last access: 30 May 2023), 2020a. 

EEA – European Environment Agency: Impervious Built-up 2018 (raster 10 m), Europe, 3 yearly, August 2020, Tech. rep., European Environment Agency, https://land.copernicus.eu/en/products/high-resolution-layer-imperviousness/impervious-built-up-2018 (last access: 27 October 2023), 2020b. a, b, c, d

EEA – European Environment Agency: Tree Cover Density 2018 (raster 10 m), Europe, 3 yearly, September 2020, Tech. rep., European Environment Agency, https://land.copernicus.eu/en/products/high-resolution-layer-tree-cover-density (last access: 24 May 2023), 2020c. 

El Garroussi, S., Di Giuseppe, F., Barnard, C., and Wetterhall, F.: Europe faces up to tenfold increase in extreme fires in a warming climate, npj Climate and Atmospheric Science, 7, 30, https://doi.org/10.1038/s41612-024-00575-8, 2024. a

Eslami, R., Azarnoush, M., Kialashki, A., and Kazemzadeh, F.: GIS-Based Forest Fire Susceptbility Assessment By Random Forest, Artificial Neural Network And Logistic Regression Methods, J. Trop. For. Sci., 33, 173–184, 2021. a

Esri Environment: Land Cover 2050 – Global, Tech. rep., Esri Environment, https://hub.arcgis.com/datasets/esri::land-cover-2050-global/about (last access: 2 September 2024), 2021. a, b, c, d, e, f

Federal Office for Agriculture and Food: Waldbrandstatistik der Bundesrepublik Deutschland für das Jahr 2022, Tech. rep., Federal Office for Agriculture and Food, Bonn, https://www.bmel-statistik.de/fileadmin/daten/0302250-2022.pdf (last access: 13 June 2023), 2023. a

Fekete, A. and Nehren, U.: Assessment of social vulnerability to forest fire and hazardous facilities in Germany, Int. J. Disaster Risk Re., 87, 103562, https://doi.org/10.1016/j.ijdrr.2023.103562, 2023. a

Feng, L., Lysakowski, B., Eisenschmidt, J., and Birkhofer, K.: The impact of wildfire and mammal carcasses on beetle emergence from heathland soils, Ecol. Entomol., 47, 883–894, https://doi.org/10.1111/een.13179, 2022. a

Fick, S. E. and Hijmans, R. J.: WorldClim 2.1 (CMIP6) – 30 s resolution climate projections: WorldClim, https://www.worldclim.org/data/cmip6/cmip6_clim30s.html (last access: 22 December 2023), 2022. a, b

Fu, Y., Li, R., Wang, X., Bergeron, Y., Valeria, O., Chavardès, R. D., Wang, Y., and Hu, J.: Fire Detection and Fire Radiative Power in Forests and Low-Biomass Lands in Northeast Asia: MODIS versus VIIRS Fire Products, Remote Sens.-Basel, 12, 2870, https://doi.org/10.3390/rs12182870, 2020. a

Ghorbanzadeh, O., Blaschke, T., Gholamnia, K., and Aryal, J.: Forest Fire Susceptibility and Risk Mapping Using Social/Infrastructural Vulnerability and Environmental Variables, Fire, 2, 50, https://doi.org/10.3390/fire2030050, 2019. a, b, c, d, e, f, g

Gislason, P. O., Benediktsson, J. A., and Sveinsson, J. R.: Random Forests for land cover classification, Pattern Recogn. Lett., 27, 294–300, https://doi.org/10.1016/j.patrec.2005.08.011, 2006. a

Gnilke, A. and Sanders, T.: Forest fire history in Germany (2001–2020), Tech. rep., Thünen-Institut of Forest Ecosystems, https://literatur.thuenen.de/digbib_extern/dn064174.pdf (last access: 30 January 2024), 2021. a, b, c, d, e

Gnilke, A., Liesegang, J., and Sanders, T.: Potential forest fire prevention by management-An analysis of fire damage in pine forests, Tech. rep., Thünen-Institute of Forest Ecosystems, https://literatur.thuenen.de/digbib_extern/dn065237.pdf (last access: 21 February 2024), 2022. a, b, c, d, e, f

