Introduction

Forests have been utilized by humans for centuries, leading to significant transformations in species compositions, structure, and the disturbance regimes of these ecosystems (Bebi et al 2017; Senf and Seidl 2021). Forest disturbance is defined as a short-term event causing crown cover or biomass loss (Hirschmugl et al 2017). The Food and Agriculture Organization of the United Nations (FAO) characterizes it as "damage caused by any factor (biotic or abiotic) that adversely affects the vigour and productivity of the forest and is not a direct result of human activities" (FRA 2020). Biotic disturbances, such as drought, forest fires, insect epidemics, windthrow, and hurricanes, as well as abiotic disturbances like illegal logging, harvesting operations, and artificial afforestation, can exert profound impacts on forests. These disturbances shape the structure and dynamics of forests, interacting with climate and other factors, including human land use (He and Mladenoff 1999; Turner 2010; Senf and Seidl 2021). Nepstad et al (2008) pointed out that climate change and the growing demand for wood products are increasing the rates of forest disturbance. Additionally, Lindenmayer et al (2004) emphasized that forest management practices can alter biological legacies and landscape structure, with potential feedback on subsequent disturbances. Therefore, gaining detailed information on forest disturbances and understanding changes in their regimes becomes crucial for the development of appropriate land management strategies. Consequently, the mapping of forest disturbances becomes even more critical, necessitating the use of reliable, operational methods for sustainable forest management (European Commission 2013; Chen et al 2021).

Numerous approaches have been developed to monitor and map forest disturbances, and one effective method is the analysis of susceptibility, which quantifies the probability of a specific phenomenon (e.g., landslides, forest fires) occurring under defined conditions (Li et al 2017). Susceptibility assessments typically utilize two quantitative methods: "knowledge base methods" and "data-based methods" (Corominas et al 2014). Alternatively, these methods can be categorized into heuristic, classical statistics, and machine learning methods (Reichenbach et al 2018; Bruzón et al 2021). The fundamental premise of susceptibility assessment is grounded in the idea that an event is more likely to occur under conditions similar to those that led to past events (Tien Bui et al 2012). While forest fire susceptibility has been extensively studied in the literature (Dolui 2023; Kantarcioglu et al 2023; Rihan et al 2023), other disturbances like severe windstorms (Garamszegl et al. 2022) and insect outbreaks (Sivrikaya et al 2023) have received limited attention. Beyond specific disturbances, susceptibility assessments have been applied to various subjects related to forests. For instance, Batar et al (2021) proposed a novel approach—weight of evidence—for mapping and assessing forest fragmentation susceptibility in the Indian Himalayan Region. Sahana et al (2018) assessed susceptibility to deforestation using a fragmentation approach and the frequency ratio model. Shabani et al. (2020) mapped forest stand susceptibility during harvesting using logistic regression and boosted regression tree machine learning models.

In susceptibility mapping studies related to forestry, the focus has traditionally been on single phenomena, such as forest fires, windthrow, or insect outbreaks. However, in recent studies, there has been a shift towards assessing susceptibility to multiple phenomena, primarily centred around various hazards. For instance, Guerriero et al (2022) conducted a comprehensive multi-hazard susceptibility assessment using the analytic hierarchy process in the Derwent Valley Mills UNESCO World Heritage Site (United Kingdom). Their study encompassed surface flooding, groundwater flooding, compressible ground, landslides, and running stages as hazardous events. Similarly, Ullah et al (2022) employed convolutional neural networks to map multi-hazard susceptibility, considering flash floods, debris flow, and landslides. Rusk et al (2022) utilized the maximum entropy (MaxEnt) machine learning method to map multi-hazard susceptibility in the Hindu Kush Himalaya. Pourghasemi et al (2020) adopted the random forest machine learning method to conduct a multi-hazard probability assessment in Fars Province (Shiraz City, Iran), encompassing floods, landslides, and forest fires. In Ankara City (Turkey), Yanar et al (2020) applied the Mamdani fuzzy algorithm to assess multi-hazard susceptibility, focusing on floods and landslides. Regarding forest disturbances, the literature has limited studies in this domain. A recent noteworthy study by Stritih et al (2021) assessed the susceptibility of forests to various natural disturbances, including avalanches, bark beetles, fires, heavy snow, windthrow, and harvests in Graubünden, Switzerland. This study utilized the random forest machine learning method.

