Narayan Kayet · Abhisek Chakrabarty · Khanindra Pathak · Satiprasad Sahoo ·Tanmoy Dutta · Bijoy Krishna Hatai

Abstract A comparative study of Frequency Ratio (FR)and Analytic Hierarchy Process (AHP) models are performed for forest fire risk (FFR) mapping in Melghat Tiger Reserve forest, central India. Identification of FFR depends on various hydrometeorological parameters (altitude, slope,aspect, topographic position index, normalized differential vegetation index, rainfall, air temperature, land surface temperature, wind speed, distance to settlements, and distance by road are integrated using a GIS platform. The results from FR and AHP show similar trends. The FR model was significantly higher accurate (overall accuracy of 81.3%, kappa statistic 0.78) than the AHP model(overall accuracy 79.3%, kappa statistic 0.75). The FR model total forest fire risk areas were classified into five classes: very low (7.1%), low (22.2%), moderate (32.3%),high (26.9%), and very high (11.5%). The AHP fire risk classes were very low (6.7%), low (21.7%), moderate(34.0%), high (26.7%), and very high (10.9%). Sensitivity analyses were performed for AHP and FR models. The results of the two different models are compared and justified concerning the forest fire sample points (Forest Survey of India) and burn images (2010-2016). These results help in designing more effective fire management plans to improve the allocation of resources across a landscape framework.

Keywords Forest fire risk (FFR) · Remote sensing · GIS ·FR · AHP · Sensitivity analysis · Validation

Introduction

Forest and wildfires are the most critical threat to the forest area of the worldwide and cause adverse ecological, economic and social effect. Forest ecosystems are increasingly threatened by fires caused by a range of natural and anthropogenic factors, and hence, geospatial assessment of forest fire risk (FFR) is very important to reduce fire impacts. In this research, geospatial and FSI training points using FR and AHP models perform forest fire risk mapping based on the multi-criteria decision-making system using GIS platform.

There are few studies available for FFR mapping based on geospatial and multi criteria decision making models.The first application on aerial infrared image scanners was used for forest fire mapping (Chuvieco and Congalton(1989). After National Aeronautics and Space Administration (NASA) launch of earth resources satellites (1972),several studies were carried out for forest fire and burnt area mapping (Tanaka et al. 1983; Agee 1998). The occurrence of forest fires is of significant environmental concern, affecting forest preservation, generating economic and ecosystem damages and human suffering (Cortez and Morais 2007). Forests are considered essential natural resources, having a role in corroborative economic activities, in maintaining environmental balances and regulating climate as well as the carbon cycle (Clark 1990; Erten et al.2004). Ecological conditions are damaged by forest fires.Due to natural or anthropogenic reasons, forest ecosystems are threatened by increasing numbers of forest fires.Therefore, geospatial assessment of forest fire risk is a critical task to decrease fire impact on ecosystems (Rawat 2003; Kayet et al. 2018a, b). A review paper on the use of remote sensing in forest fire assessment has been studied.There are three major spectral domains used in forest fire detection, monitoring, and mapping (Leblon et al. 2001).One research study has analyzed topographic and climatic factors and their impact in their study of forest fire risk mapping by multi-criteria analysis using AHP and HR models in the Minudasht forests, Iran (Pourtaghi et al.2015).

Another paper on using remote sensing and GIS data is the application to forest fire hazard mapping by Chuvieco and Congalton (1988), and is centred on the Mediterranean coast of Spain, a major fire-prone area in Europe. The study was forest fire mapping using Multi-Criteria Decision Analysis (MCDA) on the Gallipoli Peninsular National Park. Herrera (2016) carried out MCDA based on MODIS active fire data for Mexican forest fires from 2005 to 2014 and showed that most fires occurred in cedar and pine forests on the Atlantic and Pacific coasts.

Studies for current forest condition monitoring, fire risk zone mapping and degradation assessment in an Indian context were carried out by Chuvieco et al. (2003). Kant Sharma et al. (2012) determined the forest fire risk zone by Fuzzy logic and AHP approaches in the Shimla Forest Division. Geospatial technology was also used for the prediction of forest fire risk mapping in Rajaji National Park in Uttaranchal, India. IRS 1D LISS III image and DEM (Digital Elevation Model), climatic and ancillary data were also used for this work, and forest fire risk zones were identified to prepare forest fire control and management plans in the Rajaji forest at the Chilla range of Uttaranchal. Forest fire risk zoning was studied by Saklani(2008) under the guidance of ITC Netherland University.The forest fire risk is identified using remote sensing, climate data and several topographical elements as criteria.

