1 Introduction

Forests maintain biodiversity, conserve soil, and safeguard watersheds in mountainous areas [2]. However, projected temperature increases in high-altitude ecosystems are greater than those at lower elevations [86]. This trend raises concerns about wildfire risk, which could significantly deteriorate ecosystem services. In the terrestrial environment, fire acts as an ecological tool in the evolution of vegetation when employed as controlled burning; otherwise, it acts as a destructive force [56]. Fire maintains the sustainability of forest ecosystems through nutrient cycling, controlling the density and growth of new trees, and modifying the habitat for species [55]. On the other hand, it also results in harming people’s health, degrading forests, and the functioning of the ecosystem [70], and affecting locals’ livelihoods, those who are highly dependent on the forest for their survival [52]. The challenges posed by forest fires have grown significantly complex, both worldwide and locally, threatening landscapes.

Studies show that the total area burned each year worldwide is estimated to be around 4 million km2, similar in size to India and Pakistan combined [33, 49]. The growing severity of forest fires in Asia, directly linked to climate change, has led to a 3–4% increase in the burnt area annually [31, 70]. Annually, forest fires destroy thousands of hectares of valuable forest resources in Nepal.໿ Mountainous regions face a growing threat from increasingly frequent and severe forest fires, jeopardizing biodiversity, ecosystem services, and human communities. Forest fires are the primary cause of deforestation and degradation in the High Mountains. An analysis of four years of fire incidents revealed that 10–15% of total forest fire events occurred in high-altitude districts of Nepal [12]. Elevation-driven warming observed worldwide, and specifically in Nepal, is a critical driver of forest fires [62, 81]. Warmer temperatures and widespread dryness during fire season (March-May) have diminished the effectiveness of high-elevation flammability barriers, facilitating fire spread in the upslope [6]. These regions, historically characterized by long fire return intervals and stand-replacing fires, are now experiencing more frequent blazes, threatening forest composition, community dynamics, and population structures [16, 57]. In addition to other impacts, climate-driven fires contribute to the spread of invasive grasses into high-elevation habitats, further threatening native vegetation [46].

Natural resource managers have been facing a challenge in controlling forest fires to achieve the desired outcomes due to the unpredictable nature of these events in new places and the increasing damage they cause [9]. The topography, climate, remoteness, and low population at higher elevations make it difficult to control once the fire ignites [12]. Moreover, due to a lack of resources, labor, and timely control mechanisms, manual fire control systems are unsuccessful in wide areas [41]. Therefore, it becomes imperative to identify the forest fire risk area before preventing, minimizing, or controlling fire risk and adapting strategies in case a fire breaks out. Modern technology and traditional knowledge applications are crucial in preventing, controlling, and managing forest fires. The combination of geographic information systems (GIS), remote sensing (RS), and modeling is becoming an essential tool in all aspects of wildfire management. In areas with restricted physical access or historical records, remotely sensed data has emerged as one of the critical components of risk mapping, especially in developing countries like Nepal [21].

Various approaches to forest fire risk modeling exist in the literature, each targeting specific aspects of fire behavior and risk. Forest fire research frequently uses logistic regression, multiple linear regression, Multi-Criteria Decision Analysis (MCDA) tools such as the Analytical Hierarchical Process (AHP), and fuzzy AHP. More recently, machine learning techniques have risen in popularity because of their ability to predict and model complex fire data patterns, data-driven, and manage intricate relationships between many factors [73]. Among them, Random Forest (RF) is a prevalent machine learning method introduced by Breiman [15], based on ensemble learning. In ecological studies, the RF model has demonstrated impressive accuracy and the capability to model complex interactions among variables [24], including in the assessment of fire susceptibility and the creation of risk maps.

The intricate mountainous terrain of Nepal poses a challenge in exploring the relationship between forest fires and potential causative factors. Determining the relative risk, in particular, helps create better plans for managing forest fires by minimizing the risk to people’s safety, property, and natural resources [65]. Forest fire frequency in Nepal’s mountains is escalating, yet localized risk mapping, particularly in Rasuwa district, is lacking. Several researchers investigated the spatial distribution of forest fire risk all over the country and in specific landscapes. However, these may not account for the unique climatic, topographic, and socioeconomic factors of mountainous terrains, which can influence fire behavior. For instance, Parajuli et al. [59] developed a spatial forest fire risk map in Nepal’s Terai Arc Landscape (TAL) and Chitwan-Annapurna Landscape (ChAL) regions, including Rasuwa District within a broad ecological framework.