It appears that the utilization of machine learning methods in susceptibility assessment is on a trajectory of continual growth across various research domains. However, the success of developing models is critically dependent on the expertise of machine learning specialists. In traditional machine learning, models are crafted by handling each step in the process separately. In simpler terms, the efficacy of machine learning models relies on the collaboration between the right tools and the right expertise (Özdemir and Örslü 2019). Developing machine learning models presents technical complexities that extend beyond the grasp of non-experts in machine learning (Waring et al 2020). This technical complexity is attributed to various steps in the process, including: (1) pre-processing and cleaning data, (2) selecting and constructing appropriate features, (3) choosing an appropriate model family, (4) optimizing model hyperparameters, (5) designing the topology of neural networks if deep learning is employed, (6) post-processing machine learning models, and (7) analyzing obtained results (Hutter et al 2019; Waring et al 2020; URL 2023). The tasks of finding suitable models and optimizing hyperparameters prove challenging and time-consuming for human experts (Halvari et al. 2020). The heightened demand for machine learning methods that are readily applicable without the need for expert knowledge has led to the concept of automated machine learning (AutoML). AutoML is defined as meta-level machine learning algorithms functioning as a platform or open-source library for automating each step in the machine learning process (Hutter et al 2019; Halvari et al 2020).

The utilization of the AutoML framework in susceptibility assessment is relatively scarce in the existing literature, as noted by Bruzón et al (2021) and Renza et al (2021). A recent study in the field of forestry, conducted by Ren et al (2022), delved into the spatiotemporal assessment of forest fire vulnerability. Building on this, this study aims to assess forest disturbance susceptibility employing the AutoML framework, specifically utilizing Python PyCaret. In this study, three distinct forest disturbances— forest fires, insect damage, and wind damage—were considered. The susceptibility assessment was conducted individually for each disturbance, as well as collectively for multiple disturbances. Additionally, a comprehensive map illustrating the potential occurrences of forest disturbances was generated.

Material and methods

Study area

The study area selected for this research is the Izmir Regional Forest Directorate (RFD), one of 28 forest regional directorates in Turkey. Situated in the northern longitude range of 39°35ʹ38"–37°52ʹ30" and eastern latitudes of 26°11ʹ42"–28°52ʹ28"(Fig. 1), Izmir RFD encompasses almost one-third of the Aegean region of Turkey, including the cities of Izmir and Manisa. The region's forests cover an area of 2,509,336 ha, approximately 46% of the total area. The primary forest species in this region are Turkish pine (Pinus brutia Ten.) and black pine (Pinus nigra Arn.) (Kucukosmanoglu and Uzmez 2019). The distribution of forest types across the region is shown in Table 1. Elevation within the Izmir RFD varies from sea level to 2148 m, featuring slopes ranging from zero to 90%. The region is crisscrossed by a network of roads spanning 40,948 km, with a density up to 10 m/km2. The region has a typical Mediterranean climate characterized by warm and rainy winters and hot, dry summers. While the coastal areas more directly experience the Mediterranean climate, the interior part of the region, influenced by local topography, also reflects these climatic patterns. In Izmir RFD, winter precipitation is predominantly rain, and the annual precipitation fluctuates between 700 and 1000 mm. The hottest months are July and August.

Fig. 1
figure 1

Location of study area: Izmir RFD

Full size image
Table 1 Distribution of forest types in Izmir RFD
Full size table

Situation of forest disturbances in the Izmir RFD

The Izmir RFD is one of the more vulnerable forest regions to various disturbances in Turkey, with a particular emphasis on forest fires (Akyuz and Kucukosmanoglu 1997; Kucukosmanoglu and Uzmez 2019). According to forest fire statistics from the General Directorate of Forestry (GDF) for 2022, 23,938 ha were burned between 2004 and 2022 (GDF 2023). The annual breakdown of burned areas in Izmir is shown in Fig. 2. In contrast to the availability of statistics on forest fires, there is a lack of published data concerning wind and insect damage. The examination of forest disturbances in this study extended to an investigation of official reports accessible in the Izmir RFD archives covering the years 2010–2021. These records provided valuable details, including the extent of salvage logging for each disturbance (fire, insect, wind), as well as information on forest stand number and type, and the name of the forest district. Based on these records, it is evident that four main insect species—Tomicus destruens Woll. (Coleoptera: Scolytinae), Orthotomicus erosus Col. (Curculionidae: Scolytinae), Thaumetopoea pityocampa (Den. & Schiff.) (Lepidoptera: Thaumetopoeidae), and Thaumetopoea wilkinsoni Tams. (Lepidoptera: Thaumatopoeidae)—have predominantly contributed to damage in the region's forests. Sarıkaya et al. (2013) emphasized that bark beetles (Scolytinae) are particularly impactful defoliators in Izmir province, with the Mediterranean pine engraver, Orthotomicus erosus Woll. ranking high among Scolytinae species, causing significant annual economic losses.