We explore the leading causes of forest fires at the regional landscape scale using the FR and AHP decisionmaking models in a GIS framework to quantify forest fire risk mapping of Melghat Tiger Reserve (MTR) forest. The forest fire zoning results are identified by burning images(2010-2016) and forest fire training points (FSI).

Materials and methods

Study area

The Melghat Tiger Reserve (MTR) is located in the Satpura hill ranges of central India, in the Melghat forest of Amravati district of Maharashtra (Fig. 1). The MTR area of 1571.74 km2lies in the heart of the Melghat forest and was declared the Melghat Tiger Reserve on 22nd February 1974. This area has dry deciduous forests and excellent tiger habitat. The MTR is within latitude 21°0′15′′to 21°0′45′′N and longitude 76°0′57′′to 77°0′30′′E at elevations of 312 to 1178 m MSL. It is the largest of three project tiger programs in the state. The yearly maximum temperature averages 42.7 °C (Fig. 2), with annual rainfalls ＞2700 mm. Winds are generally light to moderate;the strongest winds are less than 22 km/h.

Data sources

Landsat-8 satellite images of Maharashtra state were acquired in 2016. LST (land surface temperature, TPR, and NDVI layers were computed from the Landsat 8 OLI satellite images. The Landsat data were obtained from GLOVIS (glovis.usgs.gov). Maps for elevation, slope, and aspect were prepared from Cartosat-1 DEM (Digital Elevation Model) data collected from Bhuvan (bhuvannuis.nrsc.gov.in). In the multi-criteria process, ten thematic maps (elevation, slope, aspect, TPR, NDVI, rainfall, LST,wind speed, distance to settlements, and distance to the road) were extracted from satellite images, quick bird images (Google Earth), and climate data (National Centers for Environmental Prediction (NCEP) Climate Forecast System Reanalysis (CFSR). Settlement and road layers were computed from the quick bird images. Rainfall, wind speed, and temperature layers were extracted from weather data. Forest fire training points were utilized in preparing the point vector layer. Forest fire training data (2000-2016)was acquired from Forest Survey of India (FSI). All spatial,thematic raster layers were scaled to 30 × 30 m pixel grid cell.

Factors affecting forest fires

There are many physical and anthropogenic factors which affect forest fires in the area. Physical factors include weather conditions, topography, drainage, soil textures,vegetation type and density. Anthropogenic factors are mainly settlements and connectivity of the area.

Fig. 1 Location map of the MTR forest

Topography

Topography data are one of the primary parameters of forest fires. The impact of slope, aspect, and altitude on fire behaviour have been reported by Renard et al. (2012) and Adab et al. (2013). In this study, topographic features were generated using Cartosat-1 DEM. Maps of slope degree,aspect, elevation, and TPI (topographic position index)were prepared.

Meteorology

For weather data, temperature, wind speed, and rainfall factors were determined. Wind speed and land surface temperatures are the main factors affecting forest fires(Kayet et al. 2016a, b). A LST map was generated from Landsat 8 OLI satellite images. Maps of rainfall, wind speeds, and temperature were used for criteria factors.These maps were prepared in raster format in GIS software by the spatial interpolation method.

Vegetation

Vegetation types and density are significant factors in forest fire assessment. Different types of vegetation have different carbon and chemical contents which respond significantly during a fire event. Dense and dry vegetation is more susceptible to fire in comparison to moist, sparse vegetation (Flannigan et al. 2001). For calculating NDVI,the Landsat 8 OLI imagery obtained 2016 was used.

Anthropogenic factors

These are human-related factors and in India, most forest fires are caused by human activities. Distances to settlements and distances to roads are used to prepare the criteria.

Pre forest burn mapping

The pre-fire land cover is one of the significant factors that influence post-FFR mapping (Lozano et al. 2007). This study analyzed annual burn maps (2000-2016) using Landsat images and Forest Survey of India forest training points (2010-2016) to calculate the yearly burned area from 2010 to 2016. These annual burn maps support FFR mapping.

