Spatio-temporal analysis of forest fire events in the Margalla Hills, Islamabad, Pakistan using socio-economic and environmental variable data with machine learning methods

Aqil Tariq · Hong Shu · Saima Siddiqui · Iqra Munir · Alireza Sharifi · Qingting Li · Linlin Lu

Abstract Most forest fires in the Margalla Hills are related to human activities and socioeconomic factors are essential to assess their likelihood of occurrence.This study considers both environmental (altitude, precipitation, forest type, terrain and humidity index) and socioeconomic (population density, distance from roads and urban areas) factors to analyze how human behavior affects the risk of forest fires.Maximum entropy (Maxent) modelling and random forest (RF) machine learning methods were used to predict the probability and spatial diffusion patterns of forest fires in the Margalla Hills.The receiver operating characteristic (ROC) curve and the area under the ROC curve (AUC) were used to compare the models.We studied the fire history from 1990 to 2019 to establish the relationship between the probability of forest fire and environmental and socioeconomic changes.Using Maxent, the AUC fire probability values for the 1999s, 2009s, and 2019s were 0.532, 0.569, and 0.518, respectively; using RF, they were 0.782, 0.825, and 0.789, respectively.Fires were mainly distributed in urban areas and their probability of occurrence was related to accessibility and human behaviour/activity.AUC principles for validation were greater in the random forest models than in the Maxent models.Our results can be used to establish preventive measures to reduce risks of forest fires by considering socio-economic and environmental conditions.

Keywords Forest fires · Maxent · GIS · Disaster risk reduction · Random forest machine learning · Multitemporal analysis

Abbreviations

CDACapital Development Authority

RSRemote Sensing

GISGeographic Information System

DEMDigital Elevation Model

TWITopographic Wetness Index

FWIFire Weather Index

ETM+Enhanced Thematic Mapper Plus

OLIOperational Land Imager

Introduction

Forest fire probability analysis is essential for disaster risk reduction and is useful for the prevention and management of forest fires.The majority of forest fires in Margalla Hills in Islamabad are considered accidental due to human activities, e.g., the combustion of garbage, stubble burning, discarded cigarettes.However, meteorological and environmental factors play key roles in the ignition, combustion, and spread of forest fires.Socio-economic variables are also important for predicting human activities in forest fire areas.This study considers both environmental factors (altitude, precipitation, forest type, terrain and humidity) and socioeconomic factors (population density, distance from roads and urban areas) to evaluate the risk of forest fires (Tanvir and Mujtaba 2006).Although the government has increased prevention efforts there are several factors such as lightning strikes and human behavior that cause forest fires.Every year millions of hectares of forest are burned worldwide (Flannigan et al.2009).The ecological function of fire affects the growth of ecosystems, the supply of soil nutrients and biodiversity.Forest and wildland fires are measured through basic processes which initiate regular vegetation succession.However, unpredictable fire can have significant adverse effects on humans and the environment.Forest-related research has been carried out in Pakistan, highlighting that forest fires are a significant threat, but the causes and resultant protection measures have not been determined (Khalid and Saeed Ahmad 2015).

Pakistan is a land of diverse landscapes and natural resources in the South Asian Ecological Zone.It has a diverse array of edaphic, physiographical, climatic and wildlife features.For a country like Pakistan with an agriculturalbased economy, forestry is one of the most important natural resources.However, Pakistan has one of the lowest forest cover in the world with only has 4.5% of forest area, facing severe forest depletion (Oliveira et al.2012).In Pakistan, forest fires mainly occur during dry, hot seasons in shrub areas and tropical Chir forests, areas of Himalayan subtropical pine forests (Bui et al.2017).Nine types of vegetation are found in Pakistan: swamp and littoral woodlands, tropical dry deciduous vegetation, thorn woodlands, subtropical evergreen broad-leaved forest, Chir pine forests, Himalayan humid temperate forest, dry temperate forest, subalpine forest, and alpine scrub.In this study, the area belongs to the tropical dry deciduous vegetation and Chir pine forests (Satir et al.2016).Previous studies found that forest fires in the Margalla Hills mainly occur due to the carelessness of people, and may cover about 12,605 hectares.Forest fires occurred from 2002 to 2011 in the Margalla Hills.During this period, 75% of fires were under 4 h, while 15% lasted 4-8 h (Roy et al.2014).Forest fires are caused by socioeconomic factors but many people think that forest fires are not critical and it is difficult to model these factors in space and time (Tariq et al.2021).