Although existing landscape-level assessments provide valuable regional insights, they may limit the identification of local fire susceptibility when a single model is applied across high-altitude and environmentally diverse areas. Such broad-scale analyses inherently generalize local heterogeneity in topographic, anthropogenic, and biophysical regimes, fitting a single set of relationships across the entire study extent. This study directly addresses this gap; therefore, its primary contributions are spatial precision, methodology, and decision orientation, achieved by fitting the model at the district level.

The machine learning approaches in forecasting forest fire vulnerability have been widely applied; however, their application in the high-altitude mountainous regions of Nepal, like Rasuwa, is limited, underscoring the necessity for targeted modeling initiatives in these areas [11, 80]. A catastrophic wildfire of 2012 nearly devastated the Gatlang forest, burning unchecked for months [79]. Research has documented an increasing frequency of forest fire incidents in the region [69]. Previous research conducted by Paudel et al. [61] in the same geographic region highlighted the importance of long-term monitoring of anthropogenic fire impact on alpine shrublands and urged for sustainable management strategies involving local communities; however, there remains a crucial gap in the proactive prediction of fire risk.

Identifying these research gaps, this study aims to fill the void by developing a comprehensive forest fire risk map of the Rasuwa district, utilizing earth observation, geospatial datasets, and a RF model. Thus, it gives district and municipal authorities locally relevant evidence for prioritizing resource allocation, forest management, fire prevention, and risk information at their own jurisdictional scale.

2 Methodology

2.1 Study area

Rasuwa District, located within the Chitwan-Annapurna Landscape (CHAL) between 27°96' and 28°38' North latitude and 85°12' and 85°80' East longitude, is a mountainous area spanning 154,400 hectares (Fig. 1), characterized by distinct landscapes and rich cultural heritage. Langtang National Park covers a significant portion of the district, contributing to its rich biodiversity and a unique patchwork of vegetation shaped by diverse topography and geology (Fig. 1). The diverse landscape with forests, agricultural land, and snow-covered mountains includes 31.43% forest cover and 16.63% perpetual snow. While analyzing the elevation characteristics of the study area using Digital Elevation Model (DEM), the district’s elevation ranges from 614 to 7,227 m, embracing a wide range of vegetation, including montane tropical forests in the south, subtropical, temperate, subalpine, and alpine forests in the north; in addition, about 68% of the district area is situated above 3,000 m mean sea level.

Fig. 1
Fig. 1
Full size image

The location map of study area showing land use/land cover classes, along with the boundary of Langtang National Park

The meteorological station of Dhunche, the district headquarters, records the mean annual precipitation and temperature of 1605 mm and 15.6 °C, respectively, with the lowest temperature of 8.7 °C in January and the highest temperature of 20.6 °C in July [74]. Rayamajhi & Manandhar [69] found that from 1989 to 2018, the study area experienced an annual rise in the maximum and average temperatures by 0.0532 °C and 0.0202 °C, respectively. Furthermore, the increase in monsoon precipitation by 2.1 mm and the decrease in winter precipitation by 0.5 mm in the region highlight the consequences of climate change. The district has experienced recurrent forest fire incidents in recent years. In addition, studies report the trend of increasing temperature and declining winter precipitation in the region [69], which are known to elevate fire susceptibility. Rasuwa was selected due to its varied vegetation composition, complex terrain, altitudinal variations, and dependence of local communities on forest resources. These biophysical and socio-economic factors, along with the observed change in regional climatic patterns and the increasing incidences of forest fires, make it suitable to analyze the fire risk patterns in the high-altitude environment.

2.2 Datasets used

2.2.1 Forest fire incidents

In this study, forest fire incident data between 2012 and 2024, from the Suomi National Polar-orbiting Partnership - Visible Infrared Imaging Radiometer Suite (SNPP-VIIRS) sensor with a spatial resolution of 375 m, recorded specifically on the vegetated areas (forest areas, grassland, and other wooded land) were considered. To enhance the reliability and robustness of the data, data with nominal and high confidence levels above 30% were selected for the study [32].