Guo, H., Bao, A., Liu, T., Jiapaer, G., Ndayisaba, F., Jiang, L., Kurban, A., and De Maeyer, P.: Spatial and temporal characteristics of droughts in Central Asia during 1966–2015, Sci. Total Environ., 624, 1523–1538, https://doi.org/10.1016/j.scitotenv.2017.12.120, 2018. a

He, W., Shirowzhan, S., and Pettit, C. J.: GIS and Machine Learning for Analysing Influencing Factors of Bushfires Using 40-Year Spatio-Temporal Bushfire Data, ISPRS Int. J. Geo-Inf., 11, 336, https://doi.org/10.3390/ijgi11060336, 2022. a, b, c, d, e, f, g, h, i, j, k, l, m, n, o, p, q

Hilker, J. M., Busse, M., Müller, K., and Zscheischler, J.: Photovoltaics in agricultural landscapes: “Industrial land use” or a “real compromise” between renewable energy and biodiversity? Perspectives of German nature conservation associations, Energy, Sustainability and Society, 14, 6, https://doi.org/10.1186/s13705-023-00431-2, 2024. a

Holsten, A., Vetter, T., Vohland, K., and Krysanova, V.: Impact of climate change on soil moisture dynamics in Brandenburg with a focus on nature conservation areas, Ecol. Model., 220, 2076–2087, https://doi.org/10.1016/j.ecolmodel.2009.04.038, 2009. a

Horn, K. H.: Forest Fire Susceptibility Modelling in north-east Germany, Zenodo [code], https://doi.org/10.5281/zenodo.14214917, 2024. a

Horn, K. H.: Forest fire susceptibility modelling in north-eastern Germany, Zenodo [code and data set], https://doi.org/10.5281/zenodo.14710876, 2025. a

Kemter, M., Fischer, M., Luna, L. V., Schönfeldt, E., Vogel, J., Banerjee, A., Korup, O., and Thonicke, K.: Cascading Hazards in the Aftermath of Australia's 2019/2020 Black Summer Wildfires, Earths Future, 9, e2020EF001884, https://doi.org/10.1029/2020EF001884, 2021. a

Kroll, F. and Haase, D.: Does demographic change affect land use patterns?: A case study from Germany, Land Use Policy, 27, 726–737, https://doi.org/10.1016/j.landusepol.2009.10.001, 2010. a, b

Kühn, M.: Planungskonflikte und Partizipation: die Gigafactory Tesla, Raumforschung und Raumordnung | Spatial Research and Planning, 81, 538–556, https://doi.org/10.14512/rur.1698, 2023. a

Kuhn, M. and Johnson, K.: Applied Predictive Modeling, Springer New York, New York, NY, https://doi.org/10.1007/978-1-4614-6849-3, 2013. a

Kussul, N., Fedorov, O., Yailymov, B., Pidgorodetska, L., Kolos, L., Yailymova, H., and Shelestov, A.: Fire Danger Assessment Using Moderate-Spatial Resolution Satellite Data, Fire, 6, 72, https://doi.org/10.3390/fire6020072, 2023. a

Lang, N., Jetz, W., Schindler, K., and Wegner, J. D.: A high-resolution canopy height model of the Earth, Nature Ecology & Evolution, 7, 1778–1789, https://doi.org/10.1038/s41559-023-02206-6, 2023. a

Leibert, T., Wolff, M., and Haase, A.: Shifting spatial patterns in German population trends: local-level hot and cold spots, 1990–2019, Geogr. Helv., 77, 369–387, https://doi.org/10.5194/gh-77-369-2022, 2022. a

LGB State Office for Land Surveying and Geoinformation Brandenburg: Geländemodell, https://geobasis-bb.de/lgb/de/geodaten/3d-produkte/gelaendemodell/ (last access: 26 May 2023), 2023. a, b, c

Li, H., Vulova, S., Rocha, A. D., and Kleinschmit, B.: Spatio-temporal feature attribution of European summer wildfires with Explainable Artificial Intelligence (XAI), Sci. Total Environ., 916, 170330, https://doi.org/10.1016/j.scitotenv.2024.170330, 2024. a, b