In 1960, the Government of Pakistan declared Islamabad the capital of Pakistan.All government and non-government offices were moved from the former capital Karachi.After the 1990s, Pakistan had rapid economic and urban growth (Khalid and Saeed Ahmad 2016).Due to the growth of gross domestic product (GDP) per capita, the population increased with the expansion of urban areas (Sajjad et al.2009).Since anthropogenic activities are a main cause of most forest fires in Pakistan, the increase in fire events is the direct result of high rates of urbanization.Therefore, in the present research, the incorporation of socio-economic and environmental factors was necessary for developing effective forest fire predictions.Previous studies have used socioeconomic and environmental factor data with numerous statistical methods, for example, Generalized Linear Mixed Models (GLMMs) and Generalized Linear Models (GLMs) to predict wild fires based on the spatial trend of human-caused fires (Aragão et al.2007; Plucinski 2014).The periods of study differ between regular, seasonal and wider time spans.Different machine learning methods were used to estimate the degree of forest fire occurrence, considering all variables, and using models like Maximum entropy (Maxent) and random forest (RF) models (Oliveira et al.2012; Bar Massada et al.2013; Vilar et al.2016).Machine learning algorithms were used to check random data subsets and this can be used as a nonparametric structure (Elith et al.2006).Maxent is a nonlinear regression model originally developed to use point positions and layers to predict the spatial distribution of species (Elith et al.2006; Phillips et al.2006).In several studies, it has been applied to the probability of fire ignition, because the distribution of fire caused the spread of organisms and a good outcome has been obtained (Oliveira et al.2012).The RF machine learning model is a non-parametric model of regression and classification trees based on ensemble technology.Numerous studies have implemented this model to estimate the likelihood of fire occurrence and to produce fire maps with reasonable accuracy (Ok et al.2012).In this study, multiple socio-economic factors that could influence forest fires are included in the analysis to predict and analyze the spatial distribution of forest fire probability in Margalla Hills, National Park, Islamabad, Pakistan.The objective of this research is to analyze and predict the spatio-temporal distribution of forest fires using random forest (RF) and Maxent models, by comparing the results of both models and determine the relationship between forest fire probability and socioeconomic data or environmental changes from 1990 to 2019.

Materials and methods

Study area

Margalla Hills National Park is in the northeast of Islamabad, Pakistan with an area of 15,883 hectares (Abidi et al.2012).It lies between 33°043’ N ans 72°055’ E (Fig.1) and is one of the more ecologically sensitive and protected areas in Pakistan.The Federal Government established the National Park April 27, 1980 to protect the natural environment and biodiversity around the capital city.A largescale reforestation program has been carried since 1970 and dense forests now cover 45% of the area, of which 29% are deciduous, 39% are conifers, and the remaining 28% are mixed forests.The terrain is undulating and rough, consisting of steep slopes and gullies with altitudes from 465 to 1600 m a.s.l.(Iqbal et al.2013).The rock base belongs to the Paleocene to Eocene period with the foundation of limestone (Iqbal et al.2013; Khalid and Saeed Ahmad 2015).The rock colour is greenish-brown and the calcareous substance is grayish blue.The predominant soil type is thin and consists of silt clay with a reddish color with a medium texture and well-developed profile (Malik and Husain 2003).The region has a subtropical climate with mild summers and winters.Average summer temperature is 34.3 °C, in winter 8.4 °C.It rarely snows in winter and the annual average rainfall is 1200 mm a-1(Smakhtin and Hughes 2007).

