Understanding forest fire patterns and risk in Nepal using remote sensing, geographic information system and historical fire data

Additional keywords: fire behaviour, fire danger, fire management, fire risk.

The Hindu Kush-Himalayan (HKH) region (Fig. 1a) extends over 3500 km encompassing all or parts of eight countries: Afghanistan, Bangladesh, Bhutan, China, India, Myanmar, Nepal and Pakistan. The region contains 10 of Asia’s largest river systems, these provide water and ecosystem services to the 210 million people living in mountain areas, as well as the 1.3 billion people downstream (Molden and Sharma 2013). The region is extremely fragile in terms of land cover diversity and its association with variable terrain, climate and socio-demographic interactions. Although the HKH region is rich in biodiversity, it is one of the most understudied regions of the world (Sharma and Chettri 2005). The Intergovernmental Panel on Climate Change (IPCC) has recognised the HKH region as a ‘data deficit area’ (IPCC 2007). In the last few decades, the HKH region has undergone rapid economic, social and environmental changes (Uddin et al. 2015). Under such transitions, the land use and land cover are experiencing changes due to natural and anthropogenic causes. It is clear that land cover changes in the HKH is a driving factor for changes in the ecosystems and their services (Koschke et al. 2012). In this mountainous region, forests play a key role in watershed protection, soil conservation and biodiversity maintenance (Acharya et al. 2011).

Fig. 1.  (a) Location of study area. The black boundary represents the Hindu Kush–Himalayan mountain region. (b) Physiographic zone of Nepal. (c) Elevation of Nepal. (For colour figures, see online version available at http://www.publish.csiro.au/nid/17.htm.)

Forest fire due to natural and anthropogenic factors (Adab et al. 2013) causes economic losses to people in this region and increases the emission of carbon that influences climate change (Chowdhury and Hassan 2015). Catastrophic forest fire causes the destruction of large areas in many countries in Europe, America and Australia (Bonazountas et al. 2005). Though the impacts are relatively low compared with other major disasters, like floods, landslides and earthquakes, forest fires have direct and indirect impacts that include death, and damage to buildings and infrastructures, as well as an adverse effect on people’s health (Kochi et al. 2010; Stephenson et al. 2013; Doerr and Santín 2016; Martin 2016). In tropical, mountainous countries, fire is used as a land management tool to clear forested land for agriculture (Biswas et al. 2015). Human activities in the forest also increase the incidence of forest fire (Kunwar and Khaling 2006). In the HKH region, forest fire is one of the major factors affecting the forest ecosystem of the region. Understanding the causes and spatiotemporal patterns of forest fires is key to the development of an effective fire management plan. Forest fire in Nepal burns ~400 000 ha of land annually (Bajracharya 2002). Nepal predominantly faces forest fires during the dry season from November to June each year, with the highest concentration occurring from March to May. They are assumed to be the major driver of forest degradation, due to the unplanned burning of the undergrowth in the forest affecting its natural regeneration. Besides the physical impact in the forest, fire also has an adverse effect on the local peoples’ livelihood, by limiting the sustainable availability of forest resources and the direct damage of public assets and infrastructure. There is evidence that number of forest fires is increasing in Nepal, affecting natural vegetation as well as causing a major destruction of human settlements (Parajuli et al. 2015). Tropical, subtropical, and temperate forests of Terai, Siwalik and mid-hills, particularly those with a large distribution of Shorea robusta and Pinus roxburghii, face destruction due to forest fires every year. The main reason for the high incidence of forest fires in these zones is the hot and dry weather and proximity of the forests to human settlements, roads and agricultural land. Of the total forest fire incidences, 58% are caused by deliberate burning by grazers, poachers and non-timber forest product collectors, 22% are caused by negligence and 20% occur by accident (Kunwar and Khaling 2006).

