Peatland is a wetland ecosystem in which the production of organic matter from dead plants is higher than its decomposition. Several factors influence peatland formation, including climate, humidity, topography, and geology [1] . The majority of peatlands are located in temperate and boreal zones under low-temperature conditions. However, regional environmental and topographic conditions have resulted in the formation of tropical peat swamp forests in Southeast Asia, Southern Africa, South America, and Central America.

Peatlands provide various ecosystem services. Peatlands store 500 - 700 G ton of carbon, yet it only covers 3% of the Earth’s surface [2] . The majority of this carbon is stored in temperate and boreal peatlands, but tropical peat swamp forests also store a significant amount of carbon at around 80 - 90 G ton, 69 G ton of which is stored in Southeast Asia. Furthermore, peat swamp forests in Southeast Asia provide other ecosystem services such as climate regulation, water supply, and supporting high biodiversity [3] .

When tropical peatland starts to burn, it will release a significant amount of carbon into the atmosphere (243 ton per hectare [4] ). Peatland fires have changed peatland function in Southeast Asia from carbon sink to a carbon source [5] . Moreover, in recent years, persistent peatland fires have been identified as a hazard with serious effects on human society. The total economic, social, and ecological damages and losses due to fires in Indonesia were estimated to be at least 16.1 billion USD in 2015 [6] . The effect of fires on peatland ecosystems may persist for a long time [7] .

Peatland fires have a long history, and they are not a new phenomenon in Southeast Asia. Major fire events occurred in Kalimantan in 1846, 1902, 1915, and 1972, all of which were El Niño Southern Oscillation (ENSO) years [8] . El Niño reduced precipitation drastically [9] , which makes peatland more susceptible to fire due to lowered water table [10] . Local communities in East Kalimantan and South Sumatra have traditionally conducted slashing and burning in peat swamp forests along the river bank and forest edges to convert the forest to agricultural land ( [11] [12] ).

However, fire has become a more frequent and severe problem in Southeast Asia since the late 1990s [13] . Visibility records from the airports in Sumatra and Kalimantan since the 1960s indicate higher fire frequency after industrial plantations or the Mega Rice Project in Kalimantan began [14] . Twenty percent of peat swamp forests in Malaysia Peninsular, Sumatra, and Kalimantan were transformed to industrial plantations by 2010 [15] . Fire was a tool for land preparation after logging to create plantations by palm oil and industrial forest companies [16] . Furthermore, the water table was lowered by the creation of canals to grow oil palms or acacia plantations and to create large-scale rice paddy fields, which increased fire risk [17] . Industrial plantation concessions contributed to fire ten times higher than selective logging concessions ( [18] ).

Fires may be prevented by sustainable management practices [19] . Fire occurrences in Southeast Asian peatland depend on land cover [20] . Many studies have reported that fires occurred more frequently in deforested areas than in forests or oil palm and acacia plantations, fires originated and spread from deforested areas more often than oil palm plantations and settlements in Central Kalimantan [21] . The sources of fires have shifted from peat swamp forests due to slash and burn to deforested areas since 2002 due to the Mega Rice Project [22] . In Riau, most of the burned area was a deforested area expanded by a previous fire or created due to the failure of industrial plantation development [23] . To prevent recurrent peatland fires, factors that affect fire occurrences besides climate should be elucidated.

Despite deforested areas consisting of various land cover types, such as shrubland regenerated after a fire [24] , rice paddy fields, coconut and rubber cultivation, young oil palm plantations, and bare soil, most of previous studies have not classified these land cover types as they used moderate-resolution satellite images to cover large regions. These deforested land covers may contribute to fire occurrence differently. In Jambi Province, Stolle et al. [25] demonstrated that few fires occurred in rice paddy fields, coconut, plantations, grassland, and rubber cultivation. In contrast, recurrent fires occurred in shrubland, and ferns dominated vegetation that grew after fire [26] . Young oil palm and acacia plantations in South Sumatra were susceptible to fires [27] . Understanding which types of deforested land cover are more prone to fire is necessary for sustainable development of peatland.

Although all land belongs to the Indonesian government, land is used by different landholders. Concession holders may not control all the land within their management area as migrants and local communities may claim land tenure in the concession area and forests [28] . Land discrepancies among existing conditions, concession holders, and land management are major problem in peatland management. Owing to land discrepancies, the government has difficulties in identifying the responsible parties for peatland fire [29] .

Even if the land cover type are the same, land management may influence fire occurrence. In Kalimantan, fewer fires occurred in protected forest areas than in forests allowed to selective logging or conversion to a plantation [30] .