Littell, J. S., Peterson, D. L., Riley, K. L., Liu, Y., and Luce, C. H.: Effects of drought on forests and rangelands in the United States: A comprehensive science synthesis, Fire and Drought, 135–154, https://www.fs.usda.gov/research/treesearch/download/50971.pdf (last access: 23 February 2024), 2016. a

Lizundia-Loiola, J., Otón, G., Ramo, R., and Chuvieco, E.: A spatio-temporal active-fire clustering approach for global burned area mapping at 250 m from MODIS data, Remote Sens. Environ., 236, 111493, https://doi.org/10.1016/j.rse.2019.111493, 2020. a

Lower Forestry Authority of the State of Brandenburg: Waldbranddaten der Jahre 2010 bis 2022 [data], 2023. a, b, c, d, e, f, g, h, i, j, k

Maingi, J. K. and Henry, M. C.: Factors influencing wildfire occurrence and distribution in eastern Kentucky, USA, Int. J. Wildland Fire, 16, 23, https://doi.org/10.1071/WF06007, 2007. a, b

Mallik, A. U., Gimingham, C. H., and Rahman, A. A.: Ecological Effects of Heather Burning: I. Water Infiltration, Moisture Retention and Porosity of Surface Soil, J. Ecol., 72, 767, https://doi.org/10.2307/2259530, 1984. a

Martell, D. L.: Forest Fire Management, in: Handbook of Operations Research in Natural Resources, edited by: Weintraub, A., International Series in Operations Research & Management Science, Springer Science+Business Media, LLC, Boston, MA, 489–509, https://doi.org/10.1007/978-0-387-71815-6_26, 2007. a, b, c

Matos, E. S., Freese, D., Śla̧zak, A., Bachmann, U., Veste, M., and Hüttl, R. F.: Organic-carbon and nitrogen stocks and organic-carbon fractions in soil under mixed pine and oak forest stands of different ages in NE Germany, J. Plant Nutr. Soil Sc., 173, 654–661, https://doi.org/10.1002/jpln.200900046, 2010. a, b

Meinshausen, M., Nicholls, Z. R. J., Lewis, J., Gidden, M. J., Vogel, E., Freund, M., Beyerle, U., Gessner, C., Nauels, A., Bauer, N., Canadell, J. G., Daniel, J. S., John, A., Krummel, P. B., Luderer, G., Meinshausen, N., Montzka, S. A., Rayner, P. J., Reimann, S., Smith, S. J., van den Berg, M., Velders, G. J. M., Vollmer, M. K., and Wang, R. H. J.: The shared socio-economic pathway (SSP) greenhouse gas concentrations and their extensions to 2500, Geosci. Model Dev., 13, 3571–3605, https://doi.org/10.5194/gmd-13-3571-2020, 2020. a

Milanović, S., Marković, N., Pamučar, D., Gigović, L., Kostić, P., and Milanović, S. D.: Forest Fire Probability Mapping in Eastern Serbia: Logistic Regression versus Random Forest Method, Forests, 12, 5, https://doi.org/10.3390/f12010005, 2021. a, b

Ministry for Rural Development, Environment and Agriculture in Brandenburg: Waldzustandsbericht 2023 des Landes Brandenburg, Tech. rep., Ministry for Rural Development, Environment and Agriculture in Brandenburg, Potsdam, https://forst.brandenburg.de/sixcms/media.php/9/wzb23.pdf (last access: 13 March 2024), 2023. a

Ministry for Rural Development, Environment and Agriculture in Brandenburg: Strategie des Landes Brandenburg zur Anpassung an die Folgen des Klimawandels, Tech. rep., Ministry for Rural Development, Environment and Agriculture in Brandenburg, Potsdam, https://mluk.brandenburg.de/sixcms/media.php/9/Klimaanpassungsstrategie-BB-Kurzfassung.pdf (last access: 3 July 2024), 2024. a

Natekar, S., Patil, S., Nair, A., and Roychowdhury, S.: Forest Fire Prediction using LSTM, in: 2021 2nd International Conference for Emerging Technology (INCET), 21–23 May 2021, Belagavi, India, 1–5, https://doi.org/10.1109/INCET51464.2021.9456113, 2021. a

