Open AccessArticle

Correlating Fire Incidents with Meteorological Variables in Dry Temperate Forest

Khurram Abbas

Ali Ahmed Souane

Hasham Ahmad

Francesca Suita

Zhan Shu

^1,*

Hui Huang

^3,* and

Feng Wang

⁴

School of Forestry, Northeast Forestry University, Harbin 150040, China

Service for Analysis, Research and Management of Wild Animals (SAIGAS), Faculty of Veterinary Medicine, Cardenal Herrera-CEU University, CEU Universities, 46115 Valencia, Spain

School of Water Conservancy and Civil Engineering, Heilongjiang Agricultural Engineering Vocational College, Harbin 150025, China

⁴

Heilongjiang Provincial Institute of Natural Resource Rights and Interests Investigation and Monitoring, Harbin 150080, China

Authors to whom correspondence should be addressed.

Forests 2025, 16(1), 122; https://doi.org/10.3390/f16010122

Submission received: 20 November 2024 / Revised: 7 January 2025 / Accepted: 9 January 2025 / Published: 10 January 2025

(This article belongs to the Special Issue Managing Forest Wildfires in Climate Changes: New Paradigms and Challenges)

Download

Browse Figures

Figure 1
Map of Diamer District, Gilgit-Baltistan. "> Figure 2
(a) Digital elevation model value, (b) hill shade value, (c) aspect value, and (d) slope value. "> Figure 3
(a) Diamer district’s annual burned area of forest fires, 1998–2023; (b) Diamer District’s annual forest fire occurrence of different causes from 1998 to 2023. "> Figure 4
(a) Diamer District, Julian dates of fires from 1998 to 2023; (b) Diamer District, Julian dates of huge fires from 1998 to 2023. "> Figure 5
Julian dates of the earliest and latest forest fires in the Diamer District from 1998 to 2023. blue bullets and line signifies early fire events and its regression, red bullets and line shows later fire events and its regression. "> Figure 6
Diamer District, annual climate conditions in May, June, and July from 1998 to 2023. (a) temperature; (b) relative humidity; (c) annual rainfall; (d) wind speed. "> Figure 7
Climate–fire relationships in the Diamer District from 1998 to 2023. (a) temperature; (b) relative humidity; (c) rainfall; (d) wind speed. bullets, lines, and shadows represents fire events, regression and correlation, respectively. "> Figure 8
Climate–fire relationships in the Diamer District in May from 1998 to 2023. (a) temperature; (b) relative humidity; (c) rainfall; (d) wind speed. "> Figure 9
Squared error loss across iterations for random forest and GBM models. The black curve is training set loss and the green and red curves represent validation loss for the Random Forest and Gradient Boosting Machine models, respectively. The blue dashed line represents the optimal point of iteration for the Gradient Boosting Machine model. The figure represents the model’s learning dynamics: Random Forest depicts a relatively low overall error with more stable convergence compared to Gradient Boosting Machine, showing a growing error from overfitting at the later iterations. "> Figure 10
Feature importance plot for Random Forest. "> Figure 11
Variable importance in the Gradient Boosting Model (GBM). The x-axis is the relative influence of each variable on a scale from 0 to 100%. The higher the value, the more significantly the variable contributes to the model’s predictions. In this figure, ‘forest_fire_alarms’ is the most important variable, contributing almost entirely to the prediction, while ‘MEAN_ELEVATION’ shows negligible influence. "> Figure 12
Residuals (predicted—actual burned area) for the Random Forest Model, illustrating that the x-axis shows the actual burned area (in hectares), while the y-axis indicates the residuals. The point that is closer to the red dashed line at 0 indicates better predictions. Negative residuals suggest the model tends to underestimate, especially for larger burned areas. Additional analysis suggests potential bias in the model’s handling of larger fire events. "> Figure 13
Residuals (predicted—actual burned area) for the Gradient Boosting Model (GBM), showing the actual burned area (in hectares) on the x-axis, while residuals are on the y-axis. Points closer to the red dashed line suggest better predictions. Larger residuals, particularly for larger burned areas, suggest the model underestimates burn areas, similar to the Random Forest model. This pattern reflects the challenges of accurately predicting extreme fire events using the GBM. "> Figure 14
Partial Dependence Plot (PDP) for the Log of Burned Area in the Random Forest Model, showing the log-transformed burned area (in hectares) on the x-axis, while indicating the predicted burned area on the y-axis. The plot suggests that for low log burned area values (below 3), the predicted area remains stable at around 3000 hectares. However, predictions increase sharply when the log burned area exceeds, highlighting that the model predicts significantly higher burned areas for larger fires, suggesting the model’s sensitivity to extreme fire events. "> Figure 15
Partial Dependence Plot (PDP) for the Log of Burned Area in the Gradient Boosting Model (GBM). "> Figure 16
Correlation heatmap of fire metrics, including meteorological and environmental factors. It shows the interrelation between different fire metrics and the environmental parameters of precipitation, temperature, and relative humidity. ">

Versions Notes

Abstract

Forest fires pose a significant ecological threat, particularly in the Diamer District, Gilgit-Baltistan, Pakistan, where climatic factors combined with human activities have resulted in severe fire incidents. The present study sought to investigate the correlation between the incidence of forest fires and critical meteorological elements, including temperature, humidity, precipitation, and wind speed, over a period of 25 years, from 1998 to 2023. We analyzed 169 recorded fire events, collectively burning approximately 109,400 hectares of forest land. Employing sophisticated machine learning algorithms, Random Forest (RF), and Gradient Boosting Machine (GBM) revealed that temperature and relative humidity during the critical fire season, which spans May through July, are key factors influencing fire activity. Conversely, wind speed was found to have a negligible impact. The RF model demonstrated superior predictive accuracy compared to the GBM model, achieving an RMSE of 5803.69 and accounting for 49.47% of the variance in the burned area. This study presents a novel methodology for predictive fire risk modeling under climate change scenarios in the region, offering significant insights into fire management strategies. Our results underscore the necessity for real-time early warning systems and adaptive management strategies to mitigate the frequency and intensity of escalating forest fires driven by climate change.

Keywords:

forest fires; climate factors; Random Forest; Gradient Boosting Machine; Gilgit-Baltistan

1. Introduction

Forest ecosystems are crucial in preserving biodiversity by absorbing carbon dioxide and releasing oxygen. They provide essential resources such as timber, fuelwood, fodder, and non-timber forest products, support various ecological processes, and contribute to mitigating the impacts of climate change [1,2,3,4]. Pakistan’s rural economy and environmental health are heavily dependent on forest resources. Nevertheless, a combination of natural factors and human activities renders these areas particularly susceptible to forest fires which have become the major causes of deforestation across the globe [2,5].

The forestry sector, renowned for its considerable natural capital value and resilience against climatic threats, has significantly declined over the past two decades. Forest fires inflict direct harm by annihilating plant life and disrupting carbon cycles and local ecological equilibrium, posing indirect threats to the biosphere [6,7,8]. Factors such as fuel accumulation and drought conditions exacerbate fire risks by enhancing the combustibility of forest biomass, which may ignite fires for various reasons [9,10]. Globally, the scale of destruction is alarming, with over 200,000 forest fires annually resulting in the loss of 4.7 million hectares of forests since 2010. These fires not only threaten ecological stability but also have far-reaching socio-economic and climatic consequences [6,11].

The increase in forest fire incidents correlated with temperature and late spring, conditions further aggravated by global climate change. The Intergovernmental Panel on Climate Change (IPCC) forecasts that the number of fires and their strength will increase more and more in the 21st century, posing a challenge for global forest management [12,13,14,15]. Pakistan, despite its low ranking on the worldwide Greenhouse Gases (GHS) index, is among the top seven countries most affected by climate change that amplifies forest fire occurrences [1]. The degradation of forest cover in Pakistan, particularly in areas such as the Northern Mountains and Gilgit Baltistan, is compounded by human encroachment, agricultural expansion, and climate variations. These factors lead to significant ecological, economic, and social repercussions [5,16].

There is indeed a vital lack of knowledge regarding how climatic elements work together with the development of forest fire in the dry temperate forests of Gilgit-Baltistan, especially in the Diamer District. The area, with its undeniable potential for environmental imbalance and its importance of being a place of livelihood, has not yet been a subject of comprehensive study; however, it has shown a high degree of possible impacts from deforestation and climate change [1]. By utilizing advanced predictive modeling techniques encompassing Random Forest (RF) and Gradient Boosting Machine (GBM), our research aims to assess the impact of climatic factors such as temperature, humidity, rainfall, and wind speed on forest fire patterns over the period from 1998 to 2023. The predictive forest fire models based on machine learning have some unique benefits when compared to their statistical model counterparts. These techniques are more resilient to complex data sets and give better accuracy in forecasting, hence can be utilized for the examination of the complicated behavior of forest fires under the influence of different atmospheric conditions [17,18].

