Heightened fire probability in Indonesia in non-drought conditions: the effect of increasing temperatures

In Indonesia, drought driven fires occur typically during the warm phase of the El Niño Southern Oscillation. This was the case of the events of 1997 and 2015 that resulted in months-long hazardous atmospheric pollution levels in Equatorial Asia and record greenhouse gas emissions. Nonetheless, anomalously active fire seasons have also been observed in non-drought years. In this work, we investigated the impact of temperature on fires and found that when the July-October (JASO) period is anomalously dry, the sensitivity of fires to temperature is modest. In contrast, under normal-to-wet conditions, fire probability increases sharply when JASO is anomalously warm. This describes a regime in which an active fire season is not limited to drought years. Greater susceptibility to fires in response to a warmer environment finds support in the high evapotranspiration rates observed in normal-to-wet and warm conditions in Indonesia. We also find that fire probability in wet JASOs would be considerably less sensitive to temperature were not for the added effect of recent positive trends. Near-term regional climate projections reveal that, despite negligible changes in precipitation, a continuing warming trend will heighten fire probability over the next few decades especially in non-drought years. Mild fire seasons currently observed in association with wet conditions and cool temperatures will become rare events in Indonesia.


Introduction
Large areas of tropical humid forests are being altered by human activities such as clearing and logging (FAO 2015), making these landscapes more vulnerable to drought-driven fires (Brando et (Murdiyarso and Adiningsih 2007). While fire ignition is related to human activities, the interannual variability of fire spread and frequency responds to large-scale climate fluctuations (Ceccato et al 2010, Chen et al 2016, Murdiyarso and Adiningsih 2007, Siegert et al 2001, van der Werf et al 2008a. During the 1997 El Niño drought, disastrous fires in Indonesia resulted in months-long hazardous atmospheric pollution levels (Marlier et al 2013) and carbon emissions estimated at between 4% (Levine 1999)  Original content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence.
Any further distribution of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI. emissions from fossil fuels (Boden et al 2016). Large fires were observed again during the 2015/2016 El Niño and fire-related greenhouse gases (GHG) emissions were second only to those of 1997 (Huijnen et al 2016). Despite the clear impact droughts have on fire occurrence in Indonesia, anomalously active fire seasons have recently been observed during nondrought years (Gaveau et al 2014), raising the question of whether climate and fire dynamics are any different in the absence of seasonal droughts. In non-tropical regions, temperature plays an important role in determining fire risk, often interacting with precipitation according to whether it is dry or wet (Aldersley et al 2011, Westerling andBryant 2008). We investigate whether this is the case in Indonesia, as the current role of temperature in fire dynamics is not well understood and its future impact on fires, though not uniformly distributed globally (Moritz et al 2012), is projected to increase (Pechony and Shindell 2010). The atmosphere's water vapor demand increases with temperature and with it evapotranspiration rates (Breshears et al 2013). As a consequence, faster soil and vegetation water depletion (Williams et al 2013, Zhao and Running 2010) and greater susceptibility to fires (Abatzoglou andWilliams 2016, Brando et al 2014) are observed. This has important implications for future fire activity assessments in Indonesia, as studies tend to focus on how projected changes in precipitation regimes will impact fire occurrence (Corlett 2016, Herawati and Santoso 2011, Lestari et al 2014 and less attention is given to potential compounding effects of temperature. Here, we investigate fires sensitivity to precipitation and temperature using a statistical model built from over two decades of remotely sensed fire observations and meteorological stations data. Then, we explore the role of evapotranspiration rates variability as a mechanism linking the empirically estimated fire probabilities to the climate predictors. Lastly, the fire probability model is driven by near-term precipitation and temperature projections with the goal of assessing future changes in fire activity in Indonesia.

Study area
Our domain of study is comprised of Sumatra and Kalimantan; the later is the Indonesian part of the island of Borneo. Fires in Kalimantan and Sumatra are typically lit during the dry season, defined here as the months from July to October (JASO), to clear land for planting (Murdiyarso and Adiningsih 2007).