Nguyen, Q. H., Ly, H.-B., Ho, L. S., Al-Ansari, N., van Le, H., van Tran, Q., Prakash, I., and Pham, B. T.: Influence of Data Splitting on Performance of Machine Learning Models in Prediction of Shear Strength of Soil, Math. Probl. Eng., 2021, 1–15, https://doi.org/10.1155/2021/4832864, 2021. a

Oliveira, S., Oehler, F., San-Miguel-Ayanz, J., Camia, A., and Pereira, J. M.: Modeling spatial patterns of fire occurrence in Mediterranean Europe using Multiple Regression and Random Forest, Forest Ecol. Manag., 275, 117–129, https://doi.org/10.1016/j.foreco.2012.03.003, 2012. a

Oliveira, T. M., Barros, A. M. G., Ager, A. A., and Fernandes, P. M.: Assessing the effect of a fuel break network to reduce burnt area and wildfire risk transmission, Int. J. Wildland Fire, 25, 619, https://doi.org/10.1071/WF15146, 2016. a, b

OpenStreetMap Contributors: OpenStreetMap, https://www.openstreetmap.org (last access: 14 March 2024), 2023. a, b

OroraTech: Wildfire Solution | OroraTech, https://ororatech.com/wildfire-solution/ (last access: 6 May 2024), 2021. a

Oyonarte, C., Escoriza, I., Delgado, R., Pinto, V., and Delgado, G.: Water-retention capacity in fine earth and gravel fractions of semiarid Mediterranean Montane soils, Arid Soil Res. Rehab., 12, 29–45, https://doi.org/10.1080/15324989809381495, 1998. a, b

Poggio, L., de Sousa, L. M., Batjes, N. H., Heuvelink, G. B. M., Kempen, B., Ribeiro, E., and Rossiter, D.: SoilGrids 2.0: producing soil information for the globe with quantified spatial uncertainty, SOIL, 7, 217–240, https://doi.org/10.5194/soil-7-217-2021, 2021. a

Pourtaghi, Z. S., Pourghasemi, H. R., and Rossi, M.: Forest fire susceptibility mapping in the Minudasht forests, Golestan province, Iran, Environ. Earth Sci., 73, 1515–1533, https://doi.org/10.1007/s12665-014-3502-4, 2015. a

Preston, B. L., Brooke, C., Measham, T. G., Smith, T. F., and Gorddard, R.: Igniting change in local government: lessons learned from a bushfire vulnerability assessment, Mitig. Adapt. Strat. Gl., 14, 251–283, https://doi.org/10.1007/s11027-008-9163-4, 2009. a

Rad, A. M., AghaKouchak, A., Navari, M., and Sadegh, M.: Progress, Challenges, and Opportunities in Remote Sensing of Drought, in: Global drought and flood: Observation, modeling, and prediction, edited by: Wu, H., Lettenmaier, D. P., Tang, Q., and Ward, P. J., Geophysical monograph series, vol. 265, American Geophysical Union, 1–28, https://doi.org/10.1002/9781119427339.ch1, 2021. a, b

Razavi-Termeh, S. V., Sadeghi-Niaraki, A., and Choi, S.-M.: Ubiquitous GIS-Based Forest Fire Susceptibility Mapping Using Artificial Intelligence Methods, Remote Sens.-Basel, 12, 1689, https://doi.org/10.3390/rs12101689, 2020. a

Reyer, C., Bachinger, J., Bloch, R., Hattermann, F. F., Ibisch, P. L., Kreft, S., Lasch, P., Lucht, W., Nowicki, C., Spathelf, P., Stock, M., and Welp, M.: Climate change adaptation and sustainable regional development: a case study for the Federal State of Brandenburg, Germany, Reg. Environ. Change, 12, 523–542, https://doi.org/10.1007/s10113-011-0269-y, 2012. a

Ruffault, J. and Mouillot, F.: Contribution of human and biophysical factors to the spatial distribution of forest fire ignitions and large wildfires in a French Mediterranean region, Int. J. Wildland Fire, 26, 498, https://doi.org/10.1071/WF16181, 2017. a, b, c, d, e, f

Saidi, S., Younes, A. B., and Anselme, B.: A GIS-remote sensing approach for forest fire risk assessment: case of Bizerte region, Tunisia, Applied Geomatics, 13, 587–603, https://doi.org/10.1007/s12518-021-00369-0, 2021. a, b, c, d, e, f

