Evaluating the social and environmental factors behind the 2015 extreme fire event in Sumatra, Indonesia

Fires in Indonesia release excessive carbon and are exacerbated during drier El Niño years. The recent 2015 fires were affected by an extended drought caused by a strong El Niño event. This led to severe haze conditions across Southeast Asia, resulting in adverse socioeconomic and health impacts. Here, we evaluate the social and environmental factors that contributed to the 2015 extreme fires in Riau, Jambi and South Sumatra. We developed proxy variables for plausible drivers of fire which contribute either as a predisposing condition or as an ignition source for fires. We evaluated how these variables influenced fire count at an administrative regency-level and fire occurrence at a pixel-level (1 km2). We used generalized linear mixed effect models to model fire count at the regency-level and boosted regression trees to model fire occurrence at the pixel-level. Rainfall, slope and population density were the most important variables predicting fires at both levels. Economic variables such as the proportion of small-scale (<10 ha) and medium-scale (10–100 ha) plantation landholdings, and the reported use of fires to clear agricultural lands in villages were important in explaining fire count at the regency-level. At the pixel-level, distance from roads and the number of recorded burns over peatlands were important in explaining fire occurrence. The main influence of rain on fires corroborates with previous studies, and highlights the importance of establishing an early warning system for droughts to better prevent and manage future extreme fire events. Mitigation efforts for future fires, especially during El Niño years, can focus on identifying high-risk areas using environmental data on rainfall, slope, peatlands, and previously burnt peat areas, as well as social data related to population density, access to roads, extents of small- and medium-plantation landholdings, and village-level propensity to burn land for agriculture.


Introduction
Fire is an inexpensive and convenient tool for agricultural land clearing in the rural tropics and is widely used across Indonesia by smallholders and companies (Ketterings et al 1999, Anderson and Bowen 2000, Carmenta et al 2011. Drought years associated with the El Niño Southern Oscillation (ENSO) and Indian Ocean Dipole exacerbate fires, resulting in extreme fire events in Indonesia in 1997, 2006(Field et al 2016. However, Indonesia's land-use policies, such as the logging and clear-cutting of tropical forests and drainage of peatlands for industrial plantations (Miettinen et al 2012), have produced landscapes that are susceptible to extreme fires even in non-drought years (Siegert et al 2001, Field et al 2009, Gaveau et al 2014, Field et al 2016. These fires contribute to increasing atmospheric carbon dioxide concentrations (Huijnen et al 2016), and result in adverse socioeconomic and health impacts from haze (Marlier et al 2015a). In the most recent extreme fire event of 2015, the thick choking haze extended to Singapore, Malaysia and Thailand, and cost an estimated 100 000 deaths within Southeast Asia (Koplitz et al 2016) and economic losses of USD16.1 billion (World Bank 2015) in Indonesia alone.
Fires are a consequence of a combination of predisposing conditions, which define the vulnerability of the landscape to fires, and social causes, which result in fire ignitions . Some proximate factors behind ignitions in Indonesia include land clearing (Gaveau et al 2014, Marlier et al 2015b, land conflict (Suyanto et al 2004), and the use of fire for household activities. Fires are used predominantly for land clearing as it is cheaper than manual clearing, and burnt land is of higher value to more actors compared to unburnt cleared land (Ketterings et al 1999, Purnomo et al 2017. Fires have been used as a weapon in times of land-use disputes either between local and migrant communities or between companies and communities (Suyanto et al 2004, Dennis et al 2005, and are also used by rural villagers for cooking and waste disposal.