Fig. 2
figure 2

Forest fires statistics (2004–2022) in Izmir RFD showing burned area in hectares (GDF 2023)

Full size image

Figure 3 illustrates the volumes of salvage logging attributed to forest fires, insect damage, and wind damage for each year. Examining the records of fire damage from 2010 to 2021, a 12-year period, reveals a discernible increasing trend despite occasional fluctuations in salvage volumes. The peak of salvage volume due to fire damage was in 2019, reaching 506,698 m3. Similarly, both insect and wind damages exhibited an upward trajectory during this 12-year period, accompanied by intermittent fluctuations in salvage volumes. The highest salvage volume attributed to insect damage occurred in 2018, totalling 26,625 m3. For wind damage, the peak volume was 23,686 m3 in 2015.

Fig. 3
figure 3

Time series of salvage logging volume due to forest fire, wind, and insect damage between 2010 and 2021 in Izmir RFD

Full size image

Dataset of forest disturbances

This study focused on three forest disturbances—forest fires, insect damage, and wind damage—that transpired between 2010 and 2021. Official reports from the Izmir RFD documenting each disturbance were incorporated into the GIS environment by vectorizing them as point layers. These points are referred to as "presence data (1)" (Fig. 4). Due to the absence of precise location information in the reports, presence points were digitized randomly within the damage area associated with each disturbance. The reports only provided details on forest stand numbers and the type of damage areas within the relevant forest districts. The vectorization process was executed on a base map of forest stand types obtained from the General Directorate of Forestry (GDF) using ESRI ArcGIS software. Consequently, the total number of points varied based on the affected area's size: 4347 points for insect damage, 1217 points for wind damage, and 54,593 points for forest fire damage in Izmir RFD.

Fig. 4
figure 4

Forest disturbance (forest fire, wind, and insect damage) presence points (2010 − 2021) in Izmir RFD. The lower map shows presence points randomly distributed within stands

Full size image

Forest disturbances susceptibility assessment with an AutoML framework

The susceptibility assessment was conducted using PyCaret 3.0 (https://pycaret.org/), an AutoML framework. PyCaret is an open-source, low-code machine-learning library in Python designed to automate machine-learning workflows (https://pycaret.gitbook.io/docs/). It was selected for susceptibility assessment due to its capability to facilitate complex machine learning tasks, allowing data scientists to efficiently conduct end-to-end experiments. The workflow is illustrated in Fig. 5. The application of Python AutoML (PyCaret 3.0) was executed within the Jupyter Notebook environment, a web application designed for creating and sharing computational documents (https://jupyter.org/). This environment provided a versatile platform for implementing and documenting the susceptibility assessment workflow.