Silva, C. V. J., Aragão, L. E. O. C., Barlow, J., Espirito-Santo, F., Young, P. J., Anderson, L. O., Berenguer, E., Brasil, I., Foster Brown, I., Castro, B., Farias, R., Ferreira, J., França, F., Graça, P. M. L. A., Kirsten, L., Lopes, A. P., Salimon, C., Scaranello, M. A., Seixas, M., Souza, F. C., and Xaud, H. A. M.: Drought-induced Amazonian wildfires instigate a decadal-scale disruption of forest carbon dynamics, Philos. T. Roy. Soc. B, 373, 20180043, https://doi.org/10.1098/rstb.2018.0043, 2018. a, b, c, d

Suryabhagavan, K. V., Alemu, M., and Balakrishnan, M.: GIS-based multi-criteria decision analysis for forest fire susceptibility mapping: a case study in Harenna forest, southwestern Ethiopia, Trop. Ecol., 57, 33–43, 2016. a

Thonicke, K. and Cramer, W.: Long-term Trends in Vegetation Dynamics and Forest Fires in Brandenburg (Germany) Under a Changing Climate, Nat. Hazards, 38, 283–300, https://doi.org/10.1007/s11069-005-8639-8, 2006. a

Wang, S. S.-C., Qian, Y., Leung, L. R., and Zhang, Y.: Identifying Key Drivers of Wildfires in the Contiguous US Using Machine Learning and Game Theory Interpretation, Earths Future, 9, 2020001910, https://doi.org/10.1029/2020EF001910, 2021.  a, b, c, d

Wen, X., Tu, Y.-H., Tan, Q.-F., Li, W.-Y., Fang, G.-H., Ding, Z.-Y., and Wang, Z.-N.: Construction of 3D drought structures of meteorological drought events and their spatio-temporal evolution characteristics, J. Hydrol., 590, 125539, https://doi.org/10.1016/j.jhydrol.2020.125539, 2020. a

Wu, H., Lettenmaier, D. P., Tang, Q., and Ward, P. J. (Eds.): Global drought and flood: Observation, modeling, and prediction, Geophysical monograph series, vol. 265, American Geophysical Union, print ISBN 9781119427308, online ISBN 9781119427339, https://doi.org/10.1002/9781119427339, 2021. a, b, c

Xu, F., Bento, V. A., Qu, Y., and Wang, Q.: Projections of Global Drought and Their Climate Drivers Using CMIP6 Global Climate Models, Water, 15, 2272, https://doi.org/10.3390/w15122272, 2023. a

Zhang, G., Wang, M., and Liu, K.: Forest Fire Susceptibility Modeling Using a Convolutional Neural Network for Yunnan Province of China, Int. J. Disast. Risk Sc., 10, 386–403, https://doi.org/10.1007/s13753-019-00233-1, 2019. a, b

Zhou, Z., Zhang, L., Chen, J., She, D., Wang, G., Zhang, Q., Xia, J., and Zhang, Y.: Projecting Global Drought Risk Under Various SSP-RCP Scenarios, Earths Future, 11, e2022EF003420, https://doi.org/10.1029/2022EF003420, 2023. a

Download
Executive editor
Forest fires have become a major problem in many regions of the world, including parts of Central Europe. The modelling study addresses the different factors for Forest Fire Susceptibility (FFS), making use of high spatial resolution of input data for the state of Brandenburg, Germany. An increasing susceptibility is found under rising greenhouse gas forcing scenarios when other changes are not taken into account. Extreme weather periods are of particular relevance in this respect. However, the importance of anthropogenic and vegetation parameters for modelling FFS on a regional level can outweigh the pure climatic effects. The paper also suggests some recommendations for forest management and environmental planning for a reduction of fire risk.
Short summary
In this study we applied a random forest machine learning algorithm to model current and future forest fire susceptibility (FFS) in north-eastern Germany using anthropogenic, climatic, topographic, soil, and vegetation variables. Model accuracy ranged between 69 % and 71 %, showing moderately high model reliability for predicting FFS. The model results underline the importance of anthropogenic and vegetation parameters. This study will support regional forest fire prevention and management.
Altmetrics
Final-revised paper
Preprint