To enhance the predictive capacity of the model, the spatial distribution of forest fire incidents must be calibrated or evenly distributed [14]. Therefore, the fire incidents were spatially filtered with a minimum distance threshold of 100 m between fire incidents. In the end, 757 incidents of forest fire remained after spatial thinning.

2.2.2 Forest fire conditioning factors

Initially, the most influential variables related to forest fire were selected from the literature [21, 52, 53, 60, 83] (Table 1). It is worth noting that vegetation (NDVI and landcover), topography (elevation, slope, aspect, and TWI), anthropogenic factors (distances from the settlement and proximity to road), and the meteorological factors (precipitation, temperature, and wind speed) were the main factors of forest fires in the high mountains [87]. Malczewski [50] suggests that model uncertainty can be mitigated by rescaling factors with different spatial resolutions; thus, predictor layers were resampled to 30 meters.

Table 1 Data sources, formats, and spatial resolution used for each predictor variable and forest fire incident record in this study, grouped by four different categories
Full size table
2.2.2.1 Biophysical factors

Vegetation affects forest fires and fire behavior [25]. In addition, dry and dense vegetation in closer proximity to human access is more prone to forest fires and spreads swiftly [21]. This study utilized the National Land Cover Map of Nepal for 2022, developed by the Forest Research and Training Centre (FRTC) as a part of the NLCMS project [30]. The map provides land cover classified into 11 specific categories (Fig. 1) and illustrates its variation across the country with a spatial resolution of 30 m. However, it has not differentiated the forest into distinct types, limiting the classification needed for effective forest fire risk mapping. Assessment of NDVI indicates vegetation density, health, and the soil and plants’ moisture content [40], which impacts the susceptibility to forest fire risk.

2.2.2.2 Topographic factors

Elevation plays a significant role in forest fire risk by regulating topographic, climatic, and hydrologic factors [27, 83]. At higher elevations, high rainfall, low temperatures, and high humidity result in fewer fire occurrences, unlike at lower elevations [17]. On steeper slopes, fuel preheats, and fires spread upward quickly [38]. A steeper slope influences the local wind speed and fosters an accelerating fire upslope [4]. Aspects significantly influence the microclimate of the slope: the amount of solar radiation received, temperature, moisture content, wind flow, and ultimately, vegetation growth [37]. TWI indicates the amount of soil moisture and surface saturation distributed spatially [89], influencing the initiation of forest fires [3]. The SRTM DEM, acquired from USGS Earth Explorer, was used to analyze elevation, slope, aspect, and TWI.

2.2.2.3 Anthropogenic factors

Proximity to settlements, a critical indicator, evaluates the pressure on the forest ecosystem through forest fires, deforestation, and overexploitation of forest resources imposed by anthropogenic activities [37, 38]. The spatial distribution of settlements in Rasuwa district was assessed with the “Settlements in Nepal” dataset from the UN OCHA, with a scale of 1:50,000. The dataset was rasterized using the Euclidean distance function, a spatial analyst tool in ArcGIS. People close to the forest may start fires deliberately or unintentionally [25]. Pedestrians and vehicles moving along the road may initiate the fires purposefully or accidentally. Dhakal et al. [21] point out that the road is an influential factor in the prevalence of anthropogenic fires. The Nepal Roads (OpenStreetMap Export) dataset of 2023, which is detailed and regularly updated, is utilized in this study. The dataset lacks a single specified scale; however, the vector layer was rasterized using the Euclidean distance function and resampled to 30 m resolution.

2.2.2.4 Climatic factors

Precipitation influences the variation of fuel and surface moisture [43]. Decreasing precipitation reduces fuel moisture and elevates the probability of forest fires, and vice versa. This study utilized the Bio-12 (annual precipitation) layer to model forest fire risk, a comprehensive set of bioclimatic variables. In ecological and environmental studies, the BioClim dataset is widely used because of its high resolution and accuracy [36]. Besides precipitation, rising temperatures significantly increase the evapotranspiration rate, drying out fuels and creating favorable conditions for forest fire risk. Temperature increases may also indirectly affect mountain ecology, such as forest fire lines moving uphill [6]. The MODIS Land Surface Temperature (LST), specifically, the MOD11C3 V6.1 product, which provides 8-day composite land surface temperature at a spatial resolution of 1 km, was processed using GEE. The LST dataset was refined to cover the observations spanning from 2000 to 2024, a total of 24 years; furthermore, Daytime LST values were extracted from the ‘LST_Day_1km’ band within the collection. A long-term average daytime LST was produced by aggregating the resulting images with a median reducer. While LST highlights the thermal conditions of the landscape, the wind depletes soil and fuel moisture, fans the flame with fresh oxygen, and exacerbates the risk of rapid ignition and vicious burning [42, 85].