Fig.1 Location of the study area: a Study area in Pakistan; b History of forest fires from 1990-2019; villages, roads and uses of Margalla Hills; c Topographical map of Margalla Hills

History of forest fires

The geospatial forest fire data (i.e., fire frequency, extent and area, date of occurrence, spatial location and cause) from 1990 to 2019 were obtained from the Islamabad Capital Development Authority (CDA).The number of wildfires increased during 1990-2019 (Fig.S1).In this study, we prepared three groups, i.e., 1990-1999 (1999s), 2000-2009 (2009s), and 2010-2019 (2019s), according to 10-year-data intervals.The cumulative fire incidents in these three periods were 76, 187 and 264, respectively (Fig.1).The forest fire data were used with the Maxent model.However, the RF model generated a number of random points to cover missing data which were combined with the location-based original data as required in the analysis (Arpaci et al.2014).In this analysis, the missing points were randomly distributed between individual points at fire-free locations.The data set of fire locations was divided into two subsets, training and validation (Pourtaghi et al.2016).

Environmental and socio-economic data

Previous studies analyzed a number of variables such as slope, aspect ratio, elevation, distance of fire from roads, and population density to assess the impact of environmental and socioeconomic factors on the probability of forest fire (Arpaci et al.2014; Pourtaghi et al.2015).We used the environmental and socio-economic data included forest type, forest density, topographic wetness index, slope, aspect, elevation, curvature, precipitation-summer, precipitation warmest quarter summer, temperature summer, temperature warmest quarter summer, humidity summer, daily mean wind speed, fire weather index, land use land cover, distance to water bodies, distance to roads, population density, visitors and distance from urban areas (Table 1; Fig.S2).

Table 1 Estimation of forest fire probability using input variables from 1990-2019

Table 2 Input variables (percentage contribution) during the 1999s, 2009s, and 2019s (Maxent analysis) to the probability of forest fires

Elevation data were acquired from ALOSPALSAR Digital Elevation Model (DEM) at 30-m spatial resolution.Topographic wetness index (TWI) was calculated using a corresponding equation to determine soil wetness (Timm et al.2006).Meteorological data were acquired from the Pakistan Meteorological Department, and interpolated maps were prepared using the weighted inverse distance method.In order to link geographic features with precipitation, the rainfall lapse rate was added by representing elevation to the dataset (Phillips et al.2006).Summer precipitation (prcp-summer), summer precipitation of warmest quarter (prec-wqs), summer temperature (temp-summer), summer temperature of warmest quarter (temp-wqs), summer humidity (humidity-sum), summer humidity of warmest quarter (humidity-wqs), commonly used with monthly data to detectmeteorological drought, were calculated using the ‘SPIGA’ package R (Roy and Kumar 2017).

Over a period of 26 weeks, this index tracked rainfall, temperature and humidity which have an influence with fire occurrence (Hislop et al.2018).This study considered meteorological variables during the summer because forest fires in Pakistan have mainly occurred in this season.High precipitation during the summer affects the annual average and makes the probability analysis more challenging (Bridges 2008).In this study, May, June and July define summer (Hirose et al.2004).The fire weather index (FWI) and topographic wetness index (TWI) was estimated using the “fwi.fbp” in R package and data were interpolated using the IDW method (Wang et al.2016).Socioeconomic variables included national park tourist data, population density, and trails and distance from urban areas and roads.The statistical data of the socioeconomic variables were converted into spatial data using ArcMap10.6.Population data was provided as mathematical data, while administrative shape files were obtained from the Capital Development Authority.The polygon data was spatially related to the population, and the population density was calculated to the extent of the study area boundary.In order to measure the distance from urban areas, roads and water bodies, Landsat TM, ETM and Landsat 8 data were used (Tien Bui et al.2017) and the Euclidean distance method for data estimation.For this analysis, the spatial resolution of all input variables in the random forest point format and the Maxent ASCII grid format was 500 m.In addition, the focus of this study was to produce high-resolution maps of forest fire probability.All coarser and higher resolution data sets were resampled at 1-km resolution because downscaling would not modify the properties of data with coarser resolution but would permit us to retain all the data in higher resolution.