Fig. 5
figure 5

The study workflow

Full size image

In the data preparation step, alongside the presence point data for each forest disturbance, a set of driving factors (or features)—categorized as topographic, structural, climatic, and stand features—were identified based on the literature. A total of 15 driving factors were considered: elevation, slope, aspect, topographic wetness index (TWI), distances to roads, streams, agricultural areas, and settlements, road density, stand types, degree of canopy closure, stand development stages, total annual precipitation, mean temperature of the warmest quarter, and maximum wind speed. All driving factors were created in raster data format with a 50 m × 50 m pixel size and uniform dimensions (5244 pixels in rows and 3589 pixels in columns). Supplementary Material (Figs. S1 − S15) provides an overview of these driving factors. Topographic features such as elevation, slope, aspect, and TWI were derived from a digital elevation model (DEM) from the GDF. Forest features, including stand types, canopy closure, and stand development stages, were extracted from a forest map from the Izmir RDF. Additionally, road density, distances to roads, streams, agricultural areas, and settlements were created using maps obtained from Izmir RDF. Climatic features were generated from a dataset acquired from worldclim.org (Fick and Hijmans 2017). Of the 15 driving factors, 11 had continuous data, while four (aspect, stand types, canopy closure, stand development stages) were categorical. The values of each driving factor at presence points were extracted and formatted into a CSV file for use in the AutoML process. In the data analysis step, before running the AutoML process, Pearson correlation analysis was performed to assess multicollinearity among driving factors with non-categorical data types. This step was crucial as the strength and direction of the relationship between two factors could impact model performance. The Pearson correlation matrix for each forest disturbance is provided in Supplementary Material (Figs. S16 − S18). The PyCaret AutoML process incorporates multicollinearity testing, where highly correlated factors are assessed to eliminate one from each pair. When the parameter is set to "True," features with inter-correlations surpassing the defined threshold are removed. Specifically, for each group, only the feature with the highest correlation to the y-axis is retained (more details can be found at https://pycaret.gitbook.io/docs/). In this study, a threshold value of 0.90 was set for this parameter concerning each forest disturbance. PyCaret AutoML conducted tests on 14 out of the 17 available machine learning algorithms (Table 2) and subsequently ranked their classification performance based on AUC (area under the curve) values.

Table 2 Machine learning algorithms compared in PyCaret AutoML process
Full size table

The best model based on the highest AUC, was employed to predict forest disturbance susceptibility for all pixels within forested areas. In addition to AUC, key evaluation metrics such as accuracy, recall, precision, F1 score, Kappa, and Matthews correlation coefficient (MCC) were considered for a comprehensive assessment (Wang et al 2021). To calculate these metrics, it is essential to clarify the four kinds of predicted samples for classification learning (Wang et al 2021): (1)True Positive (TP): The predicted class is positive, and the prediction aligns with the actual class; (2) False Positive (FP): The predicted class is positive, but the prediction disagrees with the actual class; (3)True Negative (TN): The predicted class is negative, and the prediction aligns with the actual class; (4) False Negative (FN): The predicted class is negative, but the prediction disagrees with the actual class.

Equations Eqs. 1 − 6 were used to calculate these parameters:

$$Accuracy=\frac{TP+TN}{TP+FP+TN+FN}$$
(1)
$$Recall=\frac{TN}{TN+FN}$$
(2)
$$Precision=\frac{TP}{TP+FP}$$
(3)
$${F}_{1}Score=\frac{2\times Precision\times Recall}{Precision+Recall}$$
(4)
$$Kappa=\frac{2\times (TP TN-FP FN)}{(TP+FP)\times (TP+FN)\times (TN+FP)\times (TN+FN)}$$
(5)
$$MCC=\frac{TP\times TN-FP\times FN}{\sqrt{(TP+FP)\times (TP+FN)\times (TN+FP)\times (TN+FN)}}$$
(6)

These equations provide a quantitative assessment of the model's performance in predicting forest disturbance susceptibility.

In the mapping and assessment step, the predicted results from AutoML were transformed into raster data format to create maps for each disturbance. To generate a multiple forest disturbance susceptibility map, a weighted overlay method was implemented in GIS. In this method, each susceptibility map served as input, with weights assigned as 0.5 for forest fires, 0.2 for wind damage, and 0.3 for insect damage. The resulting susceptibility maps were then classified into five categories: very low (0 − 0.2), low (0.2 − 0.4), moderate (0.4 − 0.6), high (0.6 − 0.8), and very high (0.8 − 1.0). In addition, a potential forest disturbance map was created by employing an overlapping method using the susceptibility map of each disturbance. Prior to the overlapping process in GIS, a binary classification was applied to each susceptibility map based on defined rules: if the susceptibility value was higher than 0.75, the pixel value was set to 1; otherwise, the pixel value was set to 0. The resulting map categorized areas into eight classes: (1) no damage, (2) forest fire damage, (3) wind damage, (4) wind + forest fire damage, (5) insect damage, (6) insect + forest fire damage, (7) wind + insect damage, and (8) wind + insect + forest fire damage. While a susceptibility map represents the vulnerability of an area to specific disturbances, a potential forest disturbances map integrates information from multiple susceptibility maps through an overlapping process, offering a comprehensive visualization of areas prone to various disturbances categorized into distinct classes. Additionally, the SHAP (Shapley Additive exPlanations) methodology (Lundberg and Lee 2017) was used as an innovative tool to illustrate the nonlinear relationship between explanatory features and forest disturbance susceptibility. SHAP values, computed through a game theoretical approach, elucidate the contribution of each feature within a model (Li et al 2020; Bugaj et al 2021). For the SHAP analysis, the "shap" Python library was utilized, and SHAP summary plots were generated to provide feature attribution information for all instances in every model.