Ignition sources are just part of the picture behind fires; the predisposing conditions of the landscape propagate fires and facilitate its spread . The rapid conversion of forests and drainage of peatlands for oil palm and acacia plantations (Miettinen et al 2012, Carlson et al 2013, Abood et al 2015 made landscapes in Sumatra and Kalimantan more prone to extreme fires especially under strong ENSO conditions (Field et al 2016, Taufik et al 2017. The likelihood of major fires occurring increases sharply during prolonged dry conditions (<4 mm day −1 precipitation) (Field et al 2016) as a consequent of the subsurface hydrological drought (Taufik et al 2017). Ecosystem degradation through logging and canalization of peatlands results in higher amounts of dry biomass above-and below-ground respectively, increasing the fuel load for extreme fires to occur (Siegert et al 2001, Konecny et al 2016. Road development also enhances landscape fragmentation and increases access to forests and peatlands for conversion . Repeated fires over peatlands results in higher fern and scrub cover, increasing its susceptibility to future fires (Hoscilo et al 2011, Miettinen et al 2017. In addition, the slope of the landscape could influence the movement of fires while flatter lands correspond to where peatlands are located. The 2015 extreme fires were unprecedented in scale and impact, and several studies have looked into understanding the environmental factors (Field et al 2016, Miettinen et al 2017 and the social factors , Purnomo et al 2017 related to this event. However, none combined both social and environmental factors, or examined the influence of these drivers at different geographical scales. Evaluating how social and environmental factors influence fires at different scales could allow for more targeted responses to prevent and mitigate future fires (Schwartz et al 2015). We build on the conceptual model from  which investigated fire as a function of predisposing conditions that were mostly influenced by environmental factors, and ignition sources that were often influenced by social factors such as economic land-use. We present a mesoscale (225 000 km 2 ) assessment of the social and environmental factors of fire in three provinces of eastern Sumatra-Riau, Jambi and South Sumatrawhich played an important role in the 2015 haze (Koplitz et al 2016). Using a combination of spatiallyexplicit socioeconomic and geospatial datasets, we addressed the following research question: What were the major contributing social and environmental factors to fire counts in the eastern provinces of Sumatra in 2015 at a regency-level, and fire occurrence at a 1 km 2 pixel-level?
2. Data and methodology 2.1. Study site Our study sites are the provinces of Riau (0.5333°N, 101.4500°E), Jambi (1.5833°S, 103.6166°E) and South Sumatra (2.9789°S, 104.7584°E), with 40 regencies and ∼6 000 villages (figure 1, STable 1). These three provinces lie east of the Barisan mountain range, and cover an elevational gradient from 0 to 3 736 m. Approximately 27.8% of the land (STable 1) is underlain by peatland close to the eastern coast, with peat depth up to 8 m (Wahyunto et al 2003). These provinces were the epicenter of fire hotspots in Sumatra during the catastrophic 2015 fires. The dry season began in June with limited fire activity but reached its peak in Sumatra by September (Field et al 2016).
2.2. Data and processing 2.2.1. Fire data for 2015 We used the global monthly fire location product from the Moderate Resolution Imaging Spectroradiometer (MODIS) Active Fire Detections MCD14ML Collection 6 (NASA FIRMS 2017) to identify active fires (FireNo) between 1 June-31 October 2015 (table 1), corresponding with the major haze event. This product is derived from an active fire detection algorithm which utilizes 1 km resolution MODIS thermal bands from MODIS sensors on board the Terra and Aqua satellites (Giglio et al 2016). It is important to note that the sensors are unable to distinguish between one or more individual fires within a pixel, and that each detected fire hotspot is not always a separate individual fire but the same fire event detected multiple times a day as long as it keeps burning (Langner andSiegert 2009, Miettinen et al 2017). MODIS sensors are known to have detection errors due to the effect of sunglint and the challenges of detecting smoldering peat fires (Liew et al 2003, Tansey et al 2008. Despite its limitations, the MODIS hotspot dataset is commonly used to understand fire distribution in

Proxy variables that influence fire
We derived 18 proxy variables grouped into five categories (Conflict, Economic, Population, Forest Degradation and Biophysical) to represent the social and environmental factors which influenced either ignition sources or predisposing conditions for fire spread in our study site (figure 2, table 1). The data sources for the derivation of these proxy variables include published sources of spatially-explicit datasets and the 2014 Village Potential Survey (Potensi Desa or PODES) from the Indonesian Central Bureau of Statistics (Indonesian Bureau of Statistics 2016) (table 1, supporting information is available online at stacks. iop.org/ERL/14/015001/mmedia). The derivation of each variable differs between the regency-level and the pixel-level analyses and these differences are described in greater detail in our supporting information. Here, we briefly introduce each proxy variable. Details on the hypothesized relationship between each variable and the outcome on fire count and occurrence are found in STable 3 under our supporting information.