Methodology

At the first step, the require data were collected from different sources, processed with scientific correction methods, and various factors prepared by attribute values. All factors were converted to raster format with same grid cells. All raster layers integrated by FR and AHP models generated the forest fire mapping (Fig. 3).

Fig. 2 Changes in primary climate factors of average temperature (a), rainfall (b), and average wind speed (c) in MTR forest area

The frequency ratio (FR) model

The FR is the probability of occurrence of a certain attribute. If we construct an event E and certain attributes attributed to F, the FR probability of F can be written regarding conditional probability as Eq. 1 (Sahoo et al.2017):

The steps for computing the FR: (1) well location determination; (2) representing cells of sub-attribute 1 of an attribute; and, (3) detailed area of spatial overlap for well areas and area of class 1 for the attribute.

To obtain the forest fire vulnerability index (FFVI), each attribute’s FR value is summed for a pixel as Eq. 2:

where index (i, k) denotes attribute and sub-attribute respectively; NFis the total number of attributes;is the number of sub-attributes for i th attribute;is the frequency ratio value of kth sub-attribute for ith attribute;is the class value of (px, py) cell for ith attribute;is the sub-attribute interval;is the indicator function for kth sub-attribute of ith attribute and defined as Eq. 3.

The FR values indicate the relative relationship to forest development. The greater the value, the higher the probability of forest fire occurrence, and the lower the value, lowers the risk of forest fire occurrence.

Fig. 3 Overlay methodology of forest fire risk mapping in MTR forest

Analytical hierarchy process (AHP) model

The analytical hierarchy process model is a theory of measurement for considering tangible andintangible criteria that have been applied to numerous areas such as decision theory and conflict resolution (Harker and Vargas 1990; Yalcin 2008). Forest fire risk mapping (FFRM) can be calculated as Eq. 4 (Dhar et al. 2015).

where indices (i, j) denote row and column location of a pixel; F the set of all features (F), k the elements of feature set; Skdenotes a set of sub-features for kth feature; l denotes element of sub-feature set; Wkthe normalized weight of kth feature;the normalized weight of lth subfeature for kth feature;denotes the class value of the cell (i, j) for kth feature;the sub-feature interval;the indicator function for lth sub-feature of kth feature and defined as Eq. 5.

The AHP model can be applied for estimations of Wkand. In the AHP model, a 1-9 scale is used, (i.e.,1 = extremely unimportant, 2 = strongly unimportant,3 = unimportant, 4 = moderately unimportant, 5 = equally important, 6 = moderately important, 7 = more important,8 = strongly important, 9 = extremely important) is adopted for constructing judgment matrices. The following steps are used for the calculation of weights and consistency ratios (C.R.):

Step I Development of judgment matrices (A) by pair wise comparison

Step II Calculation of relative weight (Eq. 6) Wk:

where the geometric mean of the kth row of judgmentmatrixiscalculatedasNFis the total number of features

Step III Strength assessment of judgment matrix based consistency ratio (C.R.), Eq. 7:

Consistency index (C.I.) is evaluated using Eq. 8:

where the latent root of judgment matrix is calculated using Eq. 9:

where W is the weight vector (column), the random consistency index (R.C.I.) can be obtained from standard tables. C.R. values less than 0.1 are acceptable for a specific judgment matrix. However, revision in judgment matrix is needed for C.R. ≥0.1. The same procedure should be followed forcalculation. Finally, index maps can be generated from the above procedures to identify forest fire risk zones based on FFRZ.

Sensitivity analysis

In this study, sensitivity analysis explores the relationships among all forest fire parameters and modeling applications.FR and AHP models were evaluated for sensitivity analysis and identifying which parameters had high impacts on forest fires (Saltelli et al. 2000). Sensitivity analysis was crucial to the validation and calibration of numerical models (Chen et al. 2010; Dhar et al. 2015). Influencing attributes can be identified by using Eq. 10:

where i is the subscript for a attribute, j is the superscript for FFR category [j = 1(very low), 2(low), 3(moderate), 4(high) and 5 (ver y high)].is the change (+/-) in forest fire zone area,is the jth type of forest fire vulnerability zone area without ith attribute, and SFj is the jth type of forest fire vulnerability area using all attributes.