Apart from anthropogenic interference, variable climatic pattern and prolonged winter drought are other causes of forest fires in Nepal. Due to the prolonged drought in 2009 fire incidence increased significantly, causing 41 fatalities and extensive destruction of human settlements and forests (GON 2013). Such incidents have highlighted the need for an efficient forest fire risk assessment, warning and monitoring system to improve the management of forest fires. There have been some studies analysing fire regimes in specific areas of Nepal (Bajracharya 2002; Kunwar and Khaling 2006; Wagle 2014), but no studies have been conducted on the spatio-temporal distribution of fire incidence or on fire risk covering the whole country. To fill this gap, this study analyses the spatio-temporal patterns of forest fires using active fire records (Giglio et al. 2003) from multi-temporal data of Moderate Resolution Imaging Spectroradiometer (MODIS) on board Terra and Aqua satellites from 2003 to 2013.

Earth observation data and models have been widely used for fire monitoring, danger forecasting and risk mapping. Geospatial models have been used in some parts of the world to map fire risk indices (Jaiswal et al. 2002; Saglam et al. 2008; Adab et al. 2013; Mohammadi et al. 2014; Sivrikaya et al. 2014). These models integrate various factors related to vegetation, topography and proximity of forests to human activities to estimate risk. MODIS data has been widely used to map fire severity and burnt area (Chu and Guo 2015; Chu et al. 2016) and develop danger forecasting models (Akther and Hassan 2011; Chowdhury and Hassan 2013, 2015). These models used remote sensing data to generate model parameters, for example surface temperature, vegetation indices and moisture indices (Leblon et al. 2001; Aguado et al. 2003; Leblon 2005; Bisquert et al. 2014). In this paper, we have generated a fire risk map using topographic structure, land cover, surface temperature and anthropogenic parameters. MODIS data has been used to generate land cover and surface temperature parameters while Digital Elevation Modelling (DEM) was used for topographic parameters. Spatial proximity analysis with road and settlement location was used to generate distance parameters representing human factors. Weights were assigned to different parameters in the model based on the historical fire data analysis.

The study area covers the whole of Nepal, which falls between latitude 26°22′N and 30°27′N and longitude 80°40′E and 88°12′E. Nepal shares an international border with China to the north, and India to the south, east and west (Fig. 1a). The total land area of the country is 147 181 km². It is divided into five physiographic zones – high mountains, middle mountains, hills, Siwalik and Terai (Fig. 1b), and administratively into 75 districts and 4057 village development committees (VDCs), which are grouped into five development regions – far western, mid-western, western, central and eastern (Uddin et al. 2015). Nepal predominantly exhibits mountainous terrain with elevation ranging from 58 m above mean sea level (amsl) in the southern plains to 8848 m amsl at Mount Everest in the north (Fig. 1c). Out of the total geographical area of 147 181 km², 39.6% is covered by forests. A distinct altitudinal gradient can be observed in the distribution of forest ecosystems, starting from less than 1000 m and exceeding 4000 m. The broad forest types in Nepal are tropical forests (below 1000 m), subtropical broadleaved forests (1000–2000 m), subtropical pine forests (1000–2200 m), lower temperate broadleaved forests (1700–2700 m), lower temperate mixed broadleaved forests (1700–2200 m), upper temperate broadleaved forests (2200–3000 m), upper temperate mixed broadleaved forests (2500–3500 m), temperate coniferous forests (2000–3000 m), subalpine forests (3000–4100 m), alpine shrub (>4100 m) (GON 2014). Forests are the main source of timber and fuel wood, and also a major source of fodder for livestock.

Nepal is home to 35 forest types, 75 vegetation types and 118 ecosystems, along with four global biodiversity hotspots (Chettri et al. 2008). However, these ecosystems are facing challenges due to natural and anthropogenic induced climate change. An increase of 0.06°C to 0.12°C in mean annual temperature in most of the country was observed from 1977 to 1994 (Shrestha et al. 1999).