By combining RF and GBM models into our methodology, the study here presents a novel technique for predicting and analyzing forest fire risks, especially in areas undergoing both the effects of climate change and environmental deterioration. We aim to refine fire management strategies and provide applicable insights into mitigating fire-induced damage in forested landscapes. Moreover, this study also has a potential impact on the global discourse on climate resilience, sustainable forest management, and the development of early warning systems to deal with the increasing risks imposed by forest fires.

2. Materials and Methods

2.1. Study Area

Gilgit-Baltistan is situated in northern Pakistan, at the intersection of Central Asia, China, and South Asia as shown in Figure 1. It is nestled among the towering ranges of the Himalayas, Karakoram, Pamir, and Hindukush mountains [1,4].

The dry temperate forests of the Diamer District in Gilgit-Baltistan are a vital natural resource, constituting approximately 30% of the region’s total forested area https://fwegb.gov.pk/forest (accessed on 18 October 2023). These forests are dominated by coniferous species such as Pinus wallichiana (blue pine), Cedrus deodara (deodar), Picea smithiana (spruce), and Abies pindrow (fir), which play significant ecological and socioeconomic roles [19]. Primarily located on the southern slopes of the Himalayan range, these forests exhibit altitudinal zonation, with pine species prevalent at lower elevations and fir and spruce species thriving at higher elevations [19]. The region’s coordinates are 35°23′07″ N and 74°34′07″ E, covering a forested area of approximately 0.1 million hectares [1]. The entire region covers an area of 72,971 km² (28,174 sq. mi) [20]. The 2017 census reported that Gilgit-Baltistan had a population of approximately 1.492 million, with an average annual growth rate of 2.87% (Districts & Places—Population Statistics, Maps, Charts, Weather and Web Information). Administratively, Gilgit-Baltistan is organized into three divisions: Diamer, Gilgit, and Baltistan, with Gilgit serving as the administrative capital [4]. Chilas serves as the administrative center of the Diamer District. The district delineated by Tangir to the north, the Gilgit District to the east, the Mansehra district of Pakistan’s Khyber Pakhtunkhwa Province and the Neelum District of Azad Kashmir to the south, and the Upper Kohistan District of Khyber Pakhtunkhwa Province to the west [4]. The Karakoram Highway also traverses the Diamer District.

The region (Diamer District) predominantly endures winter for around eight to nine months annually. However, Pakistan’s climate consists of four distinct seasons: from December to February, there is a mild, dry winter; The region has a cold, dry winter with average temperatures around 4 °C, which, although not extremely cold, can still cause significant drying of vegetation. These conditions, combined with low precipitation, lead to the accumulation of dry fuel, which can easily ignite during the next fire season. March to May is characterized by a hot, dry spring; the summer season, which is influenced by the southwest monsoon, spans from June to August and is marked by rain; and finally, the monsoon recedes from September to November [14].

The study area experiences a high peak of precipitation during May and June, while July is characterized by high average monthly temperatures. These conditions create a complex interplay between moisture and temperature that can influence fire activity according to local variations in vegetation and soil moisture levels. However, fire dynamics alone may not be explained simply by precipitation alone without regard to fuel load, type of vegetation, and other anthropogenic influences [21].

2.2. Data Sources

Our research encompassed a comprehensive array of data sources that include forest fires, meteorological and elevation data, as well as hill shade, slope, and aspect information. Additionally, we incorporated geographic data, forest fire logs, and data pertaining to burned areas. Table 1 summarizes each data type, including detailed features and entries.

2.2.1. Meteorological Data

The meteorological data utilized in our research was sourced from the meteorological department of Gilgit Baltistan and NASA’s official website. The data was formatted in CSV to facilitate its integration into the final fire calculations (accessed on 14 March 2024) [22]. This dataset encompassed complete data sets (from year 1998 to 2023) of average temperature, relative humidity, wind speed, and precipitation for each year.

Meteorological data, including temperature, relative humidity, rainfall, and wind speed, were obtained from (https://power.larc.nasa.gov, accessed on 14 March 2024) for the study area covering the Diamer District. Data were collected and processed to ensure accuracy by screening for missing values and anomalies. Monthly and annual averages were calculated for temperature and relative humidity, while total rainfall and average wind speeds were aggregated for the same intervals to analyze seasonal and temporal trends. These processed data were then used to assess correlations with other variables of interest in this study.

2.2.2. The DEM, Hill Shape, Aspect Value, and Slope Value

Figure 2 above depicts the Digital Elevation Model (DEM) values. The current dataset was sourced from NASA’s release of the complete level 3, arc-second Shuttle Radar Topography Mission (SRTM) data [23]. While the original dataset contained data gaps, particularly in regions with insufficient radar contrast, such as water bodies, snow-covered areas, and mountainous terrains, it was utilized in this study. Understanding how various altitudes affect the occurrence and behavior of forest fires in areas prone to such hazards is critical. It also contains Hill shade values, aspect values, and slope values. These topographic parameters are utilized to analyze terrain features and their environmental impacts. Hill shade simulates sunlight and shadow effects on landscapes, helping researchers study the influence of terrain on vegetation, water flow, and solar energy potential. Aspect indicates the direction a slope faces, which is critical for understanding microclimatic conditions, soil moisture distribution, and vegetation patterns. The slopes measure the terrain’s steepness and assess erosion risks, landslide susceptibility, and habitat suitability. Collectively, these values offer a comprehensive insight into the role of terrain in ecological and environmental processes (Figure 2) [24].

2.3. Forest Fires and Climate Data

The meteorological data was compiled from NASA’s official websites and Pakistan’s Meteorological Data Department, spanning the years 1998 to 2023. This dataset offers daily records of temperature, precipitation, maximum wind speed, and minimum relative humidity for each station. Before modeling, we conducted exploratory data analysis (EDA) on these variables and looked for patterns, anomalies, and correlations. The significance of relationships between climatic variables and forest fire incidents was the grounds on which the hypotheses were created. One of the main ones was that temperature and relative humidity had a direct effect on the burned area while precipitation and wind speed were not so important. The hypothesis testing used were the Pearson correlation and the analysis of variance (ANOVA) methods.

Shu Zhan and his colleagues analyzed forest fire data from the Ta He Forestry Bureau (1974–2004) to discern climate-related patterns and correlations. Inspired by the framework established by Shu et al. [25], we classified forest fires into four distinct categories according to the amount of land affected: forest fire alarm: burned area under 1 hectare; common forest fires: damage regions ranging from 1 to 100 hectares; big forest fires: encompass 100 to 1000 hectares; and enormous forest fires: burn more than 1000 hectares. These classifications were verified with historical fire patterns by using descriptive statistics which proved consistency among the datasets. All collected data was processed using Microsoft Office Excel 2022.

2.4. Data Limitations and Contextual Variables

This study utilizes climatic and topographical variables, including temperature, precipitation, relative humidity, wind speed, slope, and elevation. While these factors provide critical insights into the predisposition for fire occurrences, they do not directly account for starting factors such as ignition sources or fuel availability. The absence of data on vegetation type, biomass density, soil moisture, and anthropogenic activities is a limitation that restricts a more holistic understanding of fire dynamics. Future studies should aim to incorporate these critical variables for a more robust analysis of fire behavior.

2.5. Modeling Approach

We used two distinct models for analysis, those are the random forest (RF) and the Gradient Boosting Machine (GBM), to provide new and refined analytical approaches to studying forest fire dynamics in the Diamer District. These models were selected because they were operative in outcome patterns and important variables affecting the intensity of forest fires, and they were robust enough to handle intricate, multivariate datasets [26].

2.5.1. Dataset Preparation

The annual records from 1998 to 2023, including 26 observations and 13 variables, were encompassed in the dataset. The variables involved topographical features like slope and elevation, meteorological characteristics like annual temperature, precipitation, and wind speed, and different types of forest fire occurrences like major, general, and enormous fires. Preprocessing involved handling missing values using interpolation and detecting outliers in an attempt to obtain quality data. The transformation of the target variable, burned area (ha), involved logarithm distribution normalization. Normalization enabled feature selection and training of the model to involve comparable variables. Prior to the model application, the dataset was randomly partitioned into a training subset (70%) and a testing subset (30%) using a splitting technique. This partitioning strategy maintained a separate test set for unbiased model evaluation while ensuring that the models were trained on a representative portion of the data [27].

