ROUGH DRAFT authorea.com/68364

# “El Niño” events increase seasonal wildfire danger in Ecuador

Abstract

Weather is a key driver of catastrophic wildfire seasons. On a global scale, different phases of the El Niño-Southern Oscillation, “El Niño” or “La Niña”, modulate weather and therefore wildfire activity. Ecuador is one the most biodiverse countries in the world and wildfires produce severe impacts to its ecosystems. This investigation explores for the first time the relationship between wildfire weather and El Niño-Southern Oscillation in the tropical Andes. Wildfire weather variability has been quantified using seasonal metrics of the McArthur Forest Fire Danger Index. The calculation of this index requires weather station data that were only available since 1997. Therefore, the availability of data was extended using the Twentieth Century Reanalysis Project to cover the period 1963-2010. Bushfire weather in the seasons April-May-June and July-August-September show significant correlations with the NINO3.4 index (r=0.37,p=0.003, n=47 and r=0.4, p=0.001, n=47 respectively). Additionally, a Chi-squared test showed that every extreme wildfire season in this region was linked to an ‘El Niño’ event. Our results demonstrate that ‘

# Introduction

Wildfires are a global phenomenon that produces severe impacts (Luke 1978, Conard 1997, Morton 2003). Catastrophic wildfire events claim lives (Haynes 2008, Haynes 2010), degrade the environment (Shakesby 1993, Stephens 2004, Lane 2010) and destroy building infrastructure (Morton 2003, McAneney 2009, Crompton 2010). Wildfires also contribute to biodiversity loss (Kodandapani 2008, Pastro 2011) and affect the climate (Conard 1997). They are part of the natural evolution of vegetation ecosystems (HaynesBradstock 2012). However, human actions amplify the impact over people, nature and infrastructure (Pausas 2011). In fact, wildfires have a strong anthropogenic ignition pattern. In addition, population growth in rural areas increase human vulnerability to this hazard. The threat of wildfires have encouraged extensive research on the behavior of its precursors.

Weather is one of the key components that contributes to wildfire activity (Powell 1983, Williams 1999, Mills 2005, Haines 1988, Mccaw 2007, Dowdy 2012, Engel 2013). Wildfire weather refers to the meteorological influence over wildfires over a period of time —usually within days—. Common variables associated with wildfire weather are temperature, relative humidity, wind speed and precipitation (BoM 2009). Each variable exerts influence in the liberation of water content in vegetation. This process—Evapotranspiration—controls the availability of fuels to be burnt. Most of the research conducted on wildfires and climate focuses on this daily approach (Roads 2005). Yet, the study of wildfire predictability on a seasonal scale is also important. A key factor to explore this perspective is El Niño Southern-Oscillation (ENSO) (Swetnam 1990, Simard 1985, Kitzberger 2002, Beckage 2003, Wooster 2012).

ENSO produces negative rainfall anomalies in some wildfire-prone regions (Swetnam 1990, Simard 1985, Kitzberger 2002, Beckage 2003, Wooster 2012). The intensity of these anomalies depends on the regional influence of ENSO over climate. These rainfall deficits occur over the months before—and during–the wildfire season. ENSO’s influence may even increase by the superposition of the Inter-decadal Pacific Oscillation (IPO) (Verdon 2004). ENSO also has an influence over ignition patterns. In some regions it fosters the occurrence of lightning strikes (Beckage 2003). These facts encourage the investigation of new studies of ENSO and wildfire weather around the world.

The west coast of South America is one of the regions most affected by ENSO. The coastal areas of Peru and Ecuador experience massive rainfall episodes during strong El Niño events (Rodbell 1999). Yet, there has not been enough research on its influence over other natural hazards. Wildfires occur all over South America (Manta 2008). However, studies about wildfire weather in the northern Andean region are scarce. This study aims to describe the seasonal variability of wildfire weather in Ecuador. Also, we investigate if there is a link between this type of weather and ENSO in this region. This is important because wildfires occur every year in this country producing severe impacts (Ministerio del Ambiente 2013, Secretaria de Ambiente 2013). A better understanding of wildfire weather variability and its precursors can provide useful information. These insights could be of special interest to emergency response services and planning authorities.