The amount of precipitation and its spatio-temporal distribution is determined by the annual monsoon system. There are four climatological seasons including the pre-monsoon period (March–May) that is hot and dry, the summer monsoon (June–September) is characterised by high humidity and precipitation, the post-monsoon (October–November) with reduced rainfall, and winter (December–February) that is dry and cold (Kansakar et al. 2004). Four precipitation regimes are found in the country: the central and western high mountains have the lowest level of precipitation; the Terai and Siwalik zones, distributed east to west of the country, have an intermediate level of precipitation; the central and eastern foothills of the Siwaliks and middle mountains have a moderately high amount of precipitation; and the highest precipitation areas are located on the hilltops of the middle mountains and at the foothills of the high mountains (Kansakar et al. 2004). Due to climate change, temperatures in the Terai and mid hill region of Nepal are gradually increasing. Temperatures in this region have increased ~3°C during the summer months over the last few decades and are expected to rise by 1.4°C, 2.8°C and 4.7°C by 2030, 2060 and 2090 (Pradhan et al. 2013). Changes in temperature and precipitation regimes through Nepal in coming years are expected to increase the number of forest fires (Negi et al. 2012).

Active fire data from the MODIS instrument on NASA’s Terra and Aqua satellites dating from 2003 to 2013 was used. MODIS fire observations are made four times every day from the Terra (1030 hours and 2230 hours) and Aqua (0130 hours and 1330 hours) satellites. MODIS provides two standard fire products as part of the data package; the active fire product specifies the location of the fire and the burnt area product shows the extent of the burnt area at the time of satellite overpass. The active fire product mainly uses the brightness temperature from MODIS thermal bands (band 21, 22 and 31). The location of fire corresponds to the centre of a 1 × 1 km fire pixel representing one or more fire incidences occurring within the pixel area. A contextual algorithm compares each pixel with its background pixels to calculate temperature anomalies based on a preset threshold, therefore identifying pixels that are on fire (Giglio et al. 2003; Giglio 2010). The fire information includes the confidence at which the pixel is determined to be on fire, calculated based on the brightness temperature of the target pixel; the difference between the brightness of the target pixel and the mean of the surrounding pixels; the number of adjacent cloud pixels; and the number of adjacent water pixels. The final fire product also masks out false detection based on sun glint, desert boundaries and water (Giglio et al. 2003). During the period from 2003 to 2013, 28 449 fire incidences were recorded by MODIS with confidence levels of 1–100%. To avoid false identification we have filtered out the records with less than 50% confidence level, which reduces the total number of fire records to 21 147. The distribution of these fires across different land areas is given in Fig. 2a. From this subset, 12 269 fire incidences that occurred within forest, grasslands, shrublands and outside protected areas were selected for the analysis. Records of fires within the protected areas and agriculture lands were discarded because most of the fires in these areas are managed fires.

Fig. 2.  (a) Distribution of fire incidence in different land cover classes inside and outside protected areas. (b) Total annual forest fire in different year. (c) Percentage of fire incidences occurring in different months of the year.

Vegetation cover is the most important factor in forest fires as it represents the availability of fuel (Roy 2003). For the vegetation factor, we have used land cover data for Nepal for the year 2010 (Uddin et al. 2015). The land cover data was prepared from the Landsat TM images using an object-based image analysis method. Temperature is the second most important factor for fire ignition and spread (Roy 2003). We used Terra MODIS-derived 8 day land surface temperature (LST) data (Wan 2008) from 2003 to 2013. At first, we generated the monthly maximum LST from MODIS and then averaged the monthly maximum LST for the pre-monsoon period (March–May) for each year. This data was then averaged for the study period (2003–2013) to generate the LST factor for input into the model. Topographic characteristics are also very important factors for fire spread (Jaiswal et al. 2002). Topographic parameters (i.e. elevation, slope, and aspect) were generated from the Shuttle Radar Topographic Mission (SRTM) digital elevation model (DEM) (CGIAR-CSI 2008). Anthropogenic factors are represented by proximity of fires to settlements and roads, these were generated with the ArcGIS distance tool using settlement and road cover data of Nepal.