2.5.2. Random Forest Model

A popular machine learning algorithm for a variety of classification and regression tasks is the Random Forest model. This approach makes it easy to build up many decision trees during data training. The model usually operates on the classes (classification) or predictions based on the output of the individual tree. We set up 300 trees for the RF model, which is adequate for the calculation time taken in building the models versus the accuracy. The technique for building each tree was random so that four variables were used for any split in the tree. The Gini index was the measure of feature importance applied in determining the relevant factors for forest fire behavior. The hyper-parameters for the RF model, such as the number of trees and maximum depth, were optimized for minimizing overfitting and improvement in performance, and the robustness of the model was found through out-of-bag error estimation and evaluation metrics like Root Mean Squared Error (

R M S E

) and R-squared (R²) [26,27,28]. The formula for a Random Forest model in regression can be expressed as (Equation (1)):

\hat{f} (x) = \frac{1}{n} \sum_{i = 1}^{n} T_{i} (x)

(1)

where:

$\hat{f} (x)$ indicates the prediction of the Random Forest for input x,
n represents the number of trees in the forest
$T_{i} (x)$ indicates the i-th tree for the input x.

The performance of the RFM was evaluated using Root Mean Squared Error (

R M S E

) and R-squared (R²) (Equation (2)):

R M S E = \sqrt{\frac{1}{N} \sum_{i = 1}^{N} {({\tilde{y}}_{i} - {\hat{y}}_{i})}^{2}}

(2)

where:

${\tilde{y}}_{i}$ represents observed value
${\hat{y}}_{i}$ represents predicted value
$N$ represent the number of observations.

Based on the Gini Index or mean decrease in accuracy, the model also uses feature importance, which identifies the contribution of each variable.

2.5.3. Model Validation for Random Forest

Out-of-bag (OOB) error estimation was employed to assess the RF model’s performance, which provides a robust, reliable, and unbiased prediction of the model’s errors. In addition, the model was evaluated through Root Mean Square Error (

R M S E

) fitting for the test set.

2.5.4. Gradient Boosting Machine Model (GBM)

The Gradient Boosting Model is a machine learning algorithm that is sequentially employed to construct subsequent models, each one rectifying the errors of its predecessor. For the GBM model, 300 trees were set, a maximum of three levels per tree, and a learning rate (shrinkage) set at 0.01 to avoid overfitting. A three-fold cross-validation technique was utilized to ascertain the optimal number of trees required for generalizing the model more effectively. Due to the small size, only certain model parameters, such as the minimum number of observations per node and subsampling rates, were set to avoid overfitting. Three-fold cross-validation was conducted to tune hyperparameters such as the number of trees, learning rate, and minimum observations per node. This involved minimizing overfitting and the possibility of the model not generalizing well to unseen data. Feature importance analysis was performed to assess the relative contributions of climatic variables, such as temperature and rainfall, toward the occurrence and intensity of fires [29]. The prediction function

F_{m} (x)

is iteratively updated using the following Equation (3):

F_{m} (x) = F_{m - 1} (x) + V \cdot h_{m} (x)

(3)

where:

$F_{m} (x)$ is the prediction function at the mmm-th iteration
$V$ is the learning rate,
$h_{m} (x)$ is the prediction of the weak learner (often a decision tree) at iteration mmm.

2.5.5. Statistical Validation for Models

The predictive performance of both models was checked by implementing the

R M S E

and R² metrics. The accuracy of prediction by both models for different burnt area sizes was evaluated using residual analysis.

Additionally, we performed hypothesis testing on the model outputs to test for the existence of significant patterns. Statistical tests were performed, including t-tests for mean difference and chi-square tests for categorical variables for testing the robustness of the predictions by the models. The model shows the correlation matrix and partial dependence plots (PDPs) to visualize and interpret interactions between independent variables and fire dynamics.

2.5.6. Statistical Analysis

To further strengthen the statistical strength of this study, further analyses were carried out to cross-validate the findings of the machine learning models and examine the interrelation between climatic variables and forest fire dynamics. ANOVA was performed to compare burned areas across critical fire months (May, June, and July), which revealed significant seasonal variations in fire activity. Post hoc tests revealed that July was always significantly larger than May and June. Pearson correlation analysis was used to estimate the strength of the relationships between climatic variables: temperature, humidity, precipitation, wind speed, and burned area. The results show a positive correlation between burned area and temperature, while relative humidity and precipitation had negative correlations. These results highlight the role of temperature and humidity as the primary factors controlling forest fire activity in the regions of Shu, Di, and Huang [25].

Furthermore, t-tests were conducted to compare the distribution of burned areas under extreme climatic conditions, such as high versus low temperatures, which showed statistically significant differences. Regression diagnostics, which included residual analysis, ensured the robustness of the Random Forest and Gradient Boosting Machine models, demonstrating the ability to detect complex interactions between climatic and topographical variables without introducing a significant bias. Stepwise multivariate regression further validated these results, pointing to temperature and relative humidity as the most influential predictors of burned area, explaining over 50% of the variance in fire activity. These complementary statistical tests reinforce the outputs of the machine learning model and further help in understanding the interplay between climatic factors and forest fire dynamics, thus enhancing the precision and reliability of the conclusions of this study [26,27].

3. Results

3.1. Burned Area Trends over Time

Between 1998 and 2023, 169 fires were reported in the Diamer District. The total burned area amounted to approximately 109,401.325 hectares, with an average of 4207.74 hectares per year. Given the significant variation in the size of each forest fire, their log numbers are calculated to represent the extent of each fire’s impact on the forested area (see Figure 3a). It is noted that the smallest number was 6 hectares in 2000, whereas the largest was an impressive 1.9 × 10⁴ hectares in 2020 (see Figure 3b). Notably, 2017, 2018, 2019, and 2020 logged substantially higher burned areas than earlier years, which is a clear indicator of enhanced fire severity. This is in line with the global trends where fire frequency and intensity have been shown to increase gradually in many locations, along with climate change. The outcome is that the fire cycles in Diamer District are taking three to four years, which is the most important window for targeted management interventions. From 1998 to 2023, 169 forest fires occurred primarily in May, June, and July, accounting for 5.2 × 10⁴ ha, 2.82 × 10⁴ ha, and 2.14 × 10⁴ ha, respectively. This seasonal clustering elevates the strong impact of climatic conditions, primarily temperature and humidity, during peak fire months. These results strongly emphasize the necessity of resource and preventive deployments for this high-risk period in the Diamer District.

3.2. Periodic Analysis of Forest Fire Events

The Julian date system has been utilized to ascertain the dates of forest fires in the Diamer District. This approach segments the year into days numbered 1 through 365 and assigns a distinct identifier to each. The detailed Julian dates for both forest fires and major forest fires (those larger than 60 hectares) that occurred in the Diamer District over 25 years are summarized (see Figure 4a,b). Such changes may indicate an extended fire season, which might be related to other changes in climatic conditions, such as earlier springs and prolonged dry periods, characteristic of similar ecosystems around the world. This finding carries major implications for planning into fire prevention, indicating an adaptive strategy for managing extended fire seasons.

The Y-axis is scaled from 1 to 365 days in accordance with the format given below. This study reveals that each dot on the chart represents a specific forest fire that occurred between 19 February and 16 December. Over a 25-year span, forest fires predominantly occurred between Julian days 102 to 181 and 240 to 293, which correspond to the calendar periods 12 April to 30 June and 28 August to 20 October. These intervals correspond to the Diamer District’s spring and fall seasons, with a higher frequency of fires observed in the spring. Since 2018, the number of forest fires has increased between the Julian dates of 186 and 244, which corresponds to the period from 5 July to 1 September. Notably, 1998 and 2020 witnessed a huge concentration of fires, with 2020 being particularly striking as it saw over 20 forest fires within a mere 39-day period in the summer.

The given plot in (Figure 5) illustrates the Julian date statistics for the earliest and latest forest fires recorded from 1998 to 2023. The upper set of points represents the Julian dates of the newest forest fires, while the lower set depicts the Julian dates of the earliest forest fires. Over time, the trend lines show a modest increase in the earliest fire dates and a gradual decrease in the latest fire dates. Specifically, the regression equation for the earliest forest fires is y = 2.0718x − 4054.6 with an R² value of 0.2241, suggesting a modest upward trend. This indicates that the dates of the earliest forest fires have gradually occurred later in the year. On the other hand, the regression equation for the latest fires is y = −3.3414x + 6969.5, with an R² value of 0.1223, showing a very slight downward trend with minimal significance.

The analysis indicates that the annual onset of the first fire tends to occur marginally later, while the latest fires occur slightly earlier. This trend suggests that the situation regarding fire prevention during the summer months has become progressively more critical, with fires now capable of occurring over a wider span of dates.

3.3. Correlation of Forest Fires with Climate Variables

In Gilgit-Baltistan, winter temperatures in the range of 4 °C are cold but not freezing. Such conditions, though not extreme, lead to the drying of vegetation and the accumulation of dry fuel, which may influence fire activity. Large-scale forest fires were more common when temperatures were between 5 and 15 degrees Celsius. This rise was probably caused by meadows, grasslands, and dry, exposed forest regions that were more prone to burning. As the temperature escalated, vegetation became increasingly parched, resulting in the depletion of moisture and heightened flammability of the falling leaves and other combustible materials. Even a tiny spark may start a fire in these circumstances, and it would spread swiftly. However, the likelihood of forest fires tended to decline during extended hot weather. This decrease was ascribed to plant growth, which increased moisture retention and decreased the likelihood of vegetation burning.