Maximum entropy model (Maxent)

For possible and current estimation of the spatial dispersal of species, Maxent is a machine-learning technique commonly used to test the differences between layers and point locations.Maxent procedures present data based on the maximum entropy principle to predict the dispersal of chance or habitat suitability of animals.According to the research variables and categorical or discrete input limitations, the model predicts scattered subjects (Hastie et al.2009).Maxent produces a probability of occurrence ranging from 0 to 1 for each variable response curves.In addition, the model results provide the field under the operating characteristic curves of the receiver (AUC) (Vilar et al.2016).The AUC value ranges from 0 to 1.0, and a value of 0.5 indicates that the performance of the model is not better than that of random, while a value close to 1.0 indicates that the performance is better.There is an AUC value greater than 0.7 for a model producing decent or strong predictions (Hastie et al.2009).In addition, in terms of statistical significance, Maxent is considered highly accurate.In some studies, the model has obtained excellent results in the prediction of forest fires because the frequency of fires can be considered related to the distribution of species (Ali et al.2019).Maxent version 3.4.1 was used to evaluate the probability of forest fires in the study area at a spatial resolution of 500 m.Random points were used to acquire the probability from 0 to 1, position of forest fire is 30%, and 6500 is the limit for maximum iteration.Bootstrap with 15 simulation runs was used to minimize the uncertainty of probability.The format of the output was described as logistic.

Random Forest (RF) model

In previous studies, RF algorithms were used for the classification and regression trees based on ensemble techniques (Hastie et al.2009; Aldersley et al.2011; Rahmati et al.2016).The model created decision trees of a number of randomly selected guide samples to obtain predictions for each tree.A selection was performed later to create a subset based on the independent variables at each node.The final result of this model was the combination of the outcomes of all trees (Cutler et al.2007).While running the RF model, all factors used in tree planting should be determined with the total number of trees planned to plant.The average result over a large number of trees showed that the model had low bias and low variance.

The model retained one-third of the samples in order to verify and confirm accuracy, which leads to an unbiased generalization of errors.Out-of-bag (OOB) errors considered the proportion of misclassified (percent) components from all OOB components.Without a separate test range, OOB errors calculate the performance of the model (Tariq et al.2021).In the calibration process, the purpose of the RF model was to classify the appropriate model to evaluate the association between a dependent and independent variable so as to determine the weight of each component (Calle and Urrea 2011).In this study, the R ‘random forest’ and ‘sdm’ software packages were used (Tariq and Shu 2020).The data of forest fire occurrence was used as the dependent variable and independent variables were: forest type, forest tree condition, LULC, topographic wetness index (TWI), fire weather index (FWI), wind speed, summer precipitation, precipitation warmest summer quarter, summer temperature, temperature warmest summer quarter, summer humidity, humidity warmest summer quarter, slope, aspect, elevation, curvature, distance to water bodies, distance to roads, population density, number of visitors, distance from urban areas.

Model performance

To analyze the model output, validation data sets were used which were not used in the training process of the model.By applying the most common threshold-independent approach of the receiver operating characteristic (ROC) curve, the projecting output of the model was evaluated (Bui et al.2016).On the vertical axis, all combinations of sensitivities were plotted, and on the horizontal axis, the proportions of false negatives (1-specificity) were plotted.The area under the receiver operating characteristic curve (AUC) was used as a quantitative efficiency indicator.AUC=1 indicates a perfect estimate while an AUC ＜ 0.5 indicates a poor result (Aldersley et al.2011).The grading of model results based on the AUC metric was: 90-100% (excellent), 80-90% (very good), 70-80% (good), 60-70% (moderate) and 50-60% (poor) (Rahmati et al.2016).Sensitivity was derived on the basis of the correctly expected fraction of fire events (i.e., positive points), while ‘1-specificity’ is the fraction of incidents that have not occurred and were incorrectly predicted (Hirose et al.2004).Pourtaghi et al.(2015) used linear and mixed-effect models to calculate spatiotemporal distribution which combined both fixed and random effects and considered multiple sources of variance.