2.3 Methodological framework

The overall methodological framework of the study was divided into three parts: data acquisition and preprocessing; preparation and preprocessing of forest fire and non-fire incidents before model configuration and training; and deployment of the model and validation of the final tuned model using the ROC curve and several statistical metrics (Fig. 2).

Fig. 2
Fig. 2
Full size image

Overall methodological framework of the study. Step 1 involves data processing for eleven predictor variables and screening for multicollinearity; step 2 involves training and testing, and step 3 involves preparation of the final forest fire risk map

2.4 Multicollinearity test

When two or more predictor variables are strongly correlated, rather than independent, this is known as multicollinearity. This might result in a less accurate assessment of the impact of an independent variable on the dependent variable [23]. Therefore, a multicollinearity test was performed to examine the correlation among the predictors. This study applied the most frequently used metrics: variance inflation factor (VIF) to assess the extent of multicollinearity. A VIF above 10 suggests high multicollinearity, resulting in poorly estimated regression coefficients, and requires attention [5]. Hence, only variables having a VIF ≤ 10 were taken into account for additional analysis.

2.5 Dataset configuration

The forest fire incidents, dependent variables, are divided into two classes: the occurrence (fire points) and non-occurrence (non-fire points). After spatial thinning, 757 forest fire incidents remain for model building. For accurate binary classification and to enhance the performance of machine learning, a supplementary dataset, i.e., non-fire points, was required [82]. In this study, non-fire incidents were generated randomly and intersected with fire incidents in a way that they do not overlap with fire incidents. Furthermore, the impact of class imbalance on model performance was assessed by sensitivity analysis using three labelled fire (1) and non-fire (0) incidents: 1:1, 1:2, and 1:5. These datasets were randomly grouped, where 80% was used as a training set for model training, while 20% was used as a test set for model validation to avoid potential overfitting and for validating the predictive performance of the model with unseen data.

2.6 Model training and configuration

In the realm of predictive modeling, this study employed the Random Forest (RF) model using the ‘random forest’ package [48] in R [68]. The RF model leverages multiple decision trees to improve prediction accuracy, enhance robustness to noise and outliers, and reduce overfitting [15]. Decision trees are grown from training data generated by randomly sampling the original data to increase the diversity, and as a result, greater classifier stability is achieved [15]. The final classification decision is as follows:

$$H\left(x\right)=\text{Argmax}_{y}\sum\limits_{i=1}^{k}I({h}_{i}\left(x\right)=Y$$
(1)

\(\:H\left(x\right)\) is the combination model, \(\:{h}_{i}\) is a single decision tree’s classification model, \(\:Y\) is the output variable, \(\:I(\cdot)\) is the indicator function [90].

In this study, to minimize overfitting and enhance the predictive performance of the model, a 5-fold cross-validation framework was implemented on the training dataset. Key hyperparameters were tuned using a systematic grid search: the number of decision trees (ntree), the minimum size of terminal nodes (nodesize), the number of variables randomly selected per split (mtry), and randomization. The search space included: (mtry: 1, 3, 5, 7, 9, 11), (node size: 1, 3, 5, 7), (ntree: 200, 300, 500, 700, 1000). For each combination of hyperparameters, the model was trained on four folds and evaluated on the remaining validation fold. This study aimed to identify the optimal parameter measured by AUC on the validation fold, a standard metric for evaluating the performance of binary classifiers [28]. The best predictive performance of the model was identified by comparing AUC scores at each iteration. The final RF model was then retrained on the entire training dataset using these best hyperparameters. Predicted probabilities from the RF model were also used to compute the Brier Score as a measure of probabilistic calibration.

2.7 Variable importance and partial dependency plots

Since the interaction of numerous factors usually influences hazards, the RF model has assessed two important measures to examine the impact of different variables on the predictive performance of the model: the mean decrease in accuracy (MDA) and the mean decrease in Gini (MDG). This study assesses the variable importance using MDG, which calculates the aggregate decrease in Gini impurity, which is supplied by each variable across all the trees within the ensemble. Although MDG is computationally efficient and widely used, it might show a bias with variables with numerous categories or continuous data [77]. Despite this, it offers a useful summary of variable importance for the study’s spatial classification system.