There was a lower likelihood of forest fires during the winter months because the forest regions were covered with snow. Nonetheless, in early June, forest fires persisted at a considerable rate despite an uptick in rainfall and a reduction in dry, windy days with diminished evaporation. Approximately 109,400 hectares of forest area were affected by 169 forest fires annually on average during the 25 years. Larger burned areas, higher fire hazard ratings, and an increased frequency of fires were all observed during this time. Our 25-year analysis showed that May, June, and July experienced a higher number of forest fires almost every year, so we focused on these months to evaluate forest fire data alongside monthly average temperature, relative humidity, rainfall, and wind speed on an annual basis in the Diamer District (Figure 6).

We conducted an investigation into the correlation between forest fires and climate factors during May, June, and July from 1998 to 2023 because of the high concentration of forest fires during these three months. Our analysis reveals that the regression models and associated indices are as follows (Figure 7). The findings indicate a significant correlation (p < 0.05) between the occurrence of fires and those four different types of climate data obtained in the Diamer District.

Specifically, higher temperatures were consistently accompanied by larger burned areas, which supports the hypothesis that these are the primary drivers of fire dynamics. A much weaker influence was exerted by precipitation. While rainfall probably dampens fire spread, the amount of precipitation variability in this region is not sufficiently great to overcompensate the effects of rising temperatures. This insight fits with broader studies on dry temperate forests, where climate-induced drought conditions exacerbate fire risks.

The scatter plots (Figure 8) show the correlation between the number of forest fires in May and four environmental conditions: average temperature, average humidity, average precipitation, and average wind speed over the years. The plots for average temperature and average humidity show slightly more variation in the number of fires than average precipitation and wind speed. Moreover, the predicted values from the regression model closely follow the actual data. These visualizations suggest that average temperature, humidity, precipitation, and wind speed are not very strong predictors of forest fires and affect forest fires in a minimal sense. These results support the applicability of meteorological indicators for use in early warning systems to make more accurate fire risk predictions during critical months.

3.4. Modeling Results

3.4.1. Performance Evaluation and Model Comparison

RMSE and R-squared (R²) analyses were incorporated to evaluate the achieved performance for the RF and GBM models. The performance of the RF model was observed to be an RMSE of 5803.69, which explains only 49.47% of the variance of the burned area data. Conversely, the GBM model produced a lower RMSE of 7176.67, showing an explanation of 46.9% variance after parameter tuning and cross-validation, which was less than that of the RF model (Figure 9). Such two models were fairly accurate, though the RF model was able to perform better in capturing complex interaction among climatic and topographical variables, thus preferred for this study.

3.4.2. Feature Importance Analysis

Weather and terrain as such were the most critical variables as shown in the variable importance analysis for both models. In the feature importance plot of the RF model, which is given in Figure 10, these factors acted the most on the burned area, while in Figure 11, temperature and relative humidity had the highest importance scores, identified as the most critical drivers of burned area size. The importance scores show higher temperatures and lower humidity between May and July substantially contribute to larger burned areas. Precipitation and wind speed had relatively lower importance scores, suggesting a less direct influence on fire intensity. The feature importance plot of the GBM model elevation and temperature also contributed a lot to determining the fire intensity. This has led to the role of climatic and topographical factors in indicating fire intensity and therefore calls for a holistic approach to fire management that includes both climatic and topographical factors.

3.4.3. Residual Analysis

The anticipated accuracy of each model was verified through the creation of residual plots. As shown in Figure 12, the model’s residual plot of the RF model showed very small residuals for all values of the burned area, suggesting the model can predict even low and high fire events. On the other hand, the residual plots for the GBM model, as shown in Figure 13, had larger residual deviations mainly for higher values of burned area, which indicates the model is less effective and sensitive to extreme events. This further enhances the credibility of the RF model to use in the actual application involving fire risk prediction, mainly where unstable fire intensities prevail in regions.

3.4.4. The Partial Dependence Plots (PDPs)

The effect of the most essential features on the burned area was also represented using Partial Dependence Plots (PDPs) for each model.

For the RF model, the VP for Log_Burned_Area (Figure 14) would appropriately bring out elevation focus as a step further in the differential consideration of topography.

In the case of the GBM model, the VP for Log_Burned_Area (Figure 15) brought the effect of temperature into focus but without the clear-cut and predictable patterns within the RF model that were largely stable.

The x-axis indicates the log-transformed burned area (in hectares), while the y-axis represents the predicted burned area. The horizontal line signifies that variations in the log-transformed burned area do not notably influence the model’s predictions, indicating that the GBM model does not substantially account for this variable in forecasting burned areas. This implies a limitation in the model’s effectiveness for accurately predicting the extent of burned areas.

3.4.5. Correlation Matrix

A correlation matrix (Figure 16) was created to determine the relationship among the variables and their combined impact on the burned area. This matrix shows that precipitation, temperature, and burned areas exhibit significant correlations, providing valuable insights into climatic factors and fire intensity.

The correlation between burned area and meteorological factors is illustrated in Figure 16. Burned area is negatively correlated with precipitation (−0.38) and relative humidity (−0.42), meaning that higher moisture levels lower fire activity. A positive correlation with temperature (0.26) indicates that higher temperatures favor fire spread. These correlations are within the moderate range (±0.4 to ±0.5) but are ecologically meaningful because of the complexity of interactions influencing fire dynamics. Relative humidity sheds light on the atmospheric moisture role. The very low correlations between giant forest fires and meteorological factors might be indicative of some anthropogenic influence, including human-induced ignitions or land-use changes.

Even though the correlation coefficients are below ±0.7, it is also worth mentioning that in ecological studies, moderate correlations are significant since fire dynamics is a multifactored issue. A heatmap provides a holistic overview of how temperature, precipitation, relative humidity, and other environmental factors affect the various types of fires.

3.4.6. Model Selection

A thorough review determined that the RF model was the most suitable model for this study. In terms of interpretation of variance, RMSE, and consistency over a range of fire intensities, it performed the best at apprehending the effects of topography and environment on forest fire dynamics. By utilizing cutting-edge machine learning models to analyze forest fires in the Diamer District, this selection improves the study’s forecast accuracy and presents a fresh approach to analyzing forest fires.

3.5. Statistical Analysis Outcomes

To supplement the machine learning models, further statistical analyses were conducted to investigate the interactions between climatic variables and forest fire dynamics. ANOVA indicated that the area burned differed significantly between the critical fire months of May, June, and July (F(2, 78) = 14.23, p < 0.01). Post hoc tests found that July areas burned were statistically significantly higher in comparison to both May (p <0.05), and June (p < 0.05). This confirms it as the main fire month of the year. Pearson correlation analysis revealed a significant positive relationship of burned area with temperature (r = 0.46, p < 0.01) but negative relationships between relative humidity and precipitation (r = −0.41, p < 0.05; r = −0.22, p = 0.08, respectively). A weak association was indicated in wind speed, which seems not to play a significant role in fire activity.

Comparisons of burned area distributions under extreme temperature conditions between the two samples by two-sample t-tests were highly significant. In high-temperature periods, larger burned areas (mean = 1512 ha) compared to low-temperature periods (mean = 674 ha; t(78) = 3.87, p < 0.01) were obtained. The results of the regression diagnostics and residual analysis revealed the robustness of the machine learning models. Bias was very low, and the VIF values were acceptable, with no issues of multicollinearity. Additionally, stepwise multivariate regression validated these results indicating that the critical predictors include the temperature and the relative humidity, collectively accountable for explaining variance in the amount of area that was burnt 52.3% by those variables: Adjusted R² = 0.523, p < 0.01. Those machine learning outcomes complemented by further statistical knowledge show a great range of detailed features of the research region.

3.6. Practical and Theoretical Implications

The results can help us understand the dynamics of forest fires in the Diamer District and the drivers of these forest fires. This can be very useful for practical applications, such as recommending where to deploy resources during peak fire months and building early warning systems that incorporate meteorological data. Scientifically, this study adds to the burgeoning literature on the impacts of climate change on fire regimes in dry temperate forests. The successful application of machine learning models further cements their promise to improve the accuracy of complex ecological systems predictions.

This further points to the urgency of measures to adapt to climate, perhaps with afforestation programs for fire-resistant species, improvement of forest monitoring systems, and community-based initiatives to reduce the risks posed by forest fires. The findings further underscore the broader necessity of reducing greenhouse gas emissions that will mitigate the lingering effects of climate change on forests.

