Volume 33, Issue 8 p. 1129-1145
Research Article
Free Access

Global Fire Forecasts Using Both Large-Scale Climate Indices and Local Meteorological Parameters

Huizhong Shen

Huizhong Shen

Laboratory for Earth Surface Processes, College of Urban and Environmental Sciences, Peking University, Beijing, China

School of Civil and Environmental Engineering, Georgia Institute of Technology, Atlanta, GA, USA

Search for more papers by this author
Shu Tao

Corresponding Author

Shu Tao

Laboratory for Earth Surface Processes, College of Urban and Environmental Sciences, Peking University, Beijing, China

Correspondence to: S. Tao,

[email protected]

Search for more papers by this author
Yilin Chen

Yilin Chen

Laboratory for Earth Surface Processes, College of Urban and Environmental Sciences, Peking University, Beijing, China

School of Civil and Environmental Engineering, Georgia Institute of Technology, Atlanta, GA, USA

Search for more papers by this author
Mehmet Talât Odman

Mehmet Talât Odman

School of Civil and Environmental Engineering, Georgia Institute of Technology, Atlanta, GA, USA

Search for more papers by this author
Yufei Zou

Yufei Zou

Earth and Atmospheric Sciences, Georgia Institute of Technology, Atlanta, GA, USA

Search for more papers by this author
Ye Huang

Ye Huang

Laboratory for Earth Surface Processes, College of Urban and Environmental Sciences, Peking University, Beijing, China

Laboratoire des Sciences du Climat et de l'Environnement/Institut Pierre Simon Laplace, Commissariat à l'Énergie Atomique et aux Énergies Alternatives–CNRS–Université de Versailles Saint-Quentin, Université Paris-Saclay, Gif-sur-Yvette, France

Search for more papers by this author
Han Chen

Han Chen

Laboratory for Earth Surface Processes, College of Urban and Environmental Sciences, Peking University, Beijing, China

Search for more papers by this author
Qirui Zhong

Qirui Zhong

Laboratory for Earth Surface Processes, College of Urban and Environmental Sciences, Peking University, Beijing, China

Search for more papers by this author
Yanyan Zhang

Yanyan Zhang

Laboratory for Earth Surface Processes, College of Urban and Environmental Sciences, Peking University, Beijing, China

Search for more papers by this author
Yuanchen Chen

Yuanchen Chen

Laboratory for Earth Surface Processes, College of Urban and Environmental Sciences, Peking University, Beijing, China

College of Environment, Zhejiang University of Technology, Hangzhou, China

Search for more papers by this author
Shu Su

Shu Su

Laboratory for Earth Surface Processes, College of Urban and Environmental Sciences, Peking University, Beijing, China

Search for more papers by this author
Nan Lin

Nan Lin

Laboratory for Earth Surface Processes, College of Urban and Environmental Sciences, Peking University, Beijing, China

Search for more papers by this author
Shaojie Zhuo

Shaojie Zhuo

Laboratory for Earth Surface Processes, College of Urban and Environmental Sciences, Peking University, Beijing, China

Search for more papers by this author
Bengang Li

Bengang Li

Laboratory for Earth Surface Processes, College of Urban and Environmental Sciences, Peking University, Beijing, China

Search for more papers by this author
Xilong Wang

Xilong Wang

Laboratory for Earth Surface Processes, College of Urban and Environmental Sciences, Peking University, Beijing, China

Search for more papers by this author
Wenxin Liu

Wenxin Liu

Laboratory for Earth Surface Processes, College of Urban and Environmental Sciences, Peking University, Beijing, China

Search for more papers by this author
Junfeng Liu

Junfeng Liu

Laboratory for Earth Surface Processes, College of Urban and Environmental Sciences, Peking University, Beijing, China

Search for more papers by this author
Gertrude K. Pavur

Gertrude K. Pavur

Earth and Atmospheric Sciences, Georgia Institute of Technology, Atlanta, GA, USA

Search for more papers by this author
Armistead G. Russell

Armistead G. Russell

School of Civil and Environmental Engineering, Georgia Institute of Technology, Atlanta, GA, USA

Search for more papers by this author
First published: 19 August 2019
Citations: 15

Abstract

Fire forecasts that predict dry-season fire activities several months in advance are beneficial for fire management. On a global scale, however, the predictability of fires is limited because fires depend on multiple factors and lack a single dominant predictor to describe diverse fire characteristics across regions. Here, based on 33 local meteorological parameters (MPs) and 37 large-scale climate indices (CIs), we establish four empirical model clusters to predict global interannual fire variability. We show that across various geographic locations, the models provide reliable fire forecasts at least three months prior to the peak fire months. Compared to MPs, CIs such as the Oceanic Niño Index are comparable or even superior predictors. Globally, as well as in most continents, the El Niño–Southern Oscillation is the major driving force, explaining 17% of interannual fire variability, with strong implications for fire carbon emissions and the global carbon cycle. Other important predictors include the Northern Atlantic sea surface temperature (9%), the Southern Atlantic sea surface temperature (5%), and the Pacific/North American Pattern (3%). The predictive models reveal a strong interaction between MPs and CIs, indicating potential climate-induced modification of fire responses to meteorological conditions. We show that the newly developed predictive models can benefit future fire management in response to climate change.

Key Points

  • We developed empirical models to provide global fire forecasts 3 months prior to the peak fire months
  • Combining meteorological and climate variables enhances fire predictability
  • The El Niño–Southern Oscillation is a driving factor that affects global interannual fire variability

Plain Language Summary

Integration of climate indices and meteorological parameters enables more accurate fire forecasts globally with a longer forecast length.