Results

Forest disturbance susceptibility assessment

An AutoML framework, specifically PyCaret 3.0, was employed to assess the susceptibility to forest disturbances in Izmir RFD. The assessment considered three primary forest disturbances: forest fires, insect damage, and wind damage. Additionally, a comprehensive analysis of susceptibility to multiple forest disturbances was conducted. After comparing various algorithms using AutoML, the extra trees classifier (ET) algorithm emerged as the best model for susceptibility analysis in each disturbance category. Detailed model performances for each dataset are provided in Supplementary Data (Figs. S19 − S21). The forest disturbance susceptibility maps for each individual disturbance (Figs. 6, 7 and 8) and the map depicting susceptibility to multiple disturbances (Fig. 9) were generated based on the selected ET algorithm. Each susceptibility map was categorized into five classes, and the areal distributions of susceptibility classes in the forested areas are shown in Table 3. Notably, the results reveal that 82.5% of the forests in Izmir RFD are susceptible to multiple disturbances at high and very high levels. Additionally, 88.9% are susceptible to forest fires, primarily falling into high and very high classes. Wind damage susceptibility is observed in 66.8% of forested areas at high and very high levels, while 76.9% show susceptibility to insect damage.

Fig. 6
figure 6

Forest disturbance susceptibility map for forest fire damage (basemap: a grayscale hillshade)

Full size image
Fig. 7
figure 7

Forest disturbance susceptibility map for wind damage (base map: a grayscale hillshade)

Full size image
Fig. 8
figure 8

Forest disturbance susceptibility map for insect damage (base map: a grayscale hillshade)

Full size image
Fig. 9
figure 9

Multiple forest disturbances susceptibility map (base map: a grayscale hillshade)

Full size image
Table 3 Area (hectares and percentage) of susceptibility classes for disturbances in forested areas
Full size table

In the current study, SHAP methodology was used to explain model predictions, providing insights into the contribution of each feature to susceptibility modeling. SHAP summary plots, shown in Fig. 10 for insect damage, Fig. 11 for wind damage, and Fig. 12 for forest fire damage, effectively combine feature importance with feature effect. In these plots, the x- and y-axes represent the Shapley value and the value of the feature, respectively, with colours indicating the feature's value from low to high. Features are ordered based on their importance, revealing road density as the most critical feature for susceptibility in each disturbance. According to the SHAP analysis, key features influencing model predictions for insect damage include road density, distance to roads, forest type, annual precipitation, elevation, maximum wind speed, and mean temperature of the warmest quarter. For wind damage, important features include forest type, annual precipitation, distance to roads, elevation, mean temperature of the warmest quarter, and maximum wind speed. Similarly, for fire damage, significant features are distance to roads, annual precipitation, maximum wind speed, elevation, mean temperature of the warmest quarter, and forest type.

Fig. 10
figure 10

SHAP summary plots for susceptibility modelling of insect damage (1: Road density, 2: Distance to roads, 3: Forest type “1”, 4: Forest type “5”, 5: Annual precipitation, 6: Elevation, 7: Maximum wind speed, 8: Mean temperature of warmest quarter, 9: Distance to settlements, 10: Distance to streams, 11: Distance to agricultural areas, 12: Slope)

Full size image
Fig. 11
figure 11

SHAP summary plots for susceptibility modelling of wind damage (1: Road density, 2: Forest type “5”, 3: Annual precipitation, 4: Distance to roads, 5: Elevation, 6: Mean temperature of warmest quarter, 7: Maximum wind speed, 8: Forest stand development stage “6”, 9: Distance to settlements, 10: Distance to agricultural areas, 11: Slope, 12: TWI)