4. Discussion

4.1. Correlation Between Forest Fires and Time

Our analysis from 1998 to 2023 recorded a total of 169 forest fires. The average number of fires recorded is 6.76 per year. This shows significant fluctuations in forest fires annually, with 2020 recording a peak of 20 fires, contrasting sharply with the single incident recorded in 2000. As presented in Figure 3a, these fluctuations show that the frequency of forest fires is going up, in agreement with broader studies in the Himalayan regions and globally. The global rise in average temperatures, driven by anthropogenic climate change, has intensified heatwaves, droughts, and fire weather. Studies [30,31] link increasing heatwave frequency and intensity to human-induced climate forcing. Fire weather conditions—high temperatures, low humidity, and strong winds—have become more frequent, fueling wildfires globally [32]. These trends highlight the cascading risks of a warming climate on ecosystems and human systems. This increase in frequency is found to be associated with climatic extremes caused by global warming, due primarily to increased temperatures and less humidity [33,34]. The RF model provided insight into how these factors influence fire dynamics, aligning with the hypothesis that climatic extremes contribute to the frequency and intensity of forest fires in the region. Additionally, the RF model, leveraging its predictive prowess, was able to reflect these temporal trends more accurately than the GBM model, indicating its suitability for temporal fire prediction in diverse ecological regions like Gilgit-Baltistan [16,35].

4.2. Burned Area and Forest Fire Occurrence Patterns

The peak fire season was identified as May, June, and July. This period shows a concentration of fires, indicating a possible link between climatic conditions and human activities during these months. In addition, extreme weather events, such as heatwaves and prolonged droughts, have escalated due to altered precipitation and evaporation patterns [36]. High temperatures and reduced moisture levels in vegetation create ideal conditions for ignition and rapid fire spread, making wildfires more frequent and intense. Over time, the number of forest fires showed an increasing trend between 2015 and 2020 see Figure 2b. The cycle of years during which forest fires occurred more frequently was 4 to 5 years. As shown in Figure 4b, May represented the highest number of fires with 35 while June had 48, and July had 53. This made these months the significant fire season in the Diamer District. Such results are consistent with other similar dry temperate forests in Central Asia. Heightened fire activity during May, June, and July is attributed to the extreme temperatures in July, while the decreasing precipitation values from May to July drive the increased fire activity. Although May and June reveal peak precipitation, localized drought conditions or reduced moisture retention due to soil properties and vegetation characteristics may contribute to fire ignition and spread. It was evident that incorporating soil moisture and vegetation data in subsequent analyses would improve the interpretation of factors that influence the dynamics of fire during critical fire seasons.

While extreme temperatures are related to increased fire activity in July, the case of precipitation is more ambiguous. For example, peak precipitation values occur in May and June, but fire activity can be suppressed during these months as a result of rapid vegetation drying following rainfall, a lack of soil moisture retention, or local drought conditions. This makes it essential to include high-resolution soil moisture and vegetation data in this analysis for better resolution in the understanding of fire dynamics. A notable peak occurred in 2020 with a record number of 20 fires reported. Several factors probably contributed to this surge in fire activity. First, climatic conditions during the critical fire season (May to July) were marked by extreme temperatures and significantly reduced precipitation, creating exceptionally dry conditions that made vegetation highly flammable. This is in line with the correlations reflected in Figure 16 that indicate a negative relationship of burned area with both precipitation (−0.38) and relative humidity (−0.42), and a positive association with temperature (0.26). Third, 2020 was a year of persistent drought, which reduced further the moisture levels in both the soil and vegetation, thereby increasing the risk of fires. Third, anthropogenic factors such as deforestation, agricultural burning, and other human activities may have been major contributors to igniting fires. The weak correlations for giant forest fires with meteorological factors indicate that these fires may be human-induced ignition sources rather than caused by natural factors. The COVID-19 pandemic that occurred in 2020 led to lockdowns and restrictions resulting in reduced forest monitoring and firefighting capabilities, and this may have allowed uncontrolled fires to spread and become widespread. All of these factors—extreme weather, drought, human activities, and reduced fire management capacity can be combined to explain this unprecedented number of fires experienced in 2020 [14]. Feature importance analysis (Figure 10 and Figure 11) further confirmed the annual variation in the frequency and area burned from fires, measured in the numbers of fire incidents per year and hm² of area burned. The figure brings out the fact that, during those years when precipitations were lower and temperatures higher, fire frequency also tended to be higher. This implies that drier conditions favor higher ignition and spread of fire. The potential of the RF model to predict burned areas according to those climatic variables justifies it as a good model in practical applications in fire management [25,37].

4.3. Climate Factors and Forest Fires: Analyzing Trends and Implications

This study establishes some clear connections between climatic attributes and forest fire occurrences region, as the regression analyses show that all temperature, relative humidity, and rainfall are significantly associated with fire occurrences. Figure 7 indicates that higher temperature and lower relative humidity are the most critical drivers of burned areas being higher in size, which is in agreement with findings from studies conducted in the Malakand Division and the Indian sub-continent [12,38]. The results discussed the weaker, yet significant, role of precipitation and wind speed. Such results advance the existing body of knowledge by incorporating advanced machine learning techniques that enhance predictive capability beyond that of traditional statistical methods [39]. Although this study mainly focuses on temperature, relative humidity, and precipitation, it is important to note that thunderstorms and lightning activity can also play a significant role in forest fire ignition, especially in dry temperate forests. Lightning-induced fires are common in regions with frequent thunderstorms, and the absence of this factor in the current analysis represents an important area for further research. For instance, a study highlights Heilongjiang Province, one of China’s major forested regions, as having the highest number and concentration of lightning-caused forest fires in the country [40]. Future studies involving data on lightning strikes would give a more holistic understanding of natural ignition sources and their contribution to fire dynamics.

Although the correlations are within the moderate range of ±0.4 to ±0.5, it is ecologically meaningful in terms of fire dynamics, considering that fire dynamics involves interactions from various factors, including climate, vegetation, and human activities. Moderate correlations here can be very important to understand the relationship between them. Figure 16 shows how the environment controls fire dynamics: temperature, precipitation, and relative humidity. The burned area presents a negative correlation with the precipitation (−0.38) and relative humidity (−0.42), meaning that fire activity is suppressed with increased moisture availability. Meanwhile, the positive correlation with the temperature (0.26) implies that increased temperatures advance the ignition and spread of fire.

Curiously, these giant forest fires show poor correlations with the meteorological factors. Thereby, it has come to a point that their ignition sources may have significant contributions from anthropogenic activities; otherwise, the human-made causes of ignitions and even land-use changes can prevail. In this case, the analysis should incorporate human activity data. Although the correlations do not exceed ±0.7, their significance lies in the multifactorial nature of fire dynamics. These insights emphasize the importance of integrating climatic data and human influences when analyzing fire activity.

The impressive result of the performance of the RF model, especially in our case in analyzing the interaction of elements of topography and meteorological data and its contribution to fire management, is an asset that needs to be optimized. Figure 10 and Figure 11 highlight elevation and slope as secondary, yet essential, variables that may influence fire behavior through various changes in the type of vegetation and amount of fuel. Thus, the results drawn emphasize that successive research should consider other environmental factors, like soil moisture and vegetation cover, for improving the prediction results. For instance, regional differences have been studied for fire-driving factors in all regions of China [41]. More particularly, in Gilgit-Baltistan, forest management has to face unstable weather conditions emerging from climate change with fluctuations [25,42]. As shown in Figure 5, the extended fire season also, as illustrated by the Julian date shifts, also presents a new dimension of the risk landscape under the altered climate. In controlling the windows for this prolonged fire season, active actions call for real-time monitoring systems and localized plans for mitigation [12,13]. Extra methods that may be exploited in sophisticated machine learning methods, perhaps through neural networks, would allow for multidimensional analyses of nuanced predictions in systems of early warnings [43,44].

Through an extensive and data-driven investigation of the influence of climatic conditions on forest fire, this study sheds light on the growing problem of forest fire management. By contrasting our findings with those of previous regional studies, we highlight the current study’s worldwide relevance and the potential for more proactive and sophisticated fire management strategies globally [45,46].

5. Limitations and Future Recommendations

This study, although it is prolific, has some restrictions that will be addressed in further work. The meteorology and forest fire data are only from 1998 to 2023 and probably lack spatiotemporal resolution and coverage; this impact might affect the accuracy of any predictive work, especially in areas with larger incidents. The current study highlights the important roles of climatic factors, such as temperature, precipitation, and relative humidity, in influencing fire dynamics. However, it is essential to note that these factors act primarily as predisposing conditions rather than direct ignition sources. Critical environmental variables such as vegetation type, biomass density, soil moisture, and anthropogenic activities were not included in this study but play essential roles in initiating and sustaining fires. For example, it is the fuel from the vegetation that will ignite to propagate a fire, and soil moisture controls the flammability of the landscape. Thus, without these considerations, the conclusions from this study can only be drawn based on climatic predisposition and not as a general knowledge of fire dynamics. This study excluded a number of critical environmental factors that include, among others, vegetation type, density of biomass, fuel continuity, soil moisture, and human factors that include ignitions and changes in land use. The inclusion of such factors could significantly affect the results obtained. This study also utilizes a simplified typology for fire incidence that may not capture a comprehensive characteristic of forest fire occurrence in the region.