In addition to calculating variable importance, partial dependence plots (PDPs) were also calculated to complement the findings. This approach highlights the connection between individual features and outcomes, explicitly visualizing their effects on prediction results. A PDP illustrates the marginal impacts of one or more predictor variables on projected outcomes and is generated using the ‘pdp’ package in R [34].

2.8 Validation of the map

After generating a model, assessing its training and predictive performance is crucial. To determine the prediction efficacy of these selected approaches, validation is imperative in assessing forest fire vulnerability [37]. The models created for forest fire probability mapping were evaluated and compared employing receiver operating characteristic (ROC) and True skill Statistic (TSS). The ROC curve can be used as a numerical measure through the calculation of the area under the curve (AUC). The value of AUC ranges from 0.5 and 1.0, and a value closer to 1 indicates a model with a better predictive capability [88]. Moreover, TSS is a widely used metric in ecological and spatial modeling because it compensates for the limitations of kappa while keeping all of its advantages, and it evaluates predictive reliability by considering positive and negative prediction rates, a TSS value close to 1 indicates perfect agreement [8]. After generating a reliable forest fire risk model and following the risk assessment, the risk map was masked out with only the forest cover of the district for further analysis of risk area assessment and zonation. This step ensured that the risk analysis focused solely on areas that are susceptible to forest fires and hold ecological relevance.

3 Results

3.1 Multicollinearity assessment

Assessing multicollinearity among factors before susceptibility mapping is essential because high correlations between predictors can impact model stability, the clarity of variable importance, and the overall accuracy of predictions [23]. Thus, the initial step was to examine the relationship between the 11 predictive variables that influence fire ignition and spread before generating a reliable fire risk model. This study assessed variance inflation factors (VIF) for multicollinearity issues between predictors. However, one factor, i.e., precipitation (VIF = 20.14), was found to have multicollinearity issues; thus, the factor was eliminated in the final modeling. Since the estimated values of VIF indicate no significant linear association among the remaining factors, the remaining 10 factors were included for model development (Table 2).

Table 2 Variance Inflation Factor (VIF) values used to check multicollinearity. All values fall below the acceptable threshold of 10; eight variables are below 2.17, whereas LST (8.03) and elevation (7.52) were somewhat higher but still within a tolerable range
Full size table

3.2 Model validation

The model’s discriminative performance varied with the fire-to-non-fire ratio. The greatest accuracy was observed under the balanced 1:1 ratio. The ROC curve evaluated on the held-out 20% test dataset achieves an AUC value of 0.90 (Fig. 3), with a standard deviation of 0.0183 across cross-validation folds, indicating the reliability of the generated risk model in identifying vulnerable areas. The model achieved a True Skill Statistic (TSS) value of 0.67, which signifies its strong predictive performance with balanced capability in identifying both fire and non-fire areas. Additionally, the model’s effectiveness was assessed by Brier testing and obtained a value of 0.12, i.e., less than 0.25 for a non-informative model [76], indicating the established model had good prediction. Additional sensitivity analyses using alternative ratios (1:2 and 1:5) are provided in Supplementary Table S1.

Fig. 3
Fig. 3
Full size image

ROC curve for the random forest model. The red diagonal line represents random chance. An AUC of 0.9 indicates strong model performance

3.3 Variable importance analysis

Elevation is the most significant factor in forest fire risk mapping, followed by land surface temperature and NDVI, and TWI is the least significant factor (Fig. 4).

Fig. 4
Fig. 4
Full size image

Relative importance (%) of each predictor in Random Forest model, where X-axis shows importance and the Y-axis shows predictor variables

3.4 Partial dependence plots (PDPs)

Partial dependence plots (PDPs), (Fig. 5), illustrate the marginal effect of each predictor variable influencing the likelihood of forest fires on the expected probability of fires, while holding other variables constant.