Our three proxy variables under Conflict included the number of reported land conflicts from January 2010-October 2015 (LandConflict), the reported status on multi-ethnicity in the village (MultiEthnicity), and the reported status of brawling incidents in the village (Brawl). We assumed that multi-ethnic villages would have a higher likelihood of local conflict. Brawling incidents reported under PODES are not specific to land conflicts and can refer to any mass fights that occurred in the village. Since fires are used as weapons in conflict, we hypothesize that higher occurrences of conflict would result in more fires.
We included six proxy variables related to land-use under Economic factors-wood fiber concessions (WFC), oil palm concessions (OPC), non-species specific plantations (e.g. mix of rubber, oil palm, acacia plantations) with different size classes including smallscale (<10 ha; SmallPlant), medium-scale (10-100 ha; MedPlant), and large-scale (>100 ha; LargePlant) landholdings, and the reported practice of burning to clear agricultural lands in the past year within the village (PracBurn). The variables SmallPlant, MedPlant, and LargePlant were calculated based on the proportion of plantation area over the regency area, and these proportions do not sum to one. Fires are often used in land clearing, and we chose these proxies to elucidate associations between economic actors and fires.
Our two Population factors included population density (Pop) and the reported use of fires to remove wastes in the village (BurnTrash). We hypothesize that higher population density is likely to correlate with more fires (Cochrane 2003), and increased use of fires for waste disposal is likely to result in greater probability of accidental fires breaking out.  Table 1. Variables used as proxies for social and environmental factors related to 2015 fires (see supporting information for a detailed description of our variables).

Variable
Description Source Variable in regency-level analysis Variable in pixel-level analysis  We included four proxy variables related to forest and peatland degradation under Forest Degradation factors-spatial data on primary degraded forests (DegFor), previous recorded burns from June 2006-May 2015 over peatlands (PrevFires), peatland-cover that are susceptible to fires (SuscPeatCover) and spatial data on roads (Roads). Degraded ecosystems are more susceptible to fire disturbance (Siegert et al 2001, Konecny et al 2016; here we include degraded primary forests (DegFor) and susceptible peatland cover (SuscPeatCover) to encompass both non-peat and peat ecosystems. We included previous recorded burns over peatlands as a separate variable as the burn history of peatlands has been shown to influence its susceptibility to fires (Hoscilo et al 2011). Fragmentation of the landscape by roads increases human access and exacerbates degradation , Laurance et al 2009, which we hypothesize to correlate with increased likelihood of fires.
Our three Biophysical factors included mean monthly rainfall from May-June 2015 (Rain), spatial data on slope (Slope), and peat area (Peat). Drought rainfall is a known predictor of fires, while steeper slopes facilitate fire spread in fire behavioral studies (Rothermel 1983), though in our study site, flatter lands are more accessible to fire ignitions. Peat ignites more easily than mineral soils and has been shown to be an important predictor of fire distribution (Miettinen et al 2011).