Another limitation of this study is that the analysis of thunderstorm and lightning activity was not carried out. These are well-known to be significant natural sources of ignition for fires. In the future, the incorporation of lightning strike data and thunderstorm occurrences may enhance the understanding of how natural sources of ignition contribute to forest fires in areas where these weather conditions are common. This may complement the anthropogenic and climatic factors being analyzed for a more holistic view of fire dynamics. Models using the RF and GBM showed robustness in terms of predictive capability, although their interpretability and performance decreased when applied in entirely new conditions in terms of either ecological or climatic conditions. Also, the results are only for dry temperate forests of the Diamer District and thus regionally applicable but not directly generalized to very different ecosystems.

Such research has the potential to integrate high-resolution remote sensing data with these types of satellite observations into improved spatial and temporal accuracy in fire modeling. Additional environmental variables like the nature of vegetation, biomass density, soil moisture, and human activities can be incorporated for better knowledge building based on the overall understanding of fire behavior. Advanced techniques may include deep learning or using models combining physical process-based simulations with data-driven approaches that would help better enhance the predictive accuracy and adaptability of models to climatic variability. A real-time early warning system would be a huge step toward proactive management. The predictive models combined with historical data and live meteorological feeds in different ecosystems and regions would authenticate the result and give an insight into the differences in ecological and climatic conditions that affect fire dynamics. The inclusion of climate change projections in the models helps predict future risks in fires, thereby aiding long-term forest management and conservation. Addressing these would provide future researchers with a much better chance to build on the insights from this study into more effective fire risk mitigation.

Future studies should include more detailed vegetation data, such as fuel load and continuity of biomass, to understand better the availability and flammability of fuel sources. Soil moisture data derived from satellite observations or field measurements could be very useful in understanding the readiness of the landscape for ignition. Integrating human activity data, such as population density, land-use patterns, and ignition sources, will allow a more comprehensive analysis of the anthropogenic influences on fire dynamics. High-resolution remote sensing data and advanced typologies for fire classification must also be integrated to make the granularity and accuracy of fire modeling efforts better.

6. Conclusions

This study about the incidences of forest fire investigates the relationship between meteorological factors within the vicinity of the Diamer District of Gilgit-Baltistan region. The data clearly shows that between 1998 and 2023, fire occurrence events were significantly impacted by the temperature, humidity, and rainfall during the critical months of May, June, and July. This analysis sheds light on the great influence of temperature and relative humidity in climatic drivers controlling the dynamics of fire. In particular, it is important to emphasize that wind velocity and precipitation have essential roles, especially when those factors are involved. Indeed, temperature and relative humidity consistently correlate with the area burned, as demonstrated by Figure 16; yet wind velocity can heavily affect fire spread and intensity during an active fire event. Similarly, precipitation with its moderate correlation is a more critical factor in maintaining vegetation moisture and reducing ignition. The interplay of such factors indicates that fire dynamics are affected by a range of temperature, relative humidity, wind velocity, and precipitation, all contributing in specific ways according to the specific fire context. The structure of the analysis suggests that an association or conjuncture of these climatic factors determines the frequency and intensity of forest fire outbreaks. The Random Forest (RF) model that we employed is capable of tracing such interrelations between fire occurrences and climate change factors. Such performance speaks to the practical applicability of this model to predict fire risk based on the intricate relationships of topographic and climatic variables underlined in the present research, especially in regions of rapid climate change.

As the findings suggest, there is a need for better-targeted fire prevention measures to minimize the socio-economic and ecological impacts of forest fires, especially during peak fire periods. Over the last few years, climate change has increased the risk of fires, making them a more severe threat to the forest cover and biodiversity in the regions. It thus calls for urgency in the need to adopt adaptive management strategies, such as real-time early warning systems and resource allocation targeting risks, in order to effectively mitigate fire risks. In this regard, through RF model analysis, we have also been able to establish that temperature rise and fluctuating humidity, which are arguably some climate change factors, have a direct relationship with fire risk potentials. Implementation of these results would inform enhanced early warning systems, optimize the use of resources, and ensure better strategies for mitigating fire outbreaks. Further integrations of more environmental and human activity variables in such models will further improve the accuracy of the predictions and applications of modeling in ensuring a more sustainable approach toward fire management in such vulnerable regions.

Author Contributions

Conceptualization, K.A. and Z.S.; data curation, F.S. and H.A.; methodology, Z.S.; resources, H.H.; software, A.A.S.; funding acquisition, Z.S.; writing—original draft preparation, K.A.; writing—review and editing, F.W., Z.S. and H.A. All authors have read and agreed to the published version of the manuscript.

Funding

This research was funded by “Foreign Expert Key Support Project (Northeast Special Project), Northeast Asia Biodiversity Survey and Monitoring (D20240164)”.

Data Availability Statement

The original contributions presented in this study are included in the article. Further inquiries can be directed to the corresponding authors.

Acknowledgments

The authors would like to express their sincere gratitude to the College of Forestry, Northeast Forestry University, Harbin, China, for their financial support and the provision of necessary facilities. Finally, we would like to acknowledge the valuable and constructive comments provided by the anonymous reviewers, which have greatly improved the quality of this manuscript.

Conflicts of Interest

The authors declare no conflicts of interest.