Results and discussion

Annual burn area maps (2010 to 2016)

In Melghat Tiger Reserve (MTR) forest, burned area images (2010-2016) are shown in Fig. 4. Approximately 11.4% of the areas were burned in January and May months of the year 2010. The 22.1% areas were burned in February and June months of year 2011. In the 2012 year,26.5% areas were burned in February and May months. In 2013 year, approximately 8.5% of areas were burned in March and May months. In the 2014 year, 14.1% of the areas were burned in January and in June months. In the 2015 year, 5.0% of the areas were burned in May month. In the 2016 year, 12.5% areas were burned in January and March months. In the years 2010 to 2016 burn area shown in Fig. 5.

Forest fire risk maps validated from the FR model

Ten rating parameters (Fig. 6a-j) were used for the generation of FR-based forest fire risk (FFR) zone mapping.The result of FR-based fire risk map identifies five classes(very low, low, moderate, high and very high). The result of the spatial relationship between forest fires and forest fire-related factors using the FR model is illustrated in Table 1. FR-based forest fire risk areas were covered by very low (7.1%), low (22.2%), moderate (32.3%), high(26.9%), and very high (11.5%). In the study area, some forest compartments are very high fire risk. The fire risk potential maps derived from the FR model are shown in Fig. 7. The FFR image accuracy assessment was validated using Forest Survey of India fire points (2010-2016) and satellite image-based burn images (2010-2016). In this study, the overall classification accuracy is noted in Table 2. FR-based forest fire risk results are validated by FSI fire points and satellite image-based previous burn maps of 2000-2016.

Sensitivity analysis of FR based model

Fig. 4 Annual maps of burned areas in MTR forest (2010 and 2016)

Fig. 5 Annual forest burned areas in MTR forest (2010-2016)

Sensitivity analysis is important for FFR identification(Table 3). LST (land surface temperatures), wind speed,DS (distance to settlements), DR (distance to roads), and TPR (topographic position index), are influencing factors.

Forest fire risk maps validated from the AHP model

The idea of multi-criteria techniques has been implemented with different uncertainty levels of the AHP model. The FFR zone map was generated by normalized weighted mean overlay analysis (Table 4). Ten rating parameters were used for the generation of AHP based-FFR zone mapping. The fire risk potential maps are shown in Fig. 8 and indicates five classes (very low, low, moderate, high and very high). AHP-based fire risk areas were very low(6.7%), low (21.7%), moderate (34.0%), high (26.7%), and very high (10.9%). The overall classification accuracy is shown in Table 5. AHP-based FFR zone map accuracy assessment was based on FSI fire points (2010-2016) and current satellite-based burn images (2010-2016). In the MTR forest, some compartments are very high fire risk.

Sensitivity analysis of AHP based models

Sensitivity analysis results are shown in Table 6. LST,wind speed, DS, DR, slope and TPR, are the influencing attributes.