2.3. Data analysis 2.3.1. Fire count at regency-level To analyze the social and environmental factors of fire counts among regencies in eastern Sumatra (n = 40, SFigure 1), we used generalized linear mixed effect models (GLMMs) which allow response variables from different distributions and account for correlations between observations and nested data structures by including random factors (Zuur et al 2009). Fire count was modelled using a negative binomial distribution to account for overdispersion and the logarithmic link function. Province (Prov) was specified as a random effect and the log area (L.Area) of the regency as an offset variable to account for regency area differences. We log-transformed population density (L.Pop) and standardized all our variables for analysis. Since our aim was to evaluate the relative contribution of predictors instead of assessing the predictive power of our model, we removed variables OPC and BurnTrash which were highly collinear with LargePlant and L.Pop respectively (rho>0.70). We adopted the information-theoretic approach for statistical inference, using the Aikaike Information Criterion (AICc) adjusted for small sample size to select the best model out of our set of candidate models (i.e. the model with the highest weighted AICc) (Burnham and Anderson 2002). The set of candidate models included individual variables and every combination of the five categories, as well as three interaction terms which we hypothesized will have an effect on fires: rainfall with degraded forests, rainfall with previously burnt areas, and rainfall with susceptible peatlandcover (n = 67, STable 4 and 5). We hypothesized that models with both social and environmental factors perform better at explaining fire counts at the regencylevel. We quantified the goodness-of-fit (R 2 ) of the top three models, comprising marginal and conditional R 2 (variance explained by fixed factors, and both fixed and random factors respectively) (Nakagawa and Schielzeth 2013). GLMMs were constructed using

Fire occurrence at pixel-level
We modelled the occurrence of fires among pixels using boosted regression tree (BRT) models, which combine multiple simple regression trees to form an ensemble model. Individual regression trees (each relating a response to predictors using a 'tree' of recursive binary split) are added sequentially that best reduces a loss function using a stochastic gradient boosting algorithm to improve model accuracy and predictive performance (De'ath 2007, Elith et al 2008. We used BRTs to fit fire distribution (binary indicator of presence or absence) to the 18 predictor variables (STable 6) using the dismo package in R (Hijmans et al 2017). BRTs can handle different types of error distribution, loss functions and predictor variables, and allow for non-linear relationships between response and predictor variables (Elith et al 2008). Social variables (LandConflict, MultiEthnic, Brawl, PracBurn, and BurnTrash) were first made spatiallyexplicit to the village administrative level (see supporting information). All vector data were rasterized, and with other raster data, resampled to 1 km 2 grids (n=225 710 pixels). We ran a training set of models with 10% of the data (n=22 571 pixels) to develop the model and set aside the remaining 90% (n=203 139 pixels) as test data for independent evaluation of model robustness and optimal settings. We used default bag fraction of 0.5 (fraction of samples drawn at random to fit trees at each step) with bernoulli distribution for the response. We ran the models using a range of learning rates (0.005, 0.01, 0.02 and 0.05) which assessed the contribution of each tree to the growing model, and tree complexities (4, 5 and 6) which determined the maximum number of splits for fitting each regression tree (Elith et al 2008). We selected the model that minimized the deviance and maximized the area under the receiver operating characteristic curve (AUC). Outputs include the relative influence of predictor variable plot (measured based on the number of times a variable was selected for splitting, weighted by the squared improvement to the model as a result of each split, and averaged over all trees), and partial dependence plots for each variable (which demonstrate the effect of a variable on the response after accounting for the average effects of all other variables in the model) (Elith et al 2008). Following Müller et al (2013), we did not consider variables with a relative influence value smaller than expected due to chance (100 divided by the number of variables); 5.56% in this evaluation.
We used two different methods to assess the drivers of fire counts at the regency-level and fire occurrence at the pixel-level primarily due to differences in the sample size for these two analyses. Fire count is a continuous variable, with 40 regencies as units for analysis. Fire occurrence is a binary variable, with 225 710 pixels as units for analysis. The use of GLMMs was not feasible for our fire occurrence pixel-level analysis due to its large sample size. Likewise, we could not apply BRTs to our fire count regency-level analysis which had too few data points. As our main objective was to identify the contributing variables to fire counts and fire occurrence at different geographical scales, we applied models that were most appropriate at handling the outcome variables with its associated number of sample units used for analyses.

Regency analysis
The most parsimonious model out of our candidate set of models included economic, population and biophysical variables as important predictors for the expected log count of fires at the regency-level (wAICc=0.871, marginal R 2 =0.216, table 2), and is represented by the following equation:

Pixel-level analysis
The BRT model with the highest independentlyvalidated AUC score (0.799) and lowest deviance (0.504) had a learning rate of 0.01 and tree complexity of 6, giving an optimal tree number of 4,850 (SFigure 2). The relative influence of each predictor variable in this model is shown in figure 4. Only 5 out of 18 variables had relative influence values above 5.56%: Rainfall (25.86%), Slope (22.45%), Population density (17.26%), Roads (10.29%) and Previous Fires (6.62%). The partial dependence plots in figure 5 further show the influence of each of these five variables on fire occurrences. Overall, this suggests that fires are more likely to occur in areas where there has been low rainfall, gentler slopes and lower population density, corroborating the results  from our regency-level analysis (figure 3). Fires were also likely to occur closer to roads and over peatlands which have been burnt repeatedly.