References

Ali, E.; Azhar, M.F.; Alam, E.; Rehman, Z.; Ullah, S.; Ahmad, A.; Towfiqul Islam, A.R.M.; Zaman, W.; Javed, M.; Mittal, P. Deforestation perspectives of dry temperate forests: Main drivers and possible strategies. Front. Environ. Sci. 2023, 11, 1151320. [Google Scholar] [CrossRef]
Lin, B.; Ullah, S. Evaluating forest depletion and structural change effects on environmental sustainability in Pakistan: Through the lens of the load capacity factor. J. Environ. Manag. 2024, 353, 120174. [Google Scholar] [CrossRef]
Jodhani, K.H.; Patel, H.; Soni, U.; Patel, R.; Valodara, B.; Gupta, N.; Patel, A.; Omar, P.j. Assessment of forest fire severity and land surface temperature using Google Earth Engine: A case study of Gujarat State, India. Fire Ecol. 2024, 20, 23. [Google Scholar] [CrossRef]
Abbas, A.; Mubeen, M.; Younus, W.; Shakeel, Q.; Iftikhar, Y.; Bashir, S.; Zeshan, M.A.; Hussain, A. Plant Diseases and Pests, Growing Threats to Food Security of Gilgit-Baltistan, Pakistan. Sarhad J. Agric. 2023, 39, 510–520. [Google Scholar] [CrossRef]
Nafees, M.; Rashid, W.; Sultan, H.; Khan, N.H.; Khurshid, M.; Ali, W.; Bohnett, E. Occurrence, probable causes, and management of forest wildfires in the Northern Highlands of Pakistan. Environ. Chall. 2024, 15, 100930. [Google Scholar] [CrossRef]
Liu, M.; Jia, D. Application of remote sensing technology in forest fire fighting. City Disaster Reduct 2018, 6, 66–70. [Google Scholar]
Santoro, A.; Venturi, M.; Piras, F.; Fiore, B.; Corrieri, F.; Agnoletti, M. Forest area changes in Cinque Terre National Park in the last 80 years. Consequences on landslides and forest fire risks. Land 2021, 10, 293. [Google Scholar] [CrossRef]
Liang, H.; Zheng, C.; Liu, X.; Tian, Y.; Zhang, J.; Cui, W. Super-resolution reconstruction of remote sensing data based on multiple satellite sources for forest fire smoke segmentation. Remote Sens. 2023, 15, 4180. [Google Scholar] [CrossRef]
Ying, L.; Han, J.; Du, Y.; Shen, Z. Forest fire characteristics in China: Spatial patterns and determinants with thresholds. For. Ecol. Manag. 2018, 424, 345–354. [Google Scholar] [CrossRef]
Yang, X.; Hua, Z.; Zhang, L.; Fan, X.; Zhang, F.; Ye, Q.; Fu, L. Preferred vector machine for forest fire detection. Pattern Recognit. 2023, 143, 109722. [Google Scholar] [CrossRef]
Gong, A.; Huang, Z.; Liu, L.; Yang, Y.; Ba, W.; Wang, H. Development of an Index for Forest Fire Risk Assessment Considering Hazard Factors and the Hazard-Formative Environment. Remote Sens. 2023, 15, 5077. [Google Scholar] [CrossRef]
Bowman, D.M.; Balch, J.; Artaxo, P.; Bond, W.J.; Cochrane, M.A.; D’antonio, C.M.; DeFries, R.; Johnston, F.H.; Keeley, J.E.; Krawchuk, M.A. The human dimension of fire regimes on earth. J. Biogeogr. 2011, 38, 2223–2236. [Google Scholar] [CrossRef]
Israr, A.; Khan, S.M.; Abdullah, A.; Ejaz, U.; Jehangir, S.; Ahmad, Z.; Hashem, A.; Avila-Quezada, G.D.; Abd_Allah, E.F. Fire-Induced Vegetation Dynamics: An In-Depth Discourse on Revealing Ecological Transformations of the Mahaban and Surrounding Forests. Fire 2024, 7, 27. [Google Scholar] [CrossRef]
Rafaqat, W.; Iqbal, M.; Kanwal, R.; Song, W. Study of driving factors using machine learning to determine the effect of topography, climate, and fuel on wildfire in pakistan. Remote Sens. 2022, 14, 1918. [Google Scholar] [CrossRef]
Wasserman, T.N.; Mueller, S.E. Climate influences on future fire severity: A synthesis of climate-fire interactions and impacts on fire regimes, high-severity fire, and forests in the western United States. Fire Ecol. 2023, 19, 43. [Google Scholar] [CrossRef]
Qamer, F.M.; Shehzad, K.; Abbas, S.; Murthy, M.; Xi, C.; Gilani, H.; Bajracharya, B. Mapping deforestation and forest degradation patterns in western Himalaya, Pakistan. Remote Sens. 2016, 8, 385. [Google Scholar] [CrossRef]
Sun, Y.; Zhang, F.; Lin, H.; Xu, S. A forest fire susceptibility modeling approach based on Light Gradient Boosting Machine algorithm. Remote Sens. 2022, 14, 4362. [Google Scholar] [CrossRef]
Li, X.; Jia, H.; Wang, L. Remote Sensing Monitoring of Drought in Southwest China Using Random Forest and eXtreme Gradient Boosting Methods. Remote Sens. 2023, 15, 4840. [Google Scholar] [CrossRef]
Raqeeb, A.; Nizami, S.M.; Saleem, A.; Hanif, M. Characteristics and growing stocks volume of forest stand in dry temperate forest of Chilas Gilgit-Baltistan. Open J. For. 2014, 2014, 44948. [Google Scholar] [CrossRef]
Anwar, S.; Khan, F.A. Impact of Karakoram highway on land use and agricultural development of Gilgit-Baltistan, Pakistan. Sarhad J. Agric. 2019, 35, 417–431. [Google Scholar] [CrossRef]
Joshi, S.; Jasra, W.; Ismail, M.; Shrestha, R.; Yi, S.; Wu, N. Herders’ perceptions of and responses to climate change in Northern Pakistan. Environ. Manag. 2013, 52, 639–648. [Google Scholar] [CrossRef] [PubMed]
Meteorological Department. Meteorological Department: Government of Pakistan. Available online: https://www.pmd.gov.pk (accessed on 19 November 2024).
Shivhare, N.; Omar, P.J.; Gupta, N.; Dikshit, P.S. Runoff estimation of Banaras Hindu University South Campus using ArcGIS and HecGeo-HMS. In Proceedings of the 2016 3rd International Conference on Recent Advances in Information Technology (RAIT), Dhanbad, India, 3–5 March 2016; pp. 607–612. [Google Scholar]
Jodhani, K.; Bansal, P.; Jain, P. Shoreline change and rate analysis of Gulf of Khambhat using Satellite images. In Advances in Water Resources and Transportation Engineering: Select Proceedings of TRACE 2020; Springer: Singapore, 2021; pp. 151–170. [Google Scholar]
Shu, Z.; Di, X.Y.; Huang, H. Effects to forest fire occurrence of climate change in Ta He Forestry Bureau in Great Xing’an Mountain. Adv. Mater. Res. 2011, 183, 135–139. [Google Scholar] [CrossRef]
Muhammad, S.; Mehmood, K.; Anees, S.; Tayyab, M.; Rabbi, F.; Hussain, K.; Rahman, H.; Hayat, M.; Khan, U. Assessment of regeneration response of silver fir (abies pindrow) to slope, aspect, and altitude in miandam area in district swat, Khyber-Pakhtunkhwa, Pakistan. Int. J. For. Sci. 2023, 3, 246–252. [Google Scholar]
Razavi-Termeh, S.V.; Sadeghi-Niaraki, A.; Choi, S.-M. Ubiquitous GIS-based forest fire susceptibility mapping using artificial intelligence methods. Remote Sens. 2020, 12, 1689. [Google Scholar] [CrossRef]
Souane, A.A.; Khurram, A.; Huang, H.; Shu, Z.; Feng, S.; Belgherbi, B.; Wu, Z. Utilizing Machine Learning and Geospatial Techniques to Evaluate Post-Fire Vegetation Recovery in Mediterranean Forest Ecosystem: Tenira, Algeria. Forests 2025, 16, 53. [Google Scholar] [CrossRef]
Guo, F.; Wang, G.; Su, Z.; Liang, H.; Wang, W.; Lin, F.; Liu, A. What drives forest fire in Fujian, China? Evidence from logistic regression and Random Forests. Int. J. Wildland Fire 2016, 25, 505–519. [Google Scholar] [CrossRef]
Perkins-Kirkpatrick, S.E.; Lewis, S.C. Increasing trends in regional heatwaves. Nat. Commun. 2020, 11, 3357. [Google Scholar] [CrossRef]
Masson-Delmotte, V.; Zhai, P.; Pirani, S.; Connors, C.; Péan, S.; Berger, N.; Caud, Y.; Chen, L.; Goldfarb, M.; Scheel Monteiro, P.M. IPCC, 2021: Summary for Policymakers. In Climate Change 2021: The Physical Science Basis. Contribution of Working Group I to the Sixth Assessment Report of the Intergovernmental Panel on Climate Change; Cambridge University Press: Cambridge, UK; New York, NY, USA, 2021. [Google Scholar] [CrossRef]
Jolly, W.M.; Cochrane, M.A.; Freeborn, P.H.; Holden, Z.A.; Brown, T.J.; Williamson, G.J.; Bowman, D.M.J.S. Climate-induced variations in global wildfire danger from 1979 to 2013. Nat. Commun. 2015, 6, 7537. [Google Scholar] [CrossRef] [PubMed]
Lecina-Diaz, J.; Martínez-Vilalta, J.; Alvarez, A.; Banqué, M.; Birkmann, J.; Feldmeyer, D.; Vayreda, J.; Retana, J. Characterizing forest vulnerability and risk to climate-change hazards. Front. Ecol. Environ. 2021, 19, 126–133. [Google Scholar] [CrossRef]
McWethy, D.B.; Pauchard, A.; García, R.A.; Holz, A.; González, M.E.; Veblen, T.T.; Stahl, J.; Currey, B. Landscape drivers of recent fire activity (2001–2017) in south-central Chile. PLoS ONE 2018, 13, e0201195. [Google Scholar] [CrossRef]
Urrutia-Jalabert, R.; González, M.E.; González-Reyes, Á.; Lara, A.; Garreaud, R. Climate variability and forest fires in central and south-central Chile. Ecosphere 2018, 9, e02171. [Google Scholar] [CrossRef]
Dai, A. Increasing drought under global warming in observations and models. Nat. Clim. Chang. 2013, 3, 52–58. [Google Scholar] [CrossRef]
Wu, Z.; Hasham, A.; Zhang, T.; Gu, Y.; Lu, B.; Sun, H.; Shu, Z. Analysis of PM2.5 Concentration Released from Forest Combustion in Liangshui National Natural Reserve, China. Fire 2024, 7, 311. [Google Scholar] [CrossRef]
Barik, A.; Baidya Roy, S. Climate change strongly affects future fire weather danger in Indian forests. Commun. Earth Environ. 2023, 4, 452. [Google Scholar] [CrossRef]
Omar, P.J.; Rai, S.P.; Tiwari, H. Study of morphological changes and socio-economic impact assessment: A case study of Koshi River. Arab. J. Geosci. 2022, 15, 1426. [Google Scholar] [CrossRef]
Jiao, Q.; Fan, M.; Tao, J.; Wang, W.; Liu, D.; Wang, P. Forest fire patterns and lightning-caused forest fire detection in Heilongjiang Province of China using satellite data. Fire 2023, 6, 166. [Google Scholar] [CrossRef]
Ma, W.; Feng, Z.; Cheng, Z.; Chen, S.; Wang, F. Identifying forest fire driving factors and related impacts in china using random forest algorithm. Forests 2020, 11, 507. [Google Scholar] [CrossRef]
Gupta, N.; Patel, J.; Gupta, N.; Vishwakarma, A. Urban organic waste as a source of bioenergy for electricity generation in Bhopal, Madhya Pradesh. IOP Conf. Ser. Earth Environ. Sci. 2022, 1084, 012024. [Google Scholar] [CrossRef]
Reddy, C.S.; Bird, N.G.; Sreelakshmi, S.; Manikandan, T.M.; Asra, M.; Krishna, P.H.; Jha, C.; Rao, P.; Diwakar, P. Identification and characterization of spatio-temporal hotspots of forest fires in South Asia. Environ. Monit. Assess. 2019, 191, 791. [Google Scholar] [CrossRef]
Tošić, I.; Mladjan, D.; Gavrilov, M.; Živanović, S.; Radaković, M.; Putniković, S.; Petrović, P.; Mistridželović, I.K.; Marković, S. Potential influence of meteorological variables on forest fire risk in Serbia during the period 2000–2017. Open Geosci. 2019, 11, 414–425. [Google Scholar] [CrossRef]
Tyukavina, A.; Potapov, P.; Hansen, M.C.; Pickens, A.H.; Stehman, S.V.; Turubanova, S.; Parker, D.; Zalles, V.; Lima, A.; Kommareddy, I. Global trends of forest loss due to fire from 2001 to 2019. Front. Remote Sens. 2022, 3, 825190. [Google Scholar] [CrossRef]
Vadrevu, K.P.; Lasko, K.; Giglio, L.; Schroeder, W.; Biswas, S.; Justice, C. Trends in vegetation fires in south and southeast Asian countries. Sci. Rep. 2019, 9, 7422. [Google Scholar] [CrossRef] [PubMed]