Results

Results of the Maxent model

In the 1999s, 2009s and 2019s using the Maxent model, the average probability of a forest fire was s 0.5%, 0.6% and 0.5%, respectively.The research showed that until 2009s, the average risk of a forest fire began to rise and then declined slightly in the 2019s.In urban areas and the eastern region of the Margalla Hills where low-altitude forests were open to the public, the spatial distribution of forest fire risk was higher (Fig.2b).The Maxent findings show the percentage contribution of each input variable to the likelihood of forest fires.In the past decades, performance of Maxent model has been important for variable population (Table 2).In all periods, elevation had a high percentage contribution.In addition to these factors, in 1999s and 2009s, urban areas, and in 2019s, TWI (topographic wetness index) accounted for an important proportion.In addition, environmental variables such as wind speed, precipitation-summer, precipitation warmest quarter summer, temperature summer, temperature warmest quarter summer, humidity summer, humidity warmest quarter summer, FWI (fire weather index) and TWI showed either a diminishing importance or negligible significance over time as a percentage contribution to fire likelihood.

Fig.2 Probability of forest fires during the 1999s, 2009s, and 2019s.a RF Model, b Maxent Model

Maxent-based response curves showed the expected probability of forest fires changes through each variable.The red curves showed average replication response.A similar association with the probability of forest fires was seen in each variable (Fig.3).However, Maxent predicted a strong negative correlation between forest fire probability, elevation change and topographic wetness (TWI index).In addition, forest type was a factor of intermediate significance, with less than 11% variation between the forest condition and forest type.

Fig.3 Answer curves in the 1999s, 2009s, and 2019s illustrate the relationship between forest fire probability and the input variables (Maxent model analysis)

In general, summer precipitation was negatively correlated, but precipitation wetness quarter (Prec-wqs) produced reliable findings because it revealed both positive and negative correlations.In the 1999s, FWI was negatively correlated.However, the 2019s showed a positive correlation.This study supports the contention that the frequency of forest fires was affected by less rainfall.Population showed a positive correlation during the entire period of the Maxent analysis.This suggests that higher population density increased the probability of forest fires.However, no significant association was found between tourism and fire occurrence.In addition, as the distance from urban areas increases, the risk of forest fires gradually decreased.

Random forest results

For the Random Forest analysis, the average forest fire probabilities, which are the average raster values for the study area, were 0.782, 0.825, and 0.789 during the 1999s, 2009s,and 2019s, respectively.This study also revealed that, until 2009s, the average risk of forest fires increased and then decreased during the 2019s.Some variations over time were observed in spatial spreading of forest fire; however, the RF study indicated a great probability of fire occurrence in southeastern part of the study area and near the urban areas (Fig.2a).Over the last few decades, the probability of forest fire occurrence has been concentrated in Margalla Hills.The Random Forests results demonstrate the significance of each input variable in predicting the likelihood of a forest fire.The findings show the importance of population and elevation through all decades (Table 3).Aside from these factors, SPI-summer accounted for a significant percentage in the 1999s and 2009s, and urban accounted for a significant percentage in the 2019s.