Full size image
Fig. 12
figure 12

SHAP summary plots for susceptibility modeling of forest fire damage (1: Road density, 2: Distance to roads, 3: Annual precipitation, 4: Maximum wind speed, 5: Elevation, 6: Mean temperature of warmest quarter, 7: Forest type “5”, 8: Distance to settlements, 9: Distance to streams, 10: Distance to agricultural areas, 11: Slope, 12: TWI)

Full size image

Using the susceptibility maps for each disturbance, a potential forest disturbances map was generated through an overlapping method (Fig. 13). The extent of potential damage in hectares and as a percentage for each category is detailed in Table 4. Notably, 15.6% of the forested areas in Izmir RFD are projected to have no likely damage from the considered disturbances. At the same time, 15.1% could potentially experience damage from forest fires, and a significant portion, 54.2%, could be susceptible to all three disturbances. Specifically, a minimum 0.1% of forests may face possible damage exclusively from wind, while only 0.5% might experience damage solely from insects. The evaluation extends to different forest types, as depicted in Fig. 14, shows the percentage area of each type potentially affected by the disturbances, with colour gradations from green to red indicating the percentage within each row.

Fig. 13
figure 13

Potential forest disturbances exposure map (above: entire study area, below: area inside black rectangle)

Full size image
Table 4 Area (hectare and percentage) of potential forest disturbances
Full size table
Fig. 14
figure 14

Area distribution (percentage) of potential disturbances in forest types

Full size image

Discussion

Susceptibility assessment employing machine learning methods and GIS mapping in forestry studies, particularly for disturbances like forest fires, has become more prevalent (Dolui 2023; Kantarcioglu et al 2023; Rihan et al 2023). However, most studies have focused on a single forest disturbance, with the exception being Stritih et al. (2021) who considered multiple disturbances. They used the random forest (RF) algorithm, whereas our study used the AutoML framework comparing 14 machine learning algorithms and ranking the best based on AUC values. The ET algorithm was selected for susceptibility assessment, with AUC values consistently exceeding 0.98 for each disturbance.

The ET algorithm, similar to the RF algorithm, is an ensemble machine learning method. The key distinction is that the ET algorithm uses the entire dataset, reducing model bias, and is computationally more efficient. The primary reason for using AutoML is its ability to automate the entire machine-learning process, including data pre-processing, model selection, and hyperparameter tuning (Hadi et al 2023). Although alternative AutoML frameworks exist, such as Auto-sklearn, AutoKeras, PyCaret, and TPOT (Kumar et al 2022), PyCaret has been predominantly used in the literature (Bruzón et al (2021).

Additionally, susceptibility assessment requires driving factors (or predictors). In this study, 15 driving factors were selected based on reported disturbance drivers in the literature. For example, topographic and stand features (excluding canopy height) were used, like the approach by Stritih et al (2021). Climate factors such as annual precipitation, mean temperature of the warmest quarter, and maximum wind speed also were incorporated. Importantly, all factors were applied to all disturbances for a comparative analysis of their effects across different disturbances.

The creation of a multiple forest disturbances susceptibility map involved a weighted overlay analysis of each forest disturbance susceptibility map. Different approaches in the literature, including those by Pourghasemi et al. (2020), Yanar et al. (2020), Rusk et al. (2022) and Ullah et al. (2022), can be considered for such analyses.

Moreover, a potential disturbances exposure map was generated for the first time in this study, incorporating wind, insect, and forest fire damage. This map illustrates forested areas that could potentially be damaged by the disturbances considered. In the literature, such maps were predominantly created for multiple hazards. Pourghasemi et al (2020) generated a multi-hazard map that included forest fires as a disturbance but did not include wind and insect damage. The potential forest disturbances map was overlapped with the stand type map to estimate the extent to which areas of each could potentially be damaged by the disturbances. According to the results of the overlay analysis, Turkish pine species and black pine species, either purely or in mixed form, face the greatest damage from all three disturbances. For instance, only 4.3% of the forest type “Pure Turkish pine species,” one of the main species in the Izmir RFD, could potentially escape damage from the disturbances, while more than 70% may face potential damage from all the disturbances. Forest fire damage could be observed in approximately 95% of the area covered by Pure Turkish pine species. Küçük and Bilgili (2008) noted that Turkish pine and black pine are highly sensitive to fires. Regarding insect damage, it has been reported that Turkish pine species in particular, are susceptible to insects such as Orthotomicus erosus (Sarıkaya et al 2013).