Discussion
We found that biophysical and population variables strongly influenced fire count and occurrence at the regency-and pixel-level respectively (figures 3; 4). However, economic and forest degradation variables had varying impact, with economic drivers being more important at explaining fire count at the regency-level whilst forest degradation factors were more relevant at explaining fire occurrence at the pixel-level. Conflict variables showed no effect on the 2015 fires at both spatial scales. This suggests that mitigation efforts for future fires, especially during ENSO years, can focus their campaigns in high-risk areas that can be identified using updated data on recent rainfall, population density, roads and data on specific land-use or landcover types.
The strong effect rainfall has on fire at both scales is unsurprising. Both analyses showed lower mean monthly rainfall in the preceding two months are likely to result in more fires or higher fire occurrence, concurring with previous studies where a fall in rainfall induced by the El Niño in 2015 had a major contribution to the extreme fires (Field et al 2016, Fernandes et al 2017, Sloan et al 2017. Our BRT model showed an increase in fire occurrence below a mean monthly rainfall of 150 mm/month. This value is close to the critical rainfall threshold of 200 mm/ month based on a study on fire activity from 1982-2010 in Borneo (Sloan et al 2017). Low rainfall leads to less recharge in groundwater levels in peatlands, rendering peat soils highly flammable, which facilitates fire ignition and spread Hooijer 2016, Taufik et al 2017). The effect of slope on fires is also clearly negative at both scales of analyses, which showed that more fires occurred on flat ground. This is likely due to the ease of accessibility compared to steep areas for agricultural activities or resource extraction and also corresponds with the distribution of peatlands in our study site.
Our results from both regency-and pixel-level analyses also revealed that areas with low (but not zero) population densities contribute to fires. This is contrary to what was expected since fires in these landscapes are predominantly anthropogenic and a higher population density is likely to increase the probability of fires (Cochrane 2003, Dennis et al 2005. The inverse relationship between fires and population pressure could indicate that more fires occur in remote areas where land is available for agricultural activities and resource extraction. As population densities increase towards city levels, fires are less likely to occur. This relationship between fires and population corroborates with Cattau et al (2016a) which found that fire ignition first increased, then decreased as distance from settlement increases. Our finding that low population densities have higher fire likelihoods might suggest that prioritizing villages in remote areas in the training of fire prevention awareness and interventions, such as the Fire-Free Village Program, are important in combatting fires (Fire Free Alliance 2016).
When the regency was used as a unit for analysis of fire counts, economic variables were included as important factors in addition to biophysical and population variables. Of the economic variables that were included in our results, the proportion of smallscale (< 10 ha) and medium-scale (10-100 ha) plantation landholdings within a regency showed a negative association with log count of fires. This corroborates with Miettinen et al (2017) where fire hotspot density was found to be lower in small-holder areas compared to industrial plantations. Unsurprisingly, the reported use of fires for clearing agricultural lands in villages showed a positive association with fires, indicating the persistence use of fires for agricultural land-clearing and the importance of managing fire use at the villagelevel. The role of agency in Indonesia's fires is a contentious issue due to serious ramifications of extreme fire events such as the 2015 fires. Determining who is accountable for Indonesia's fires is complex due to overlapping land claims between industrial plantation companies, medium landholdings and small farmers (Gaveau et al 2014(Gaveau et al , 2016. While cross-sectional studies like ours and Miettinen et al (2017) show less fires within small to medium landholdings, Sloan et al (2017) focused on long-term trends of fires in Borneo and showed how fires resulted from interactions between large plantation concessions and economic land-use activity by small and medium landholdings. Longitudinal studies on fire patterns and land-use in Sumatra could reveal a more in-depth understanding of shifting drivers of fire activity which could be an area for future work.