This article comprises six sections. In the first section we provide background and justify the importance of this study. Section two describes the Ecuadorian Andean region as our study area. Section three presents details about the datasets employed in this research. Section four explains the different methods applied. Section five shows the results of our investigation. Finally, section six presents a discussion of the results and concluding remarks.

## Study area

### Geography

Ecuador is located in the north-west side of South-America (see Figure 1). The country has a diverse geography. It comprises a continental territory between the latitudes 01o 28’ N and 05o 02’ S and longitudes 75o 11’ W and 81o 04’ W (Instituto Geográfico Militar (IGM) 2013). The Galapagos Islands are also part of the country. These islands are approximately 1000 Km from the continent. The national territory has a total area of 256.370 Km2 (Instituto Geográfico Militar (IGM) 2013). The Andes mountains divide the continental Ecuador into three areas: the Coast, Andean and Amazon regions. Each of these regions has diverse geographical features, climate, and ecosystems. The highlands are the most wildfire-prone region because of its specific natural characteristics.

The Andes mountains are the most important geographical feature of Ecuador—and South-America— (Insel 2010). This mountain range crosses Ecuador from North to South. The Ecuadorian Andes comprises two flanks of mountains: the eastern and western ’Cordilleras’. Between these two flanks—that include active volcanoes—there are several inter-Andean valleys. Many of these Andean plateaus are over 3000 m above sea level. The steep elevations have an important effect on local climate. The lapse rate create the conditions for different Bio-climatic zones to exit. This enhances biodiversity. Yet, the original vegetation of the Andean valleys has almost disappeared. Most of the endemic vegetation was replaced by the Australian specie Eucalyptus globulus around 1860 (Ministerio de Ambiente del Ecuador 2012). Other introduced species are the Pinus radiata (from California) and the Pinus patula (from Mexico). This occurred for commercial reasons (Anchaluisa 2013) with great ecological impact. Further, these species are prone to wildfires in the Andean dry season.

### Climate

The tropical Andes climate has several drivers. The Amazon forest, oceanic currents, and the topography are among the main drivers (Martínez 2011). The Amazon forest—through Evapo-transpiration—produces a great amount of water vapor. This water precipitates along the western Cordillera because of an orographic precipitation process. This precipitation is also produced in the eastern Cordillera. Air masses advected from the Pacific Ocean produce this effect. The influence of the Pacific Ocean in the climate variability is important. The Pacific Decadal Oscillation (PDO) is the main driver of decadal climate variability in this region (Martínez 2011). On the other hand, the inter-annual variability depends on the El Niño Southern Oscillation (ENSO) (Vuille 2000, Villacis 2003, Martínez 2011). These influences create several climate regions within the Andes.

Most of the Ecuadorian Andes—between 1500 to 3000 m—have a “Semi-Humid Mesothermal” climate (Pourrut 1995). Mean temperatures in this region range from 8 oc to 20 oc. Maximum temperatures span from 22 oC and 30 oC. Minimum temperatures range from -4 oc to 5 oC. Air masses from the Pacific Ocean and the Amazon region create three seasons (Pourrut 1995). Two wet seasons during the periods February-May and October-November. A dry season spans from June to September. Total annual precipitation ranges from 800 mm to 1500 mm along the Andes. Relative humidity varies from 65 to 85 percent.

### Fire

Wildfires mostly occur in three geographical regions of Ecuador. The Andean region has large extensions of introduced Eucalyptus forests (Anchaluisa 2013). Thus, the availability of fuels makes it the most wildfire-prone region. The fire season in the Ecuadorian Andes spans from July to November (Secretaria de Ambiente 2013). Yet, the most critical months are July, August and September (Estacio 2012, Secretaria de Ambiente 2013). The coast and the Galapagos Islands experience wildfire episodes from January to May (Secretaria de Ambiente 2013).

Wildfires produce severe impacts in Ecuador. Yet, information about these impacts is scarce and disperse. Wildfires produced 21.570 Ha of area burnt in Ecuador in 2012 (Ministerio del Ambiente 2013). The Metropolitan District of Quito (MDQ) is particularly vulnerable to wildfires. The vulnerability is high because 15.4 percent of the national population lives here. Additionally, vegetation covers 60.46 percent of its territory (Secretaria de Ambiente 2013). Wildfires destroyed 2700 Ha of the MDQ vegetation in 2009 (Estacio 2012). In 2012, the area burnt by wildfires increased to 4.882,16 Ha (Secretaria de Ambiente 2013). This wildfire season cost the MDQ more than 50 million dollars (Secretaria de Ambiente 2013).