Table 3 Importance for forest fire probability of each input variable during the 1999s, 2009s, and 2019s (RF Analysis)

Comparison and validation of the models

Both Maxent and RF-based findings revealed that the spatial distribution of Islamabad forest fires was concentrated around towns and in some rural areas.However, over the past few decades, the trends in forest fire probability changed.The total spatial distribution of forest fires was comparable although the RF results indicated a higher average probability compared to those computed by using the Maxent model.In addition, overall probability was highest during the 2009s in both models.The results of both Maxent and RF models showed a significant influence of forest type (vulnerable), elevation, population density and its distribution in forest fire occurrence.The vector importance of urban areas to the likelihood of fire occurrence was also assessed and the results show that the most influential factor was population density.During last two decades, elevation was one of the essential variables.For climate variables, especially those related to precipitation, the Maxent model result showed little significance, while the RF was highly significant for summer precipitation, summer precipitation in the warmest quarter, summer temperature, and summer temperature in the warmest quarter.AUC values were used to calculate the efficiency and accuracy of the models and the findings demonstrated remarkably good statistical precision, g indicating that the frequency of forest fires in Islamabad was mostly due to human activities and this was simulated at the national level at a spatial resolution of 500 m.Using the Maxent model, the AUC fire probability values in the decades of the 1999s, 2009s, and 2019s were 0.532, 0.569, and 0.518, respectively; while using the RF model, the AUC fire probability values were 0.782, 0.825, and 0.789, respectively.Therefore, the RF model was more accurate than the Maxent model in calculating forest fire probability in Islamabad.

Discussion

Effect on forest fires by environmental and socio-economic variables

Islamabad has developed rapidly and awareness of urban sprawl has increased significantly, particularly during the 2009s, alongside the socio-economic growth of the country from the 1999s to the 2019s (Smakhtin and Hughes 2007).During this period of exponential growth, there was a rising trend in the occurrence of wildfires.We determined the relation between the probability of forest fires and socioeconomic variables in this study.Since most of the wildfires are considered to be a function of human activities, forest elevation and population density are helpful to understand the relationships.A higher population and lower forest elevation indicates more human activity (Kim et al.2019).These aspects are considered important as an increase in population contributes to a rise in activities and forests are less available at high elevations.

This study shows that there was a clear association between the probability of forest fires and variables such as population density, and elevation, especially during the 2009s.Over the three decades, from both the Maxent and RF data, the variable value and percent contribution of population density and elevation was consistently substantial.However, the percentage involvement and significance of population growth was the highest during the 2009s when it lead to a large metropolitan sprawl.Both models showed that forest fires occurred mainly in urban areas and in the eastern village zone with respect to their regional distribution.In both models, the value of TWI and elevation has diminished over the decades and this is compounded by global urbanization that has led to a reduction in the terms of distance from urban areas over time.

In both models, the topographic wetness index (TWI) revealed a distinct contribution between environmental variables over all periods and a poor association to the likelihood of the Maxent consequence.Since TWI works as an indicator of soil moisture over a long period of time, the probability of forest fire is higher under dry soil conditions.Precipitation was also negatively correlated with fire probability during spring according to the Maxent analysis.With the Maxent model, there was less than 10% difference found between the different types of forest, but previous studies indicate that coniferous forests have a much greater correlation to fire risk and the quantity of available fuel is high during the season (Won et al.2014).

Spatial distribution and precision machine learning models

According to distribution, the forest fire probability is more localized in or near urban areas as shown in both the Maxent and RF studies.This suggests a narrower variety of spatial differences in burning danger and means that near cities fires occur more frequently.Compared to naturally-occurring forest fires, Islamabad forest fires are often due to different human activities, suggesting a lower spatial self-correlation (Nalder and Wein 1998).This has constrained the accuracy of model estimations, but with the help of models based on machine learning applied to forest fires, statistically significant results have been produced.In the Maxent analysis, the AUC value decreased over the three decades, attributed to larger collections and because of spatial autocorrelation.This would breach the assumption of independent observations (Nalder and Wein 1998; Jenkins et al.2016).