Data and methods
We model burned area anomalies as a function of precipitation and temperature anomalies for the period 1995-2015. We propose a physical mechanism to explain the modeled burned area anomalies response to the climate predictors by evaluating how evapotranspiration rates respond to contrasting climate conditions. Then, we used projections of regional temperature and precipitation to anticipate the likely response of fires to future climate in Indonesia.
3.1. Fire probability as a function of climate 3.1.1. Global fire emission database We use the burned area (BA) product from the fourth generation of the Global Fire Emission Database (GFED4) (Giglio et al 2013) covering a period of 21 yr, including the two major fire events of 1997 and 2015 in Indonesia. GFED4 provides global monthly BA, in hectares, at 0.25 o spatial resolution from 1995 to 2015. Interannual variability is evaluated by spatially averaging JASO BA over the entire study area then subtracting the 1995-2015 climatology and dividing by its standard deviation (StdBA). Anomalous BA is also calculated locally (at each grid cell) for the July-October season by removing the local 1995-2015 climatology. Because fire ignition in Indonesia is human-induced (Murdiyarso andAdiningsih 2007, Stolle et al 2003), pixels with BA equal to zero hectare were masked out, as we assume that an absence of fires over the entire JASO season is related to an absence of fire-related activities, not climate.

Precipitation and temperature data
The gridded Global Precipitation Climatology Centre (GPCC) rain-gauge only dataset is used at monthly time-steps from 1966 to 2015 (Schamm et al 2014. Precipitation is linearly regridded to 0.25°resolution to match the GFED4 burned area dataset. Domain averaged time series of standardized precipitation (StdPr) is calculated by spatially averaging the 1966-2015 JASO precipitation over the entire study area then subtracting the baseline climatology and dividing by its standard deviation. JASO StdPr is also calculated for every grid cell within the domain and the 1995-2015 period. The length of StdPr domain averaged time series is longer than the anomalies calculated at the grid cell level but the baseline climatology is the same . The individual grid cell StdPr was used to model fire (GFED4-BA) probability while the whole study area StdPr was used to model near future variability and change.
Near surface air temperature is provided by the National Oceanic and Atmospheric Administration (NOAA) Climate Prediction Center (CPC). This dataset results from two large meteorological station networks that are combined and interpolated globally to a 0.5°degree spatial resolution grid and monthly time-steps (Fan and Van den Dool 2008). Temperature anomalies (TempAn) for JASO 1966-2015 were calculated for the whole study area by spatially averaging JASO temperature and removing the Environ. Res. Lett. 12 (2017) 054002 baseline period  climatology. This time series is used model near future variability and change. TempAn is also calculated locally by subtracting each grid cell's 1995-2015 JASO climatology and it is used to model fire (GFED4-BA) probability.
3.1.3. Statistical analysis of fire probability Fire probability, measured as the probability of positive BA anomalies, is modeled as a function of temperature anomalies (TempAn) and precipitation standardized anomalies (StdPr). The JASO BA anomalies are calculated with respect to the 1995-2015 baseline period at each grid cell and categorized as 1 for positive (active fire season) and 0 for negative (mild fire season) values. Every grid cell that registered JASO BA different from zero (and corresponding TempAn and StdPr) is retained, totaling a sample of 13 848 in length.
Fire was modeled using logistic regression and expressed as: where P(BA ¼ 1) is the probability of an active fire season (BA ¼ 1) modeled as a linear combination of predictors X n (StdPr and TempAn) with slopes b n and an intercept a. The logistic regression model was fit using Matlab's fitglm function (Mathworks 2015b) for a binomial distribution. We used the variance inflation factor (VIF) test for multicollinearity (Belsley 1991). The best model was selected by evaluating values of the Akaike and Bayesian Information Criteria (AIC, BIC) for each individual covariate and an interaction term between StdPr and TempAn (table 1). AIC and BIC are standard statistical measures for model selection and the lowest values indicate the best fit (Burnham andAnderson 1998, Raftery 1995). To test fires sensitivity to potential trends in the predictors, a model was fitted to de-trended TempAn and StdPr (table S1 available at stacks.iop.org/ERL/12/054002/mmedia).

MODIS evapotranspiration (MOD16)
Evapotranspiration (ET) rates, which is a function of temperature and moisture availability, accelerate in an anomalously warm environment resulting in fast soil and vegetation water depletion (Breshears et al 2013, Zhao and Running 2010) and greater susceptibility to fires (Abatzoglou andWilliams 2016, Brando et al 2014). With the goal of establishing a physical mechanism explaining fire behavior, we investigate how ET rates respond to interactions between temperature and precipitation and how it relates to fire probability as estimated by our statistical model. We use MODIS global evapotranspiration product (MOD16), obtained at a 0.5°spatial resolution and monthly time step available for the period January 2000 to December 2014 (Mu et al 2011, Zhao et al 2006. Seasonal JASO standardized ET (StdET) is calculated by subtracting the climatology in each grid cell and dividing it by its standard deviation. The final product is re-gridded linearly to 0.25°resolution to match the spatial resolution of the GFED4-BA product.