Figure 1. Map of Diamer District, Gilgit-Baltistan.

Figure 2. (a) Digital elevation model value, (b) hill shade value, (c) aspect value, and (d) slope value.

Figure 3. (a) Diamer district’s annual burned area of forest fires, 1998–2023; (b) Diamer District’s annual forest fire occurrence of different causes from 1998 to 2023.

Figure 4. (a) Diamer District, Julian dates of fires from 1998 to 2023; (b) Diamer District, Julian dates of huge fires from 1998 to 2023.

Figure 5. Julian dates of the earliest and latest forest fires in the Diamer District from 1998 to 2023. blue bullets and line signifies early fire events and its regression, red bullets and line shows later fire events and its regression.

Figure 6. Diamer District, annual climate conditions in May, June, and July from 1998 to 2023. (a) temperature; (b) relative humidity; (c) annual rainfall; (d) wind speed.

Figure 7. Climate–fire relationships in the Diamer District from 1998 to 2023. (a) temperature; (b) relative humidity; (c) rainfall; (d) wind speed. bullets, lines, and shadows represents fire events, regression and correlation, respectively.

Figure 8. Climate–fire relationships in the Diamer District in May from 1998 to 2023. (a) temperature; (b) relative humidity; (c) rainfall; (d) wind speed.

Figure 9. Squared error loss across iterations for random forest and GBM models. The black curve is training set loss and the green and red curves represent validation loss for the Random Forest and Gradient Boosting Machine models, respectively. The blue dashed line represents the optimal point of iteration for the Gradient Boosting Machine model. The figure represents the model’s learning dynamics: Random Forest depicts a relatively low overall error with more stable convergence compared to Gradient Boosting Machine, showing a growing error from overfitting at the later iterations.

Figure 10. Feature importance plot for Random Forest.

Figure 11. Variable importance in the Gradient Boosting Model (GBM). The x-axis is the relative influence of each variable on a scale from 0 to 100%. The higher the value, the more significantly the variable contributes to the model’s predictions. In this figure, ‘forest_fire_alarms’ is the most important variable, contributing almost entirely to the prediction, while ‘MEAN_ELEVATION’ shows negligible influence.

Figure 12. Residuals (predicted—actual burned area) for the Random Forest Model, illustrating that the x-axis shows the actual burned area (in hectares), while the y-axis indicates the residuals. The point that is closer to the red dashed line at 0 indicates better predictions. Negative residuals suggest the model tends to underestimate, especially for larger burned areas. Additional analysis suggests potential bias in the model’s handling of larger fire events.

Figure 13. Residuals (predicted—actual burned area) for the Gradient Boosting Model (GBM), showing the actual burned area (in hectares) on the x-axis, while residuals are on the y-axis. Points closer to the red dashed line suggest better predictions. Larger residuals, particularly for larger burned areas, suggest the model underestimates burn areas, similar to the Random Forest model. This pattern reflects the challenges of accurately predicting extreme fire events using the GBM.

Figure 14. Partial Dependence Plot (PDP) for the Log of Burned Area in the Random Forest Model, showing the log-transformed burned area (in hectares) on the x-axis, while indicating the predicted burned area on the y-axis. The plot suggests that for low log burned area values (below 3), the predicted area remains stable at around 3000 hectares. However, predictions increase sharply when the log burned area exceeds, highlighting that the model predicts significantly higher burned areas for larger fires, suggesting the model’s sensitivity to extreme fire events.

Figure 15. Partial Dependence Plot (PDP) for the Log of Burned Area in the Gradient Boosting Model (GBM).

Figure 16. Correlation heatmap of fire metrics, including meteorological and environmental factors. It shows the interrelation between different fire metrics and the environmental parameters of precipitation, temperature, and relative humidity.

Table 1. Summary of collected datasets.

Type of Data	Year	Data Source
Meteorological data	1998–2023	https://www.pmd.gov.pk?/ (accessed on 14 March 2024) https://power.larc.nasa.gov (accessed on 14 March 2024)
Wildfires data	1998–2023	https://www.fwegb.gov.pk (accessed on 18 October 2023)
DEM data	2023	https://earthexplorer.usgs.gov/srtm (30 m) (accessed on 28 September 2024)
Slope data	2023	https://earthexplorer.usgs.gov/srtm (30 m) (accessed on 28 September 2024)
Burned area	1998–2023	https://www.fwegb.gov.pk (accessed on 12 January 2024)

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

© 2025 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/).

Share and Cite

MDPI and ACS Style

Abbas, K.; Souane, A.A.; Ahmad, H.; Suita, F.; Shu, Z.; Huang, H.; Wang, F. Correlating Fire Incidents with Meteorological Variables in Dry Temperate Forest. Forests 2025, 16, 122. https://doi.org/10.3390/f16010122

AMA Style

Abbas K, Souane AA, Ahmad H, Suita F, Shu Z, Huang H, Wang F. Correlating Fire Incidents with Meteorological Variables in Dry Temperate Forest. Forests. 2025; 16(1):122. https://doi.org/10.3390/f16010122

Chicago/Turabian Style

Abbas, Khurram, Ali Ahmed Souane, Hasham Ahmad, Francesca Suita, Zhan Shu, Hui Huang, and Feng Wang. 2025. "Correlating Fire Incidents with Meteorological Variables in Dry Temperate Forest" Forests 16, no. 1: 122. https://doi.org/10.3390/f16010122

APA Style

Abbas, K., Souane, A. A., Ahmad, H., Suita, F., Shu, Z., Huang, H., & Wang, F. (2025). Correlating Fire Incidents with Meteorological Variables in Dry Temperate Forest. Forests, 16(1), 122. https://doi.org/10.3390/f16010122

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Menu

Correlating Fire Incidents with Meteorological Variables in Dry Temperate Forest

Abstract

1. Introduction

2. Materials and Methods

2.1. Study Area

2.2. Data Sources

2.2.1. Meteorological Data

2.2.2. The DEM, Hill Shape, Aspect Value, and Slope Value

2.3. Forest Fires and Climate Data

2.4. Data Limitations and Contextual Variables

2.5. Modeling Approach

2.5.1. Dataset Preparation

2.5.2. Random Forest Model

2.5.3. Model Validation for Random Forest

2.5.4. Gradient Boosting Machine Model (GBM)

2.5.5. Statistical Validation for Models

2.5.6. Statistical Analysis

3. Results

3.1. Burned Area Trends over Time

3.2. Periodic Analysis of Forest Fire Events

3.3. Correlation of Forest Fires with Climate Variables

3.4. Modeling Results

3.4.1. Performance Evaluation and Model Comparison

3.4.2. Feature Importance Analysis

3.4.3. Residual Analysis

3.4.4. The Partial Dependence Plots (PDPs)

3.4.5. Correlation Matrix

3.4.6. Model Selection

3.5. Statistical Analysis Outcomes

3.6. Practical and Theoretical Implications

4. Discussion

4.1. Correlation Between Forest Fires and Time

4.2. Burned Area and Forest Fire Occurrence Patterns

4.3. Climate Factors and Forest Fires: Analyzing Trends and Implications

5. Limitations and Future Recommendations

6. Conclusions

Author Contributions

Funding

Data Availability Statement

Acknowledgments

Conflicts of Interest

References

Share and Cite

Article Metrics

Article Access Statistics

Further Information

Guidelines

MDPI Initiatives

Follow MDPI