MODIS active fire data were used to analyse the spatial and seasonal patterns of fire incidence in Nepal. The data were analysed to reveal the patterns in forest fires across physiographic zones, development regions, forest types, climatic seasons and social elements such as human settlements. To explore the seasonal and temporal patterns, year and month wise forest fires, peak fire month and number of fire days were calculated per VDC.

Forest fire risk is the likelihood in an area for ignition and spread of fire (Jaiswal et al. 2002). Various indices have been proposed in literature to estimate fire risks using natural (vegetation, humidity, elevation, slope, aspect) and anthropogenic (proximity to settlements and roads) parameters (Jaiswal et al. 2002; Filippidis and Mitsopoulos 2004; Saglam et al. 2008; Cáceres 2011; Adab et al. 2013; Mohammadi et al. 2014; Sivrikaya et al. 2014; Zhang et al. 2014). These indices are grouped into three categories: structural and long-term indices; dynamic and short-term indices; and combined indices (Adab et al. 2013). Chuvieco and Congalton (1989) introduced a structural fire index (SFI) using vegetation, elevation, slope, aspect and proximity to roads to estimate fire hazard. This method was later adopted by others when estimating SFI (Jaiswal et al. 2002; Saglam et al. 2008; Sivrikaya et al. 2014). In this study we used vegetation type (land cover), elevation, slope, seasonal mean land surface temperature and proximity to settlements and roads to calculate the fire risk index (FRI). The general expression of the FRI is given in Eqn 1:

where W_i is the relative weight of a variable and C_i is the rating for different classes within each variable. We used variables for vegetation (land cover), temperature (surface temperature), topography (elevation, slope), and anthropogenic (proximity to settlements and roads) factors in the equation as they all contribute to fire behaviour (Roy 2003). The relative weights for variables were chosen based on the literature (Chuvieco and Congalton 1989; Jaiswal et al. 2002; Saglam et al. 2008; Adab et al. 2013; Sivrikaya et al. 2014) and ratings among the different classes within each variable were chosen based on historical data analysis. Higher ratings were assigned to classes with relatively higher occurrence of historical fire incidence within that class, compared with other classes. The expression used in this study is given in Eqn 2:

where, LCR is land cover rating, TR is temperature rating, SDR is settlement distance rating, RDR is road distance rating, ER is elevation rating and SLR is slope rating. The weights and ratings for different variables and classes are given in Table 1. To assign ratings for each variable, we used the exploratory data analysis method. For example, most of the forest fire occurred in the broadleaved closed forests followed by broadleaved open forest, shrubland, grassland and needle-leaved forests. The ratings for these land cover classes were assigned accordingly. A similar approach was taken to assign ratings for the other variables: surface temperature, distance from settlements and roads, elevation, and slope. The availability of fuel and high summer temperatures are the key factors in fire risk (Roy 2003). The availability of fuel is indicated by the abundance of flammable materials on the forest floor. When analysing the fire risk, the fuel ratings are assigned based on vegetation types (Xiangwei et al. 2011). Considering this, we have assigned the highest weight to the fuel factor (i.e. LCR) in line with other previous studies. Temperature is assigned the second highest weight. Globally, 90% of fires are ignited by humans (Arndt et al. 2013) and this also holds true for Nepal (Kunwar and Khaling 2006). Hence, we used proximity of fires to settlements and roads as an indicator for human activities and assigned the third highest weight to these factors. Finally, topographic factors (elevation, slope, aspect) are also included in the equation as they influence fire behaviour. Wind, precipitation and temperature are influences by elevation and all can affect fire ignition and spread. Slope influences how fast fire can rise and aspect determines the amount of solar radiation present that influences temperature (Biswas et al. 2015). We only used elevation and slope as topographic factors, because analysis of historical fire data showed no significant variation in fire incidence for the different aspects.