Fire occurrence may be also affected by landholder type. Land utilized by smallholders has generally been managed intensively and more protected against fires than concession areas [31] , because smallholders cleared their land at smaller scale than concession companies. Oil palms are planted by either a registered large-scale concession company, an unregistered company, or smallholders in Indonesia [32] . The registered companies implemented a zero-burning policy to obtain a sustainable management certification. This certificate is mandatory for all palm oil companies in Indonesia. In contrast, unregistered companies do not follow the policy and a company has used fire to develop plantations in an Indonesian national park [33] .

The proximity to a road or canal could be another important factor affecting fire occurrence. Road intensity affects spatial fire distribution [34] . In Jambi, forests within 1 - 5 km of the road suffered fires almost five times more than forests over 20 km away from the road [25] . Similarly, in Kalimantan, most forest fires occurred within 5 km from the forest edge [30] , and many fires occurred near canals [22] . Roads and canals are the main methods of access to plantations. Additionally, canal may also lower the level of the water table by drainage, thus creating a fire-prone environment.

This study aims to analyze relationship between fire occurrence and land cover type, land management, landholders, and proximity to roads and canals. We focus on Riau Province in Sumatra Island, Indonesia, because Riau has extensive peatlands, frequent fires, and has experienced rapid deforestation. The year we selected was 2014 due to severe impacts of fire (Figure 1) on the economy, social, and ecological aspects in Riau, where damage and loss totals an estimated 935 million USD [35] . We focus on four questions:

1) Are fire occurrences more frequent in shrubland than other land cover types?

2) Do fires occur more frequently in peat swamp forests allowed to convert to a plantation or other uses than protected peat swamp forests?

3) Is smallholder cultivation better for fire prevention than registered and unregistered companies?

4) Does closer proximity to road and canal result in more frequent fires?

Riau Province is located on the eastern coast of Sumatra island, stretching from the Barisan Hills downwards to Malacca Strait (2˚35'N - 0˚58'S, 100˚13'E - 103˚50'E). Riau has a tropical climate, with annual mean precipitation and temperature of 2400 mm and 26˚C, respectively, between 2012 and 2014 [36] .

A map of land cover after a severe fire between January and March 2014 was created using post-fire Landsat 8 images, acquired from April to November 2014. Landsat 8 has a spatial resolution of 15 × 15 m [37] . The images from the period with less than 50% cloud cover were used [38] . Satellite images were downloaded from the US Geological Survey National Center for Earth Resources Observation and Science via the Global Visualization Viewer (GLOVIS) data portal (http://glovis.usgs.gov/). This study was conducted with Landsat 8 pre-processing (atmospheric correction), land cover classification, and an accuracy check of land cover classification [39] .

Atmospheric correction was conducted to reduce atmospheric distortion. To enhance image information Quantum GIS 2.14.2 Essen was used by transforming radiance at the sensor into surface reflectance values [40] . Physics-based derivation of surface and atmospheric properties of hyperspectral and multispectral data was presented by image enhancement process. This is based on atmospheric radioactive transfer, input of atmospheric parameters, and calibration of the instrument accuracy. Spectral differences were enhanced with the algorithm that divided spectral band (numerator) by another band (denominator) [37] .

A supervised classification through maximum likelihood algorithm method in ArcGIS 10.2 software to classify land cover into 11 types (Table 1) was used. Training areas were used to define the spectral reflectance patterns of each land cover type in supervised classification. Classifier would use the pattern of training area to group the pixels of certain category with the same spectral patterns [41] . Pixel of an unknown category has a certain probability of belonging to a particular category in the maximum likelihood algorithm classifier. All categories had equal probabilities follows the Gaussian (normal) distribution function.

For the training areas, ground truth data were used. Geographical coordinates and land cover types were recorded in the field for 1301 sampling points during 2013-2015. To avoid pseudoreplication, stratified random sampling was conducted by the Sampling Design Tool [42] in ArcGIS 10.2. Ten percent of sampling points that were at least 5 km apart from each other were randomly sampled for each land cover type. The sampling points were not sufficient for some land cover types (shrubland, settlement, mangrove, and water body), so points were added using Google Earth images acquired from 2013 to 2015.

A training area was created for each sampling point with the assistance of visual inspection of Landsat images through displaying RGB combination (bands 654) and the image from Google Earth. The training area was checked by the 2013 land cover map (the Ministry of Forestry and Prayoto unpublished data).

Accuracy assessment is an important step due to the land cover maps that derived from remote sensing commonly contain various errors as a result of method of satellite data capturing or the classification procedure [43] . The accuracy assessment usually used error matrix that represents the number of sample units (i.e., pixels, clusters of pixels, or polygons) in a set of numbers of rows and columns assigned to a particular type, relative to the actual type in verification data. For verification data, stratified random sampling was conducted for the remaining ground truth data that was not used for creating the training area. The Sampling Design Tool [42] was used to sample 50% of the points that were more than 5 km apart from each other for each land cover type. Google Earth was used to add additional data for the categories with few sampling points. User’s accuracy was calculated by dividing the number of correctly identified ground truth points by the total number of verification points.