Near-term precipitation and temperature projections
Using the fire probability model (section 3.1.3) and selected near-term simulations of TempAn and StdPr, we assess future fire risk in Indonesia. The fire probability predictors, observed temperature and precipitation, are normally anti-correlated over equatorial land areas, but this feature is not consistently represented in General Circulation Models (GCMs). This is attributed to various soil moisture, cloud and precipitation parameterizations (Berg et al 2015, Koster et al 2009, Trenberth and Shea 2005, Wu et al 2013 in the models. Therefore, GCMs direct temperature and precipitation outputs are not adequate to assess these variables' covariability in Indonesia. This dependence between precipitation and temperature requires they are modeled jointly and a multivariate modeling scheme is used here (Greene et al 2012). The method consists of creating synthetic climate sequences that retain the statistical properties of the original series. The simulations are based on annual time steps of JASO StdPr and TempAn for the study area and period 1966-2015. The time series are first de-trended by removing the variables' local response to anthropogenic climate forcings (Greene et al 2011). The 50 yr  de-trended JASO StdPr and TempAn (henceforth referred to as DetTempAn and DetStdPr) retain oscillations on the scale of years to decades. Multiyear and decadal oscillations can result from periodic predictable components explained by Table 1. Logistic regression coefficients (b), standardized errors (SE), and statistical significance (P) for regression predictors of above average burned area anomalies. Bayesian information criterion (BIC) and Akaike information criterion (AIC) for single variable models (TempAn and StdPr) and the multi-variable model with an interaction term (StdPr:TempAn where y is the climate vector (DetStdPr and DetTempAn) at time t and tÀ1. The time steps are years as the variables have one JASO entry per year. A is a matrix of coefficients and u t is a stationary white noise (serially uncorrelated) vector. The VAR(1) was fitted to the 1966-2015 JASO regional series using Matlab's Econometric Toolbox VAR model functions (Mathworks 2015a).
The VAR (1)   . We found a highly significant correlation between StdPr and StdBA (R ¼ À0.91, P < 0.05), suggesting that the intensity of StdPr is a very good indicator of the intensity of burned area anomalies. This shows that over 80% of BA variance is explained by StdPr meaning that, for prediction purposes, the effect of any other variable on fires seems to be of secondary importance. Nonetheless, we test whether moisture availability, assessed as precipitation anomalies, impacts fire response to temperature by correlating StdBA and climate variables for two subsamples: one comprised of dry JASOs (StdPr < 0) and one of wet JASOs (StdPr > 0) taken from the 1995-2015 record (figure 2). For the 9 dry JASOs, the correlation between StdBA and StdPr reproduces that of the entire sample and the correlation between StdBA and TempAn is not significant (R ¼ À0.19, P > 0.6) showing in fact the opposite behavior expected of a warm environment leading to more fires. For the remaining 12 wet JASOs, the correlation between StdPr and StdBA is also statistically significant (R ¼ À0.59, P < 0.05), but in contrast to dry JASOs, TempAn becomes a much more important variable with a significant positive correlation to StdBA (R ¼ 0.62, P < 0.05). These results show that StdPr is the best indicator of StdBA anomalies under dry conditions, but that TempAn correlates more strongly Environ. Res. Lett. 12 (2017) 054002 with StdBA in a wetter dry season. This indicates that burned area response to temperature is controlled by the precipitation regime.

Fire probability model
The logistic regression model with the best fit (lower AIC and BIC values) includes an interaction term between StdPr and TempAn (table 1) and can be rewritten from equation (1) The presence of an interaction term in the model demonstrates that the nature of positive BA anomalies response to temperature varies according to the level of the precipitation anomaly, consistent with the correlation analysis presented in the previous section. The interaction effect between the predictors is further illustrated by plotting fire probability as a function of TempAn and contrasting wet and dry conditions (figure 3). We set in the model the values for dry (StdPr ¼ À0.67) and wet (StdPr ¼ 0.67) to represent the 25% drier and wetter tails of each grid cell precipitation distribution. In dry conditions (solid red line in figure 3) there is a nearly constant increase in fire probability per unit increase in temperature anomalies (9% to 14% per 0.5°C). In wetter than normal JASOs (solid blue line), negative TempAn is related to low fire probability, but the chances of observing an active fire season increases sharply from approximately 15% under normal temperature to above 40% at anomalies of 0.5°C. This behavior is consistent for a range of positive StdPr (online supplementary figure S2). The probability of an anomalously active fire season is mainly determined by whether it is dry or wet if TempAn is near or below normal (solid lines in figure 3), but as TempAn becomes positive, it turns into a much stronger driver especially in wet conditions. In the case of the model fitted to the de-trended variables, we find that fire probability is always higher in dry than in wet environments (dotted lines in figure 3), regardless of TempAn. This is characteristic of a response in which the interaction between the variables is negligible (online supplementary table S1). But more importantly, when it is wet, the response of fire probability to TempAn is much weaker than in the model fitted to the original variables (blue lines in figure 3). This suggests that fires in non-drought conditions would be less influenced by temperature anomalies in any given year were not for the compounding effect of increasing trends. In contrast, fire probability responds very similarly to TempAn in both models (red lines in figure 3) showing that temperature trends have no particular enhancing effect on fires under dry conditions.