Table 1.  Weights and ratings assigned to different variables and classes for fire risk calculation
amsl, above mean sea level

From 2003 to 2013, a total of 12 269 forest fires were recorded by MODIS sensors in Nepal in the unprotected forest, grass and shrub land with a confidence level of 50% or more. The frequency of fires was higher during recent years, that is, after 2009 (Fig. 2b). Fires start to occur in winter (December–February) with most of the fires (~89%) occurring during the pre-monsoon season (March–May) when the weather is hotter and drier (Fig. 2c). Using monthly fire incidences within VDC boundaries, a map of peak fire month per VDC was prepared (Fig. 3a). From this map, it is apparent that peak fire months vary across different landscapes of Nepal. April is the peak fire month in most parts of the country, with the exception of some areas in the eastern development region and at high altitude (Fig. 3a). In the eastern parts of the Terai, March is the peak fire month while in western Terai and most of the Siwalik hills, fire incidences peak in April. We observed a higher number of forest fire incidences in the year 2009, which corresponds with the driest year of the study period experiencing 50% less rainfall than average during the winter season, followed by the late start of the monsoon period for the year (Wang et al. 2013). The annual length of the fire season was calculated using the first and last fire incidence that occurred for individual districts. These annual lengths of the fire season were then averaged, for the 2003–13 period, to generate a map of all districts showing their mean number of fire days. The districts falling within the mid and far western and central parts of the Terai and middle mountains experienced more fire days, whereas the eastern parts of these zones observed fewer fire days (Fig. 3b). This is related to the precipitation regime in Nepal, where the monsoon season has a later onset and shorter duration in the western part of Nepal when compared with the eastern part of the country (Kansakar et al. 2004).

Fig. 3.  Spatial distribution of peak fire months and number of fire days based on incidence data of 2003–2013: (a) peak fire months of different village development committees (VDCs); (b) number of fire days by district.

The analysis of the forest fire profiles across the five physiographic zones of Nepal revealed that highest number of forest fire incidences were recorded in the Siwalik zone (~37%) followed by low hill (~26%), Terai (~20%) and middle mountain (~13%). Very few fire occurrences were recorded in the high mountainous zone (Fig. 4a). About 80% of forest fires occur in the broadleaved forest. The broadleaved closed forest alone accounted for ~60% of total forest fires (Fig. 4b). This is because the tropical broadleaved forest experiences heavy leaf fall during summer (i.e. March–June) which results in the accumulation of a large amount of leaf litter, fuelling frequent and prolonged occurrences of fire during summer. The incidence of fire increases with the rise in surface temperature. Approximately 72% of fires occurred when the surface temperature rose above 30°C (Fig. 4c).

Fig. 4.  Distribution of forest fire incidences across different (a) physiographic zones; (b) temperature zones; (c) land cover classes; (d) elevation ranges; (e) slope ranges; (f) aspect classes; (g) proximity to settlements; and (h) proximity to roads.

Comparatively a higher number of forest fires occurred in lower elevation areas. About 65% of the fires were recorded in the areas below the elevation of 1000 m, whereas 85% of the incidences occurred in areas below 2000 m (Fig. 4d). Hardly any fires were recorded above 4000 m. Fire incidences also varied across different slopes. More than 30% of fires were recorded in plain lands (i.e. slope less than 5%). In the hill region, a slightly higher number of incidences were recorded in moderate slope areas (i.e. 15–35% slope), while very few incidences were recorded in areas with a slope of more than 50% (Fig. 4e). It is assumed that south facing slopes are susceptible to fire, but no significant patterns were observed in the fire records for the different aspects in Nepal (Fig. 4f).

The effect of human influence on the forest fires is indicated by the incidence of fires in proximity to settlements and roads. It is apparent from the fire records that more than 41% of fires in Nepal are recorded within 1 km of a settlement (i.e. within the same pixel of the satellite image) (Fig. 4g). The number of fires decreases with the distance from settlements. Similar patterns were also observed when fire locations were analysed in relation to the proximity of roads. About 40% of fire incidences were recorded within 1 km of a road (i.e. with one pixel) (Fig. 4h). The number of fires decreases with the distance from roads. Analysis of these two proximity factors indicates a high degree of influence by human activity on the incidence of fires.