Fig. 5
Fig. 5
Full size image

Partial dependence plots showing how predicted fire probability changes across the range of each predictor

The risk of fire peaks between 100 (East-Southeast) and 200 degrees (South-Southwest) of aspect. Likewise, within 3000 m of both settlements and roads, the fire risk is significant and stabilizes beyond that range. The risk of forest fire is most likely at lower elevations (from 1,000 to approximately 2,500 m), gradually diminishing as altitude increases. Conversely, the PDP largely shows a horizontal trend between 2,500 and 4,000 m elevation, with noticeable fluctuations. Moreover, the risk of fire is stabilizing around 4,200 m. Similarly, NDVI values between 0.4 and 0.6 showed a high possibility of fire risk. The presence of flammable grass and shrubs, and conifer species that can be prone to catching fire at this range, indicates an environment where fires are a risk, especially during dry periods or in areas affected by human activities, such as farming or grazing. The risk of fire is minimal on lower slopes, while it is highest on slopes ranging from 30 to 60 degrees; after that, it begins to decline slightly. Similarly, as the temperature rises, the likelihood of fire risk also increases, demonstrating a positive relationship between fire risk and temperature. At lower TWI levels, the likelihood of fire occurrences is higher and decreases as TWI increases, indicating the negative relationship between fire occurrences and TWI. Below 5 m/s, the probability of fire is at its peak; beyond this threshold, it stabilizes, showing the non-linear plateau effect between fire and wind speed.

3.5 Forest fire risk map (FFRM)

The forest fire risk map generated using the RF model was categorized into five risk classes through the Natural Breaks classification method (Fig. 6). The classification is based on the dispersal of the data, determining optimal natural gaps between the data by detecting an intrinsic pattern embedded within the data [39]. Furthermore, for ecological relevance, fire incident locations were overlaid on the final susceptibility map. Results show that 87.47% of all recorded fire incidents fall within the very high-risk class, while less than 1% occur within the low and very low class combined. Based on these findings, classification of class represents a meaningful and operationally significant fire risk area rather than purely statistical grouping. Based on fire risk assessment, 29.66% of the area is at High and Very High risk of forest fire. Moderate-risk areas, making up 25.25% of the region, are where fire is likely to occur occasionally. Meanwhile, regions with low or very low fire risk, consisting of 45.09% of the area, experience minimal fire susceptibility (Table 3).

Fig. 6
Fig. 6
Full size image

Forest fire risk map of Rasuwa district, classified into five risk levels from very low to very high

Table 3 Risk classes with their corresponding percentage contribution and area (in hectares) for the study region. High and very high-risk cover around 30% of the district’s forest
Full size table

The risk map illustrates that the west side of Langtang National Park in Rasuwa district falls under the Very High and High-risk classes. The regions of the west side have low elevation, high LST, and difficult terrain that hinders controlling the fire, which are the probable reasons for falling into highly risky areas.

3.6 Forest fire risk zonation

The forest fire risk zonation revealed differences in risk areas among various local bodies. Kalika Rural Municipality had the highest proportion of high-risk areas at 53.7%, while Gosaikunda had the least at 19.5%. However, the highest risk areas were located in Gosainkunda Rural Municipality with 6,235 hectares, then Aamachhodingmo with 4,340 hectares, and Uttargaya with 4,225 hectares (Table 4).

Table 4 Risk area distribution across local bodies, showing the extent of risk classes within each local administrative unit
Full size table

3.7 Descriptive analysis of fire incidents

The temporal analysis of fire incidents over the last 13 years reveals that approximately 91.86% of cases occurred from January to April, indicating these as the highly fire-prone months (Fig. 7). April recorded 283 counts, the highest number of fire incidents, while the months from July to October recorded no fire incidents in the past 13 years. Similarly, the annual fire statistics from 2012 to 2024 reveal a substantial decline in 2015 before increasing to more than 100 incidents in 2016. The escalated occurrences of 223 in 2023 from 24 in 2022 demonstrate the increasing trend of fires, especially in recent years. Overall, across 13 years, the Rasuwa District had a total of 835 fire incidents. Spatial analysis of fire incidents showed that forest fires occurred at an elevation of 4,525 m. In addition to this, analysis indicated that nearly 30% of incidents occurred at elevations exceeding 3000 m, demonstrating their occurrence even at high altitudes.