The role of evapotranspiration rates (ET)
The role of temperature and moisture availability, measured as precipitation anomalies, on ET rates is evaluated by analyzing a Environ. Res. Lett. 12 (2017) 054002 cumulative probability in wet conditions (StdPr > 0) reveals that for most part anomalously low StdET rates are related to TempAn < 0 (dashed green line in figure 4), whereas in cases of wet and TempAn > 0 nearly 65% of the sampled StdET rate is anomalously high (solid green line in figure 4). Thus, provided that moisture is available, anomalously positive StdET rates are consistently higher in warm than in cool conditions. In dry (StdPr < 0) conditions the distribution of StdET for both TempAn < 0 and Temp > 0 (orange lines in figure 4) is not as contrasting, indicating that the evaporative regime in this case is less dependent of temperature and constrained instead by moisture availability. The higher StdET rates associated to positive TempAn in normal-to-wet conditions provides evidence consistent with the heightened modeled fire probability found for a warm and wet environment. probability modeled for wet (StdPr ¼ 0.67) conditions and the 100 realizations of 2020-2029 JASO TempAn point to higher fire probability in general, but more importantly, low fire probability related to the combined effect of negative TempAn and wet conditions is much less likely in the 2020-2029 simulations (thin blue line in figure 6) than currently observed (thick blue line in figure 6). Also, the distribution of simulated TempAn in association with the 2020-2029 normal-to-wet conditions (StdPr > 0) shifts to mostly positive values in contrast to the distribution used to fit the fire probability model (histograms of figure 6). In other words, the chances of observing wet and warm conditions in the next few decades are much greater than in the current climate. At the same time, wet accompanied by cool temperatures (TempAn < 0) will become nearly inexistent in the near future, turning mild fire seasons into a rare event in Indonesia. Fires are less sensitive to TempAn in dry conditions and the 2020-2029 simulated fire probabilities mostly overlap the function representing current characteristics (not shown).

Discussion and conclusions
Drought related fire is a fairly well understood phenomenon in Indonesia but the impacts of an increasingly warm environment on fire dynamics has been less explored. The lesser interest in temperature and fire dynamics finds support in fire intensity responding mostly to the severity of precipitation Environ. Res. Lett. 12 (2017) 054002 anomalies during the July-October dry season. Indeed, in a dry environment the effect of temperature is modest and our model estimates a nearly constant increase in fire probability per unit increase in temperature anomalies (9% to 14% per 0.5°C). However, this characteristic is not linear across contrasting precipitation patterns and in wet conditions temperature is a far more important variable determining fire probability. A sharp increase in modeled fire probability (15% to 40%) is found when temperatures go from near normal to 0.5°C anomalies. This is in part due to an observed regional warming trend. In a wet environment, fires are highly sensitive to anomalously positive temperatures only if trends are retained in the data used to model fire probability. In other words, fires would be less sensitive to any given year anomalous temperatures were not for compounding effect of the multiyear trend. In dry conditions, fires are driven mainly by the severity of precipitation deficit and the added effect of temperature anomalies is similar regardless of recent trends. Higher evapotranspiration rates in wet and warm as opposed to wet and cool conditions are consistent with increased fire probability. High temperature leads to high atmospheric vapor pressure deficit and ET rates, and regardless of changes in precipitation, a warming trend can induce vegetation water stress and greater susceptibility to fires. This has important implications for Indonesia, as near-term climate simulations point to a continuing increase in temperature, while precipitation variability remains largely unchanged. Our findings add yet another layer of complexity to fire manage-ment, suggesting that fire prevention measures should no longer be restricted to dry years. In addition, further understanding of how human activities and various land use interact with increasingly higher temperature could help identify effective interventions to prevent and mitigate fire occurrence and spread in Indonesia. On longer timescale, strategies to reduce GHG emission through forest restoration and conservation should take into account the effect of a warmer environment on fires, as currently observed mild fire seasons associated with wet and cool conditions will become rare events in Indonesia.