Unfortunately, wildfires in Ecuador are anthropogenic in nature (Ministerio del Ambiente 2013). In fact, the MDQ reported that humans cause 95 percent of wildfires in the city (Secretaria de Ambiente 2013). Burning forests to convert them to agricultural use is the main cause of wildfires in the country (Ministerio del Ambiente 2013, Secretaria de Ambiente 2013, Rodas 2015). Forcing land use changes for urbanization purposes is another reason people start wildfires (Secretaria de Ambiente 2013). Rural communities burn agricultural waste, which is a traditional practice that fosters wildfire occurrences (Estacio 2012). Rural people also burn regular waste because of the lack of waste management services. (Secretaria de Ambiente 2013). Finally, wildfires occur because of the damaging acts of arsonists (Ministerio del Ambiente 2013, Secretaria de Ambiente 2013).

# Data and methods

## Data

### Weather station

This study required weather station (WS) data from the Andean region in Ecuador. A limitation with this dataset was that only six stations had enough records for this study. The selected period of analysis was 1997-2012. This period had almost complete daily information. The variables for this research are air temperature (Tmax), relative humidity (RH), wind speed (W) and precipitation (P). Each of the stations provided data on an eight-hourly basis. The hour selected for the computations was 1 p.m. local time (GMT 18:00). Table 1 describes each of the stations.

### Reanalysis

This research also used reanalysis data from the Twentieth Century Reanalysis Project (20CR) (Compo 2011). The 20CR relies on a data assimilation algorithm of surface pressure values. Sea-surface temperature and sea ice distributions are boundary conditions in this model. The 20CR yields an ensemble mean of realizations. This produces the best possible representation of the state of the atmosphere. This dataset yields six-hourly data on 24 vertical levels. The 1000 hPa-level data is the first vertical level. Thus, it is the most likely to represent surface values. The hour selected for calculations with this data was 4 p.m. local time (GMT 06:00). This is the time closest to the hour adopted for most wildfire potential computations (3 p.m.) with the Australian Forest Fire Danger Index (FFDI)—to be explained next—. The 20CR has data with a resolution of 2o x 2o. Thus, the extracted data is a grid in the Ecuadorian Andean region.

This dataset presents limitations common to all reanalysis products. There is greater uncertainty in regions were there are less observations. This is a particular problem in Africa, some regions in Asia and South America. There is less reliability in the first half of the century due to lack of observations. Although there have been studies that reflect this data represents well climate variability during this period (Ashcroft 2014). Yet, studies for South America are not available. Additionally, some variables are better represented than others.

### ENSO Indices

A common ENSO metric is the Southern Oscillation Index (SOI). The SOI represents the difference in normalized surface pressure between Taihiti and Darwin, Australia. We adopted the SOI’s calculation method proposed by the National Oceanic and Atmospheric Administration (NOAA 2015). The index reflects the changes in air-mass fluxes in the tropical Pacific Ocean. These fluxes, are a manifestation of changes in sea surface temperature (SST). The SOI has an open ended scale with positive and negative values. A negative SOI value represents above-average warm SST conditions in the eastern tropical Pacific Ocean. This represents an ’El Niño’ phase of the phenomenon. El Niño brings wet conditions to the coasts of southern Ecuador and northern Peru. The opposite pattern occurs with a positive SOI value, representing a ’La Niña’ phase.

Another ENSO index is the Oceanic Niño Index (ONI) (NOAA 2015a). The ONI measures monthly SST values in the tropical Pacific Ocean. The measurements take place in the region from 5 oN to 5 oS latitude and 120 oW to 170 oW longitude. This is the ’Niño 3.4’ region. The index requires the calculation of 3-month running means and its climatology. The ONI is the difference between these two values. An ’El Niño’ condition exits if there is at least an anomaly of +0.5 oC in this computation. An anomaly of -0.5 oC indicates a ’La Niña’ phase.

## Methods

### Data Pre-processing

The computation of wildfire potential usually requires continuous daily data. Thus, only six stations in the Ecuadorian Andean region were suitable for this study. The selected stations are representative of the ecosystems that are prone to wildfire. The available records in these stations spanned the period 1997-2012. Yet, the period of analysis is short. Thus, reanalysis data was bias-corrected to extend the availability of records.