Fig. 7
Fig. 7
Full size image

Temporal fire pattern in Rasuwa district. The left figure demonstrates monthly incidents, and the right figure shows annual fire incidents from 2012–2024

4 Discussions

This study integrated remote sensing, geospatial data, and the RF algorithm to assess forest fire vulnerability, incorporating 10 causative factors in the Rasuwa district of Nepal. This study used VIIRS 375 m active fire data (VNP14IMG/VJ114IMG) from NASA over 13 years, which detects fire incidents that the Moderate Resolution Imaging Spectroradiometer (MODIS) overlooks due to its coarser resolution, i.e., 1 km. The trend of fire incidents in the Rasuwa district is increasing, with significant spikes in 2016, 2021, and 2023, which are among the highest forest fire counts in the country. This study observed that forest fires occur most frequently between March and May, when Nepal experienced a prolonged drought with nearly six months of no precipitation and heavy fuel accumulation on the forest floor. This observation aligns with prior research [21, 52, 59]. Furthermore, a higher number of fire incidents was observed in January, which aligns with the findings of Mishra et al. [53], who argued that a large portion of burnt areas at high elevations occurred from November to January. These winter fire incidents can be attributed to warm and dry winter conditions that reduce moisture levels in both the atmosphere and fuels [54], along with human activities such as land clearing, agricultural and pasture land burning [45], and tourism-related campfires at higher elevations.

The degree of influence of forest fire-causing factors may differ by the area and circumstances [26]. In this study, the RF model assigned the greatest importance to elevation, which aligns with the findings of Bar et al. [11], conducted in the Western Himalayas, India. The PDP curve for elevation reveals a distinct nonlinear relationship with the fire risk probability. From lower altitudes to 2,500 m, elevation governs substantial fire risk through fuel availability (resinous litter of conifer species), temperature gradient, and anthropogenic activities, which set fire either deliberately or accidentally [66]. Kumari et al. [44] argued that the elevation range between 2,500 m and 4000 m consists of subalpine forests, subalpine shrubland, or grasslands; in dry conditions, the vegetation can still facilitate the fire spread. This study highlights that even high-altitude regions, traditionally considered low-risk, can be susceptible. This highlights the necessity of including alpine and subalpine zones in fire risk assessments, where topographic and land-use complexities can create localized fire-prone environments. Following elevation, temperature received the highest importance score, indicating its significant impact. The observation aligns with the studies that conclude temperature and humidity have a greater impact on forest fires in higher-altitude areas [35], where they accelerate fuel drying and fire ignition. In Particular, temperatures above 15 °C had a strong positive impact, suggesting a higher risk of fire in warmer regions.

NDVI is a widely used metric for quantifying vegetation health and signifies fuel load [40] and has received great importance after the elevation and LST. This study found that needle-leaved open forests (0.4–0.6 NDVI) are the most prone to fire. With their high resin, low moisture content, and structural traits, coniferous trees make them highly flammable and cause rapid fire spread [20]. Alongside vegetation and topographic factors, anthropogenic activities increase the prevalence of forest fires. The road network and settlements near the forest influence the forest fire risk, as more incidents were recorded near these areas [21, 37]. The findings of this study revealed a progressive decline in risk as the distance from roads and settlements increased, with the highest fire likelihood within 3,000 m. Consistent with Pokhrel et al. [64], these fire risks may be linked to intentional burning by herders for seasonal pastoral land management, rather than permanent settlements and roads. The 30-60-degree slope showed the highest fire risks, and above it exhibited the lowest. This pattern may be attributable to fuel availability and human accessibility, as vegetation and human activity decline with increasing slope [10, 47]. While slope affects the fire spread rates, vegetation on the southern aspect witnessed higher fire incidents in this study due to faster humidity loss and increased flammability from sunlight exposure in the northern hemisphere, corroborating the findings of Tomar et al. [84]. Wind speed directly affects the rate of spread until other limiting factors like fuel and terrain take over [63]. In this study, wind-driven increases in risk appeared to plateau above 4 m/s, likely because fuel discontinuity, moisture variability, and topographic constraints limited further spread; in high-elevation areas, sparse vegetation may also have reduced fire occurrence despite strong winds [7].

While precipitation is a key climatic driver of the fire regimes, it was excluded from the final model due to severe multicollinearity issues (VIF = 20.14). The collinearity cluster reflects the well-established orographic precipitation mechanism of the Hindu Kush Himalayan system, wherein monsoon moisture interception and condensation are controlled primarily by topographic position [13], often limiting the independent explanatory power of precipitation. The TWI captures terrain-driven moisture accumulation at the local slope scale, which represents precise soil moisture availability [75]. Moreover, NDVI integrates cumulative water availability, fuel moisture, and vegetation condition [18]. However, these proxies cannot fully capture the event-based dynamics of precipitation, especially its role in short-term fuel moisture availability and fire weather interactions, and this limitation should be addressed in future work.