Forest types such as “Mixed broadleaved species,” “Broadleaved species dominated coniferous species,” and “Other coniferous dominated broadleaved species” have the largest percentage area classified as "no damage." As Jactel et al (2017) highlighted, mixed forests are more resistant to natural disturbances that are relatively small-scale and selective in their effect. Interestingly, this study revealed that the forest type “Pure broadleaved species,” as a pure forest stand type, also has the largest percentage area of potentially no damage by disturbance. Additionally, open forested areas (i.e., gaps in stands) exhibited the highest percentage of no damage, as expected. Furthermore, wind and insect disturbances had in a lower percentage than forest fires. This is likely due to these areas often covered with grasses and small shrubs and lacking trees or containing a few scattered individuals.

The SHAP method was used to assess feature importance on model predictions and explain non-linear relationships between predictor variables and forest disturbances susceptibility. SHAP serves as a valuable tool for enhancing the interpretability of machine learning models, aiding in comprehending the relationships between input and output, and identifying the most influential features in driving predictions (Li 2022). However, as Li (2022) emphasized, the use of SHAP to provide reliable estimates requires accurate and robust machine learning models, and SHAP values should be interpreted with special attention.

Iban and Sekertekin (2022) compared SHAP outputs for different machine learning models, and their findings showed that the top three contributing factors for modeling fire susceptibility were elevation, temperature, and slope. In the present study, road density, distance to roads, and annual precipitation were identified as the most crucial factors. The rankings also highlighted the significance of maximum wind speed, elevation, and mean temperature of the warmest quarter. As expected, lower road densities and greater distances to roads decrease the susceptibility to forest fires. In addition, high-speed winds increase the susceptibility to fires. Despite the findings of Iban and Sekertekin (2022), who found lower temperatures increased the risk of wildfires, our study revealed a positive impact of higher temperatures on susceptibility to forest fires. Similarly, while distance to settlements was moderately important, in contrast, distance to roads emerged as the top-ranked feature. This discrepancy could be attributed to the exclusive consideration of highways and state roads in our analysis. In the susceptibility analysis for insect damage, road density and distance to roads emerged as the most important features. The third-ranked feature was forest type 1, i.e., pure Turkish pine species. As noted previously, Turkish pine is prone to insect attacks (Sarıkaya et al 2013). Another noteworthy feature is forest type 5 (forest gaps), identified as no insect damage was observed in these areas. Consequently, potential insect damage based on predicted susceptibility maps was observed in very limited areas of forest gaps. Regarding wind damage, maximum wind speeds showed the expected impact, with higher values increasing susceptibility. However, wind speed was ranked less important than other factors such as road density, annual precipitation, and distance to roads.

Conclusions

This study emphasizes the importance of integrating multiple forest disturbances into susceptibility assessments through machine learning, utilizing the AutoML framework. This approach facilitates the comparison of different algorithms, aiding in the selection of the best-performing one. The study reports on multiple forest disturbances susceptibility assessments, considering wind, insect, and forest fire damages, resulting in the creation of four forest disturbances susceptibility maps. This comprehensive analysis provides valuable insights and may be extended to other forested areas by incorporating additional disturbances and selecting an appropriate number of driving factors. The methodology proposed, encompassing topographic, climate, and stand features, allows for large-scale susceptibility assessments of forest disturbances. It provides a foundation for assessing different scenarios based on anticipated climate change. The potential forest disturbances exposure map, overlaid with forest types, offers an understanding of the interactions between different forest types and possible disturbances. Susceptibility assessment contributes to a better awareness of the factors driving forest disturbances, with novel methods like SHAP enhancing the relationship between input and output of models. This method, emphasizing the importance of specific factors, may be preferred in susceptibility analysis of forest disturbances using machine learning algorithms. PyCaret AutoML further streamlines the identification of the most important driving factors. All generated maps, including susceptibility maps, potential forest disturbance exposure maps, and maps highlighting important driving factors, serve as valuable base maps for forest managers in decision-making processes. These maps consolidate information about various disturbances within a specific area, providing essential insights to develop strategies for combating disturbances and contributing to sustainable forest management practices.