During 1999s, the AUC value was the highest and throughout the 2009s it was the lowest for RF analysis.The significance has not diminished over the decades, given the larger sample size.This suggests that, with larger populations, the strong ability of the RF model is to solve spatial autocorrelation problems.Two final outcomes of the model were combined on the basis of spatial distribution, and the significance of the primary variables were identical.In achieving higher AUC values however, the RF model was superior, while the risk of forest fires was high and overestimated, compared to the outcome of the Maxent model.AUC values were comparatively low in the case of the Maxent outcome, but fire probability was accurately calculated.With a valid test of each variable using both the Maxent model and the RF model, it is possible to solve problems of overestimation and forecast precision, and acquire overall forest fire distribution and validate the effect of socio-economic drivers on wildfires.

Constraints and limitations

The data on wildfires used in this study were based on Forest Service field surveys.The agency uses field data to track the frequency of forest fires and it is important to consider data reliability.One alternative to field surveys, based on scientific studies, is fire detection using satellite imaging.In a potential study of satellite-derived data and field data and recognizing mistakes, a comparison is proposed.For spatial analysis, socio-economic data sources used in this study have not been generated.However, to demonstrate the associations between the probability of forest fires and socioeconomic shifts, the data were used as variables, translating them into grid maps.This can cause misunderstandings, which can lead to confusion about the model.Any of the variables are interpolated owing to the lack of such details.When looking for missing values, interpolation is helpful in filling the gaps, but the precision of the model could be reduced because of the ambiguity produced in creating prediction estimations.

In the current study, climate data was interpolated to make grid maps in order to obtain Prec-summer, Prec-wqs, Temp-summer, Temp-wqs, FWI and TWI.From the Maxent model, these showed a contradictory relationship with fire probability.The incomprehension that interpolation creates can be attributed to this.Not only did the models interpolate temperature statistics, but socio-economic statistics as well.In spite of including the tourists to the national park, access to the forest area by local residents were also included.Data for total visitors visiting all forests were not available, so the alternative was to use the total count of visitors to national parks.By interpolation, the number of people visiting forest regions from the 1999s could be approximately estimated.

As a consequence, an association with the danger of fires was seen by a number of national park tourists to be insignificant.This can be explained in a way that only the number of visitors to the national parks is limited to the source results.Comparisons between the decades could only be made using the variables available or created after the 1999s.Since then, usable datasets have improved, but for full-period comparisons they have not been used.For distances from roads, an association dataset can be used to portray urbanization but this was only available as of 2015.Isolation from inhabited areas was used instead to illustrate the impact of accessibility on the frequency of burning.As society becomes more complex, new ideas need to be establish in order to manage uncertainties.Future studies need to integrate more complex socio-economic variables and multiple model approaches.In many forests, for instance, because of low humidities and surface moisture, south-facing slopes have ideal conditions for fire outbreaks.

These variables also support the growth of pine species, which contribute to higher fuel accumulation (Costafreda-Aumedes et al.2018).On south-facing slopes, up to 50% of studies have been conducted on forest fires, according to the 2015 CDA Forest Catastrophe White Paper issued by the National Institute of Forest Science.This report provides an analysis of the characteristics of forest fires from 2007 to 2015.However, this report was not included in this analysis of the input variables because the model was based on a spatial resolution of 500 m, an area in which only one function of a grid can be depicted.In a future analysis with downscaled model outcomes, this function should therefore be used.

Reduction of forest fire risk

This study considers both socio-economic and environmental variables, and the correlation between socio-economic growth and the risk of forest fires is demonstrated.The average probability of forest fires during the 2009s was highest for both models, above over the 2019s, considering continuing socio-economic growth.We believe that this is the consequence during the 2009s of improved and more sophisticated forest organization.The CDA developed and introduced the Basic Strategy for the Prevention of Forest Fires (2006-2015) after a large-scale forest fire in 2015, and continued to expand the budget throughout the country to deter forest fires.The findings show that successful management of human activities is critical in reducing the occurrence of forest fires, and that socio-economic factors are progressively being taken into account when reducing forest fires.