A forest fire risk index map was generated using Eqn 2. At first, we classified raster maps representing each of the parameters into the classes described in Table 1. The classified maps were then converted into rating maps for each parameter based on the relative rating for each class, as described in Table 1. The rating maps were then combined using Eqn 2 to generate the risk index map. The risk index generated by Eqn 2 was categorised into four risk classes, namely low, medium, high and very high (Fig. 5a). The classes are assigned using the natural breaks method, based on the Jenks Natural Breaks algorithm (Jenks 1977) with manual refinement in ArcGIS software (ESRI 2016). To validate the risk index we overlaid the fire incidences from 2003 to 2013 on the risk class map (Fig. 5a). The analysis shows that more than 80% of the fires occurred in high and very high risk areas justifying the model output.

Fig. 5.  (a) Forest fire risk map. (b) Districts of Nepal classified by fire risk class based on percentage of area under high and very high risk zones. (c) VDCs of Nepal classified by fire risk class based on percentage of area under high and very high risk zones.

The district and VDC level risk maps were generated by overlaying district and VDC boundary maps with the risk class map. Districts and VDCs were ranked based on the percentage of district or VDC area within the high and very high risk zones. The resulting maps were then classified into five risk categories: very low (<5%), low (5–14%), medium (15–29%), high (29–44%) and very high (≥45%) (Fig. 5b, 5c). The analysis shows that 31 out of 75 districts fall into high and very high fire risk classes while 22 districts fall into low or very low risk classes. Most of the districts in high and very high risk classes are located in the Terai and Siwalik zones. More districts in the western and central development regions fall in the high and very high risk classes than in the eastern development region (Fig. 5b). The low or very low risk districts are in the hills and high mountain zones. A list of districts with very high risk to forest fire is defined in Table 2. From the VDC wise risk map (Fig 5c), it is apparent that the districts in the mid and far western development region and in the central region within the Terai and Siwalik zones have a higher number of VDCs in the high and very high risk classes. The department of forest in Nepal has very limited resources for the monitoring and management of fire. The ranking of districts and VDCs based on fire risk will allow them to effectively mobilise resources in the districts and VDCs with higher risk and give more preference for distribution of fire-fighting equipment to local forest offices based on the fire risk ranking of districts.

Table 2.  List of districts with very high forest fire risk

The MODIS active fire data has some inherent limitations. The Terra and Aqua satellites pass over most locations on Earth twice a day. Combined, they detect fire incidences four times in 24 h. The Terra satellite passes the same location in the morning at ~1030 hours and in the evening at ~2230 hours. The Aqua satellite passes the same location in the afternoon at 1330 hours and at night at 0130 hours. As a result of this timing, short duration fires that occur between satellite overpasses are missed. MODIS active fires are detected by calculating the thermal anomalies on a pixel 1 × 1 km in size and the location of the fire is reported as the centre of the pixel. Due to the coarse resolution, multiple fire incidences within one pixel area are reported as a single incidence. The detectability of a fire depends on the flame size but the surface heterogeneity, existence of fire in surrounding areas, clouds, heavy smoke and sun glint all affect the detectability of fires (Giglio et al. 2003, 2008; Giglio 2010).

In this study we analysed historical forest fire data to identify the spatial and seasonal distribution of forest fires in Nepal. The analysis provides a reasonable estimate of the occurrence of fires in various land covers, topographies and near different settlement features. Although MODIS fire data has some inherent limitations as described in section 4, the analysis will be useful for identifying fire regimes in Nepal. Using the vegetation, temperature, topography and human activity factors, we have estimated the fire risk index for Nepal. The fire risk maps will be helpful to understanding the pattern of risk across the country. The district level risk maps will be advantageous for forest managers at the national and regional levels to allocate resources among districts based on their relative risks, while the VDC level risk maps will help the district managers to develop strategies for fire management within the district and distribute firefighting equipment to areas that are in higher risk zones. The risk maps will also help in the setting up of effective firefighting infrastructure.