To elucidate relationship between fire occurrence and land cover type, pre-fire land cover was identified for areas burned by the fire from January to March 2014. The burned area in the post-fire land cover map can be classified into two types: 1) areas burned by the fire in January to March 2014; and 2) areas burned before January 2014 or after March 2014. The former burned area was identified by the presence of Moderate Resolution Imaging Spectroradiometer (MODIS) fire hotspots during January 29^th to March 28^th. Hotspot data were downloaded from the National Aeronautics and Space Administration via Fire Information for Resource Management System (FIRMS) data portal (http://firms.modaps.eosdis.nasa.gov/download/request.php/). MODIS fire hotspot data shows the coordinates of the center of 1 × 1 km pixel where persistent fire was detected from a MODIS image using an algorithm ( [21] [44] ). For the areas burned by the fire in January to March 2014, the pre-fire land cover was determined through visual interpretation of Landsat 8 images acquired from September to December 2013, with Google Earth data as supportive data.

According to Indonesian forestry laws, peatland in Riau was divided into conservation forest, protection forest, production forest (limited production forest, regular production forest, and convertible production forest), and non-forestland. Conservation forest was assigned for maintaining biodiversity and ecosystems, and can be utilized for research purposes. Protection forest was assigned for protecting water systems, preventing flooding, soil erosion, seawater intrusion, and maintaining soil fertility, and can only be utilized for research purposes and non-timber forest products. Limited production forest was assigned for selective logging. Regular production forest was assigned for producing wood through the clear-cutting system, planting, and harvesting industrial forests. Convertible production forest was assigned for wood production or conversion to non-forestland. Non-forestland was assigned for non-forestry activities, such as agriculture [38] . A land management map for 2011 was obtained from the Ministry of Forestry.

Based on land management, the government has granted forestry concession and non-forestry concession. Industrial forest plantation concessions were granted for logging and acacia plantations in regular and limited production forests. Non-forestry concessions (cultivation right and forest release area) were granted for oil palm plantations in Riau on convertible production forest and non-forestland.

We categorized landholders into registered companies, unregistered companies, smallholders, cooperation between a company and smallholders, and unidentified landowners [23] . We identified the landholder of the area that was burned by fire in January to March 2014 using pre-fire land cover information, concession boundary data in 2013 that were issued by Ministry of Forestry and National Land Agency, and images from Landsat and Google Earth that were captured in 2013. Registered and unregistered company plantations exhibit more orderly clusters, with regular canals, roads, and palm oil mills. If the regular pattern was located inside the concession and covered by acacia, oil palm, coconut, or sago palm, it was categorized as a registered company. If the regular pattern of oil palms was located outside the concession, it was categorized as an unregistered company. Smallholder plantations were categorized by land parcels of irregular shape, varying size and direction, and covered by oil palm, coconut, and sago palms. The boundaries of those grids were constructed by visual interpretation. These patterns are visible on the Landsat 8 and Google Earth images. Usually, acacia plantations were developed inside the concession area; however, occasionally the companies cooperate with the local community to plant acacia outside their concession area [45] . If an acacia plantation was found outside the concession, these were categorized as cooperation between a company and smallholders. For shrubland, we could not identify landholders by our method as migrants may encroach the shrubland under concession.

To elucidate relationship between fire occurrence and land cover type, land management, landholders, and accessibility, geospatial analysis was conducted using ArcGIS 10.2. The burned area map was overlaid onto the land cover map, land management map, and the map of landholder. We conducted buffer analysis for the roads and canals to test whether proximity affects fire occurrence. The probability of a burned area was compared between areas with different proximity to roads and canals. A buffer area was created every 1 km up to 5 km for the canal and every 1 km up to 10 km for the road.

Data of forest patch sizes were analyzed with R statistical soſtware. Analysis of variance (ANOVA) was used to determine significant differences between forest group at confidence level of p < 0.05.

The user’s accuracy of each land cover was 65% - 100% (Table 1) and the producer’s accuracy was 75% - 100%. The overall accuracy and the Kappa coefficient was 83% and 0.81, respectively. It indicated that image classification was good [46] .

Plantations and agricultural land covered 47% of peatland (Table 2; Figure 2). Oil palm plantations were the most dominant type (17%), followed by coconut (13.9%), and acacia (13.5%). Peat swamp forests covered only 29.2% of peatland.

Peatlands in Riau were categorized into conservation forest (Table 3; 5.7%), protected forest (0.3%), production forest (77.5%), and non-forestland (16.5%). Half of the peatland has been granted concessions, namely cultivation rights (10.2% of peatland), industrial forest plantation (30.4%), and release of forest area (8.5%).

Peatland utilization is expected to be determined by the type of land management and concessions. However, there was a discrepancy between actual peatland utilization and the types of land management and concessions. For instance,