The bias-correction was necessary for several reasons. First, 1000 hPa reanalysis data does not represent surface level values. Second, there was a 3-hour difference between the time of the two dataset’s values. Third, reanalysis data in the highlands may differ in quality to sea-level outputs. Thus, we plotted probability density functions (PDFs) of the reanalysis and weather station datasets. This allowed to determine the differences between their data distributions. Applying linear scaling techniques was a suitable approach to bias-correct the required variables.

The aim of the study was to estimate wildfire potential for the entire Ecuadorian Andean region. Thus, a calculation of daily averages for each wildfire weather variable yielded regional values. With the reanalysis dataset the procedure is straightforward. These data is spatially and temporally continuous. However, the weather stations data had some missing days. Therefore, calculating averages over daily anomalies made the results less sensitive to missing data.

### Forest Fire Danger Index (FFDI) calculation

Ecuador currently does not have any scientifically designed wildfire potential index. Therefore, an alternative approach was to use an index suitable for this region. The FFDI is an Australian wildfire potential index (Noble 1980). This index uses daily values of temperature, relative humidity and wind speed. The index also incorporates a drought factor. This factor represents the availability of fuels. This is measured indirectly using soil moisture indexes such us the Keetch–Byram drought index (Keetch JJ 1968). The FFDI equation is expressed as follows:

$\label{eqn:FFDI} FFDI = 2e^{-0.45+0.987*Ln(DF)-0.0345RH+0.0338T+0.0234v}$

The variables represented are drought factor (DF), relative humidity (RH), temperature (T) and wind speed (v). This empirical index is valid for the southeastern forest ecosystem of Australia. In this wildfire-prone ecosystem, the eucalyptus is the dominant vegetation. On the other side of the world, Latin-America—including Ecuador—experienced a massive introduction of this specie (Anchaluisa 2013). Thus, is reasonable to expect that this index may represent the Ecuadorian wildfire potential.

The application of this index with weather station and reanalysis data yielded daily values. With weather station data the calculation comprised the period 1997-2012. The bias-corrected reanalysis data allowed to extend the method for the period 1963-2012. For each year, a sum of the daily FFDI values provided a seasonal metric of wildfire potential. This cumulative FFDI corresponds to the wildfire season July-August-September (JAS). Finally, we calculated the 90th percentile on the seasonal FFDI time series. This determined which were extreme wildfire seasons during the period 1963-2012.

### Investigation of ENSO influence over wildfire seasons

We used several approaches to investigate ENSO’s link over wildfire seasons. Firstly, it was important to know the influence of ENSO in the months leading to the wildfire season JAS. Thus, we applied linear correlations of ENSO’s indexes (seasonal averages) with FFDI values. These averages comprised three-months seasons starting from October-November-December (OND) of the previous year. The correlation also included the SOI and Niño 3.4 values for the concurrent JAS season.

Secondly, we explored the relationship of each wildfire variable with ENSO. This comprised exploring the link of concurrent FFDI variables (JAS) with antecedent (AMJ) and concurrent ENSO indexes values—SOI and Niño 3.4—. The computation of linear correlations allowed to establish these relationships.

A third approach compared the extreme wildfire seasons with ENSO’s categories of intensity. We used the categories proposed by the Australian Bureau of Meteorology. Finally, we expanded this analysis using a $$\chi^2$$ test. In this test, we used categories of seasonal FFDI and ENSO strength. The categories for seasonal FFDI were extreme (FFDI >90th percentile), medium (50th percentile<FFDI<=90th percentile), and low (FFD<=50 percentile). The numbers of ENSO years were added within three categories: very strong El Niño, Weak-Moderate-Strong El Niño, and La Niña. The strength of ENSO years corresponds to the Australian Bureau of Meteorology classification available at http://www.bom.gov.au/climate/enso/enlist/.

# Results

Figure 2 shows PDFs for each wildfire weather variable. These results represent the deviations between original WS and 20CR data. All 20CR PDFs have deviations from the WS data. The daily deviations from the mean including all seasons are 8.9 oC (Tmax), 11 (RH), 1.6 m/s (W) and 4.5 mm (P).