Forest fires are a major concern in the western part of Langtang National Park in Rasuwa District. Various studies and reports show the region’s vulnerability, including a news report on a 25-hectare wildfire in the park [51]. Notably, this study revealed that the areas up to 4,000 m in altitude are prone to fire risk, further reinforcing that forest fires are becoming more common at higher elevations. The study conducted by Dhungana et al. [22] highlighted fire occurrences in the high-elevation mixed conifer forest of Rasuwa, emphasizing their significant impact on the soil physicochemical properties. In a similar way, wildfires alter the properties of soil substrates and lower the precipitation threshold that causes soil failure and the formation of landslides [1]. The landscape change by post-earthquake, including landside scars and progressive vegetation recovery, may altered the fire regimes and increased interaction with downstream processes such as flooding [58]. Taken together, the observed correlations of fire–landslide–flood cascade process indicate that these hazards are not only independent but are also interconnected through common environmental controls [72]. Furthermore, these multi-hazards generate upstream–downstream consequences. These findings further reinforce the importance of a multi-hazard approach in mountain ecosystems. Thus, the created risk map is also a useful tool for integrated multi-hazard risk assessment.

Our result highlights the imminent risk of forest fire in Rasuwa, with 29.66% of the area in high and very high risk. While the forest fire risk map generated by Parajuli et al. [59] identifies Rasuwa district as part of a wider fire risk hotspot, particularly the southwest part. Our district-level map reveals a more complex pattern. Although the district is considered a hotspot, fire risk is highly inconsistent, with some areas showing low risk. Validation is essential for assessing the predictive accuracy and robustness of the generated fire risk map, which ensures the reliability of the results for forest fire risk assessment and management planning [29]. The model demonstrated the strong predictive performance, as indicated by an AUC and a TSS.

This study has a few limitations. The RF, operates as a black-box model since it does not assess regression coefficients or confidence intervals, limiting the clear understanding of its decision-making mechanisms [19, 67]. Furthermore, the model achieved high predictive accuracy in the study area, and there is uncertainty regarding model generalizability. The model’s rule may be highly tuned to the local spatial dependencies of the current dataset [71]. Thus, future research should use explainable AI techniques, such as SHapley Additive exPlanations (SHAP) values, to clarify individual variable importance and address interpretability challenges inherent in this algorithm [78]. The use of static predictor variables such as NDVI and landcover cannot fully represent temporal landscape changes across the 2012–2024 fire record, which remains a source of uncertainty and a limitation of the present study. Additionally, the study’s reliance on multiple, coarse-resolution datasets and the absence of ground-verified fire incident data may have compromised the spatial precision of the results. Wind speed data derived from the Global Wind Atlas provides long-term modeled averages and consistent spatial coverage; however, it may not capture localized wind variability in complex mountain topography. Future studies should integrate high-resolution data and ground truthing through field surveys or community-based participatory mapping for improved model accuracy and robust validation.

5 Conclusion

This study employed a combination of GIS and RF using environmental, topographical, biophysical, and anthropogenic variables to comprehensively analyze forest fire risk in the Rasuwa district of Nepal. According to the study, a total of 19523.04 hectares of Rasuwa district’s forest are at very high and high risk of fire, with Kalika and Uttargaya Rural Municipality being key vulnerable areas. The model’s high validation performance, with an AUC of 0.9, demonstrates its reliability in the spatial delineation of forest fire susceptibility. The resulting risk map provides actionable insight for decision-makers to prioritize fire management interventions and resource allocation in high-risk zones, thereby supporting ecosystem resilience. Our study demonstrated the efficacy of GIS and machine learning algorithms in forest fire risk assessment; however, the study had certain limitations. Further research could strengthen the methodology by incorporating fire causation factors, other explainable algorithms, and ensemble learning techniques to improve accuracy. Additionally, utilizing multi-source, high-resolution remote sensing data, including detailed socio-economic and vegetation-specific variables, and ground-truth validation would bolster the credibility and robustness of the findings at finer scales.