Moreover, forest fire probability analyses and the detection of key input variables can mitigate future forest fire incidents, and sustainable forest management can be done (Khalid and Saeed Ahmad 2016).In fire-prone areas, probability maps can be used to minimize the risk of fires; many steps can be used to manage forests more efficiently and in a sustainable way, including management of stand density to reduce water competition, increase plant diversity to modify burn rates, select of tree species to mitigate the effects of climate change and future fire danger and fire risk determination (Hastie et al.2009).

Climate change destabilizes the Earth’s temperature equilibrium and has far-reaching effects on humans and the environment.With global warming, the energy balance and thus the temperature of the earth change due to the increased levels of greenhouse gases.Because human-induced warming is superimposed on a naturally varying climate, temperature rise has not been, and will not be, uniform or smooth across the country or over time.Climate change encompasses not only rising average temperatures, but also extreme weather events, shifting wildlife populations and habitats, rising seas, and a range of other impacts.Since forest fires are threats that can largely be avoided, preventive and precautionary steps, especially those related to human activities in forests near urban centers, should be implemented in a timely manner.

Conclusion

Fire risk mapping is a suitable tool to identify and locate areas around roads and near urban areas that are vulnerable and supports roadside vegetation management.For example, herbicide application and pruning in areas of special risk such as slopes, embankments, and clearings could help reduce the propagation and intensity of forest fires.At the same time, fire risk maps help resource managers and practitioners to develop fire emergency plans.

This study measured the impact of socio-economic changes, such as urban sprawl, on the risk of fire by estimating the probability of forest fires using a machine learning algorithm such as RF and Maxent over three periods in Islamabad, which has been increasingly urbanized and has seen a rapid increase in the occurrence of forest fire incidents.In terms of the overall risk of forest fires, it was higher in the 2019s, and the identification of fire risk by both models was highest in the areas around cities and in eastern hilly areas.The factors with the largest contribution over the three decades are population and elevation.In particular, population density, especially during the 2009s, was the most important and positively correlated predictor for the frequency of fire in this study.

The results demonstrated good predictive precision for sporadic human-caused forest fires in Islamabad, and 500 m spatial resolution images were used.The results of both the Maxent and RF models revealed an important relationship of forest type (vulnerable), elevation, population density, and distance to urban areas to the probability of wildfire in terms of vector importance.In fact, most critical predictor is population size in both models and over all decades.With the RF results, the model accuracy was better than that of the Maxent model but the probability of forest fire was overestimated.This study reveals that the spatial distribution of the risk of fire in or around urban centers has been increasingly dispersed over the decades, and that the risk over time has a clear association to population.This suggests that, in order to reduce the occurrence and effect of forest fires in the Margalla Hills, prevention and preparedness steps for forest fire control must be enforced.

This work lays the foundations for a completely automated fire risk assessment application in the Margalla Hills.The results can be used for a course of action and for planning fire prevention.In the future, to determine the risk accurately, it will be necessary to combine the fire probability maps with seasonal risk using indices and the periodic dynamics of forest fires.Such actions can help build an application with which it will be possible to determine the risk of forest fires and fire spread during extreme events.

AcknowledgementsWe would like to pay special and heart whelming thanks to the Statistical Bureau of Pakistan, Islamabad for providing us population census data and the Capital Development Authority (CDA), Islamabad for digital and all ancillary data.We are also thankful to Dr.Shazada Adnan, of the Pakistan Meteorological Department for providing all meteorological data related with this research.We also admire Dr.Muhammad Imran of Institute of Geo-information and Earth observation (IGEO), University of Arid Agriculture, Rawalpindi, for his support at various stages of the field work.We highly regard the reviewers and editors of the Journal for providing helpful inputs that improved the manuscript.

Author’s contributionAT planned and conceptualized, the methodology, a systematic review, the testing and examination, the tools, curation of data, and preparation of initial draft.HS reviewed, compiled and supervised all work.IM reviewed the manuscript and compiled final manuscript.SS and AS prepared all revisions and reviewed all the manuscript.