The relationship between parameters and fire

In this study, LST (land surface temperatures) has the highest FR and AHP values and is a good indicator associated with water content of the fuel. The forest fire influences the maximum range of LST. For aspect, FR and AHP fires were most abundant on southwest facing,southeast facing, and east facing slopes. Thus, slope facing aspects are susceptible to fire. The AHP and ratio values of fire were lowest on flat areas, in the northeast and north.Wind speed 1.8 to 2.2 m s-1(6.5-7.9 km/h, respectively)FR and AHP value is medium, 2.2-2.6 high(7.9-9.4 km h-1, respectively), and 2.6-2.9 very high(9.4-10.4 km h-1). Wind is a main environmental factor in the occurrence and extension of forest fires and 2-3 m/s(7.2-20.8 km h-1) highly related to forest fires. In the case of slopes, FR and AHP ratios are lower in higher slopes than lower ones. For the slope class of ＞15, the value is found too high and low for the class of 45-60 slopes. Steep slope increases the rate of fire spread due to more effective connective preheating and ignition. The FR and AHP values decrease as elevation increases; at 459-556 m, FR and AHP values are very high, but the AHP and FR values low do not increase at higher elevations (805-1114 m).Healthy vegetation contains high chlorophyll contents, and unhealthy low chlorophyll. Healthy vegetation is not favorable to forest fires, so the frequency ratios for healthy vegetation becomes very low. The highest FR and AHP values fall into the unhealthy vegetation class. The criteria of topographic position index indicates that flat areas are not as favorable for fire occurrence as are ridges and gentle slopes. Topographic position index FR and AHP ratios for flat areas were high. Ridges and gentle slopes have medium and low ratios, respectively. The relation between rainfall and forest fires shows that for rainfall between 1354.33 to 1606.85 mm, and 1606.83 to 1787.94 mm, FR and AHP ratios values are high and then decreased. The FR and AHP values decreased for rainfall 1787.49 to 1924.42 mm. It is easily to predict fire incident increases in the case of dry areas, compared to wet areas with high rainfalls. There is a negative relationship between fire and distance to roads. In this study area, 5 km distance from road area gets lower FR and AHP ratios, indicating a relatively safe zone. A 1-k distance from the road has higher FR and AHP ratios,indicating an unsafe zone of fire possibility. There is also a negative relationship between fire possibility and distances to settlements. At a 5-km distance, FR and AHP values are very low which indicates a relatively safe zone of forest fire possibility. A 1-km distance from the settlement areas, the FR and AHP ratios are high, indicating an unsafe zone for forest fire possibility. The settlement roads is the most effective for combatting forest fires in this area. Most of the forest fires are mainly caused by activities, so distance to settlements and distance to the road are significant parameters for forest fires.

Fig. 6 Criterion maps a LST, b elevation, c wind speed, d distance to road, e aspect, f rainfall, g slope, h distance to settlements, i NDVI,and j TPR

Fig. 6 continued

Table 1 Spatial relationships between conditioning factors and forest fire locations using the FR model

Table 1 continued

Fig. 7 Forest fire risk map based on frequency ratio (FR)

The spatial correlation between FR and AHP results

FR- and AHP-based fire risk pixel values are used to perform the polynomial regression fitting in different forest risk classes. The FFR class pixel values are collated from separate forest compartments. Very low class values of the two models are a good correlation (R2= 0.95; Fig. 9a). FR and AHP low pixel values are a good moderatedcorrelation (R2= 0.83; Fig. 9b). A moderate correlation(R2= 0.82) was observed between the FR and AHP moderate class (Fig. 9c). The correlation between FR and AHP high pixel class values are the moderate correlation(R2= 0.83; Fig. 9d). The FR- and AHP-based forest fire very high-risk class values are a good correlation(R2= 0.95; Fig. 9e). Moreover, the correlation between FR and AHP class values shows significant correlation in the forest compartments.

Table 2 Forest fire accuracyassessment (FR)

Table 3 Area (%) under different class change of each parameter (FR)

Table 4 The AHP based weights of sub-criteria for FFR mapping

Fig. 8 Forest fire risk map based on the Analytic Hierarchy Process (AHP) model

Table 5 Forest fire accuracy assessment (AHP)

Table 6 Area (%) under different class’s change of each parameter (AHP)

Fig. 9 a-e FFR zone spatial correlation value between FR and AHP models

Conclusions

In this study, we quantified forest fire risk in the Melghat Tiger Reserve forest as a process of topographic, meteorological, vegetation, and anthropogenic attributes. The north, northeast, and southwest parts are in highly susceptible fire zones because of high air and land surface temperatures, unhealthy or degraded vegetation, dry stands,and high human activities. Therefore, the Tiger Reserve should be concentrated on dense forest areas which are not fire prone, and where human activities are prohibited. To minimize fire risks in the Melghat forest, it is necessary to minimize human activities and introduce a plantation program to improve the health of the forests. Therefore, the local Government and fire management system should take actions to predict, prevent and control future fires in the very high-risk forest areas. The methods can be applied to early warning, and to fire suppression resources planning and allocation work in the forest area.

AcknowledgementsThe authors are thankful to DFO of Melghat Tiger Reserve (MTR) Forest, Forest Survey of India (FSI), and Forest Department of Maharashtra for their financial support and providing necessary data. The authors would like to thanks the Indian Institute of Technology Kharagpur and Vidyasagar University for its constant support and providing the wonderful platform for research. The authors also thanks Chai Ruihai (corresponding editor) for editing the paper.