When the pixel (1 km 2 ) was used as a unit for analysis of fire occurrence, forest degradation, in addition to biophysical and population variables were included as important factors. Roads have been known to contribute to forest degradation (Laurance et al 2009) and provide access for lighting fires . Our BRT model demonstrated that fire occurrence increased between 0-2.5 km from roads, before decreasing as distance to roads increases up until 25 km before rising again. This suggests that proximity to roads increases the probability of fires being lit. At greater distances from roads, accessibility could be restricted and the likelihood of fire occurrence decreases. However, remote fires such as that recorded under our BRT model at >25 km from roads could still be lit, and access to such sites could be through rivers or drainage canals which were not included in our analyses.
The proportion of peatlands in the regency showed a positive association with fires under our regency-level analysis but fell below the randomness threshold under our pixel-level analysis. Instead, peatlands which experienced repeated burns from June 2006-May 2015 was identified as an important variable for fire occurrence at the pixel-level. Our BRT model demonstrated that fire occurrence increased in peatlands which have been burnt between one and six times before plateauing at 15 burn events. The relationship behind fire occurrence and burn history may be related to post-fire changes in vegetation cover over peatlands, relative peat depth and chemical properties of peat soils. Peatswamp forests that were burnt in an initial fire could result in intense fires due to the consequent woody vegetation cover. These initial fires leave behind standing and fallen timber which contribute as above-ground fuel for subsequent fires. As the vegetation cover of peatlands transitions from woody to non-woody vegetation through repeated burning, the volume of woody fire fuels decreases, possibly leading to less fire occurrences (Hoscilo et al 2011). Konecny et al (2016) also found that the frequency of fires over peatlands decreased the relative burned area depth of peat soils, potentially influencing the mass of peat soil available for burning. Past fires could also alter the chemistry of peat soils, reducing the amount of labile, easily-combustible carbon constitutes and increasing the resistance of peat soils to fires (Milner et al in prep.) Our results suggest that burn history of peatlands should be considered when monitoring Sumatra's fire landscape, since peatlands with few burn events show a higher risk of being burnt in the next fire.
Although conflict has been found to be an underlying cause of fires in Indonesia (Suyanto 2007, Dennis et al 2005, our results suggest, similar to , that social conflicts did not play a major role in the 2015 fires. Clear land ownership and tenure in Indonesia has always been lacking, where adat or customary laws are often overlooked when granting land concessions (Shivakumar et al 2015). Combined with the lack of a transparent system for redress and compensation when conflict arises, fires are started as retaliation to destroy crops and reclaim land, both between smallholders and companies, and between locals and transmigrants (Suyanto et al 2004, Dennis et al 2005, Suyanto 2007). Conflict, however, is an inherently social phenomenon with all its accompanying complexities, and relying on reports of land conflicts from newspaper articles risk simplifying this issue due to the bias of under-reporting or selective reporting in the media (Abram et al 2017). We attempted to reduce this bias by obtaining information from three national newspapers, six provincial newspapers, and supplementary reporting from three local non-governmental organizations. We also included variables such as village-level multi-ethnicity and brawl incidents as proxy variables for conflict but these were also unimportant in influencing fire count and occurrence in our analyses. While our proxies for conflict could be inadequate, the unimportance of conflict relative to other variables in explaining fire count and fire occurrence in the 2015 extreme fires could be a result of the overwhelming influence of the El Niñorelated drought in that year.

Conclusion
Our study aimed to gain a better understanding of the variables that had the greatest influence on the 2015 extreme fire event in Sumatra. We showed that rainfall, slope and population density were the most important variables predicting fires. The main influence of rain corroborates with previous studies, and highlights the importance of establishing an early warning system for droughts to better prevent and manage future extreme fire events (Field et al 2016, Tacconi 2016, Lee et al 2016, Miettinen et al 2017. Economic variables such as the proportion of smallscale and medium-scale plantation landholdings and the reported use of fires to clear agricultural lands in villages were important factors in explaining fire count at the regency-level. Distance from roads and the number of recorded burns over peatlands were important contributors to fire occurrence at the pixellevel. Mitigation efforts for future fires, especially during ENSO years, can be focused on high-risk areas identified using environmental data on rainfall, slope, peatlands, and previously burnt peat areas, as well as social data related to population density, access to roads, extents of small-and medium-plantation landholdings, and village-level propensity to burn land for agriculture.