Figure 3 presents time series of the same variables. The values correspond to seasonal averages for Tmax, RH and W during JAS. P corresponds to a seasonal sum. The figure shows Pearson’s linear correlation coefficients between 20CR and WS data for 16 seasons. The results were statistically significant only for W (r=0.52,p=0.04).

The bias-correction process comprised the application of linear scaling techniques for every wildfire weather variable. However, the bias-correction of all variables did not improve the seasonal FFDI calculation (not shown). Only the temperature bias-correction significantly improved the FFDI results. Thus, for this study we only show the bias-correction results for this variable (see Figure 4).

Figure 4 shows PDFs of original and bias-corrected Tmax data. The application of additive linear scaling successfully centers the mean of the 20CR. Additionally, Figure 5 presents original and bias-corrected results of the seasonal time series for Tmax. After the bias-correction, 20CR maximum temperature data has a comparable magnitude with WS data. The linear scaling did not affect the correlation coefficients between the two types of data.

Figure 6 presents the results of the FFDI calculation with original and bias-corrected data. The results show PDFs of WS and bias-corrected 20CR data. The bias-correction increases the similarity between the distribution of data.

Figure 7 presents seasonal (JAS) time series of FFDI. The figure displays three time series. FFDI calculated with WS data comprises the period 1997-2012. Two more time series show FFDI calculated with reanalysis data. The calculation used original (20CR) and bias-corrected (20CR(bc)) reanalysis data. These time series span the period 1963-2012. The FFDI time series calculated with original reanalysis data over-estimates seasonal wildfire potential. The overestimation is—on average—of 65 units during the period 1997-2012. These data—WS and 20CR—have a moderate linear correlation (r=0.53,p=0.04). FFDI calculated with bias-corrected reanalysis data under-estimates the FFDI computed with observed values. The magnitude of the underestimation is less than the initial over-estimation (47 units). The bias-correction does not significantly affect the linear correlation between WS and 20CR(bc) data (r=0.53,p=0.05). These results encouraged us to extend the FFDI calculation for the period 1963-2012.

Having an extended FFDI dataset allowed us to investigate the influence of ENSO over seasonal wildfire variability in Ecuador. Table 2 shows linear correlation’s ’R’ coefficients of seasonal ENSO metrics with FFDI. The table shows how the evolution of ENSO in previous—and the concurrent season—is associated with wildfire potential. Previous seasons start one year earlier on October-November-December (OND), and continue in three-months intervals including the wildfire season July-August-September (JAS). These correlations show that the influence of ENSO is associated to wildfire potential in the preceding (AMJ) and concurrent (JAS) season. ENSO does not seem to have a direct relationship further back in time. This is valid using the SOI and Niño 3.4 indices.

These results encouraged further analysis into the concurrent (JAS) and antecedent (AMJ) seasons. Table 3 presents the results of linear correlations applied to every wildfire variable in the concurrent season, with ENSO indices in the antecedent and concurrent season. The relationships among FFDI variables are physically consistent. The R coefficients for these variables with ENSO indices are moderate are statistically significant at the 5 percent level. This holds valid for all variables with the exception of precipitation.

This investigation also explored the relationship between ENSO and extreme wildfire seasons. Our definition of extreme wildfire seasons filtered years with cumulative FFDI above the 90th percentile for the period 1963-212. Table 4 shows that the most extreme seasons occurred during the years 1982, 1991, 1997, 2009 and 2012. Each season was compared with historical categorizations of strength of ENSO events. With the exception of the year 2012, extreme wildfire weather occurred during El Niño events.

Table 4 expands on the categorical analysis of seasonal wildfire potential and ENSO strength with a $$\chi^2$$ test. The table shows that most very strong El Niño events—two out of four years—have a connection with extreme wildfire seasons. Within the range of Weak-Moderate-Strong El Niño events, most years—13 out 22—occur during low wildfire potential—seasonal FFDI<=50th percentile—. However, two years belong to the extreme wildfire weather category. Finally, most La Niña years occur during medium and low seasonal FFDI—23 out 24 years—.

# Discussion

This study has investigated the variability of seasonal wildfire potential in Ecuador. The analysis included the examination of the relationship of wildfire weather and ENSO phases. Using a seasonal metric of wildfire potential—cumulative FFDI during the season July-August-September—, this study detected that five extreme seasons occurred during the period 1963-2