1 Introduction

Wildland fires are important natural processes, but they can lead to economic loss, air pollution, ecological damage, biodiversity loss, and threats to human safety (Bowman et al., 2009; Johnston et al., 2012; Marlier et al., 2013). For decades, efforts have been made to identify factors affecting fires in hopes of reducing the negative impacts of fires. Meteorological conditions constrain vegetation growth and control fuel moisture, which directly affects fuel loading and fire propagation (Anderson, 1982; Schroeder & Buck, 1970; Van der Werf, Randerson, et al., 2008). Based on meteorological parameters (MPs), fire weather indices (FWIs) have been derived and applied to provide information for fire management (Anderson, 1982; Schroeder & Buck, 1970; Van Wagner, 1987). By using forecasted MPs from numerical weather prediction models, these FWIs can provide an early warning of the fire danger any time from 4 to 6 hr (Van Wagner, 1987) to several days (UNISDR, 2018) in advance. A recent study revealed that projections of fire season length determined by FWIs show significant correlation with the actual annual burned areas recorded by individual countries (Jolly et al., 2015). Generally, MPs and their derivatives are closely related to fire activities.

Large-scale climate indices (CIs) can also be good indicators of fires occurrence and severity. A wide range of studies have reported relationships between CIs and fire activities in regions across the world (Andela & Van Der Werf, 2014; Armenteras-Pascual et al., 2011; Chen et al., 2016; Chen et al., 2017; Chen et al., 2011; Dixon et al., 2008; Duffy et al., 2005; Field & Shen, 2008; Goodrick & Hanley, 2009; Gudmundsson et al., 2014; Hessl et al., 2004; Holz et al., 2017; Marcos et al., 2015; Mariani et al., 2018; Preisler & Westerling, 2007; Riaño et al., 2007; Shabbar et al., 2011; Swetnam & Betancourt, 1990; Van der Werf, Dempewolf, et al., 2008; Westerling et al., 2002). One typical example is the successful projection of South American fire season severity using CIs of adjacent sea surface temperature (SST) anomalies (Chen et al., 2011; Chen et al., 2013). Some CIs are tightly linked with some MPs, as is commonly referred to as “teleconnection,” which provides one potential pathway linking fires to CIs. However, a holistic picture of CI-fire connections is complex, involving multiple climatic, ecological, and anthropogenic pathways through associations with various fire stages (e.g., fire ignition and propagation) and various fuel properties and conditions (e.g., fuel load, continuity, and moisture content). For fire ignition, some climate patterns such as the El Niño–Southern Oscillation (ENSO) and the Indian Ocean Dipole are found to affect lightning activity, which is directly related to the occurrence of lightning-ignited fires in tropical and temperate regions (Bovalo et al., 2012; Goldammer & Price, 1998; Mariani et al., 2018). Changing climates in cold areas indicated by certain CIs can alter the depth of spring snowpack, which is important because a deep snowpack tends to mediate lightning-ignited fires by decreasing the proportion of effective lightning strikes (Lutz et al., 2009; Mote, 2006). In regions subject to prescribed burning, land clearing, or agricultural waste burning, meteorological conditions prior to or during the fire seasons may affect burning decisions of land managers (Chen et al., 2016; Hessl, 2011; Huang et al., 2018). Many CIs are, by design, indices of large-scale wind patterns, and their variations are often associated with regional changes of wind speed and directions and atmospheric humidity that affect fire propagation (Viegas, 1998; CPC 2017a; CPC 2017b). Some other CIs are near-planetary signals with temperature anomalies and precipitation redistribution, which have more profound impacts on fires (Dai & Wigley, 2000; Knight et al., 2006; Page et al., 2008). On a seasonal time scale, meteorological variations related to some climate patterns may alter fuel loading and moisture levels of ecosystems (Bastos et al., 2013; Chen et al., 2013); on yearly to decadal time scales, climate change induced by large-scale climate oscillations may modify the ecosystem's net primary productivity, species composition, fuel volume, structure, fuel continuity and condition, etc., thus shifting fire regimes (Beckage et al., 2003; Flannigan et al., 2000; Hessl, 2011; Mariani & Fletcher, 2016; Nemani et al., 2003; Taylor & Beaty, 2005).

Generally, CIs and MPs constrain fire activities on different spatial and temporal scales. MPs are associated with local-scale vegetation growth and fuel flammability over a short term (i.e., from days to months), while CIs regulate mesoscale ecological processes, such as changes in vegetation characteristics over a long term (i.e., from seasons to decades; Bowman et al., 2009; Stocks et al., 1998; Westerling et al., 2006). Given that local fire occurrence and activity are subject to both larger- and finer-scale processes, the predictability of interannual fire variability should be enhanced by including both MPs and CIs in one model.

Here, we present an approach to predict the interannual variability of global fire activity based on MPs and CIs. By aggregating satellite-detected fire counts over a 5° × 5° spatial resolution grid during the fire season each year, a fire index was derived to represent the spatial and interannual variations in global fire activity (section 2). We developed four model clusters to simulate interannual variations in the values of the fire index and to perform global fire forecasts. Each model cluster was an ensemble of regression models of individual grid cells and was developed based on a different set of model variables (CIs, MPs, or both). It should be noted that the term “fire forecast” in this study refers to producing a prediction of the value of the index for an upcoming fire season at a given location. We show that the model clusters can predict the values of the index for individual grid cells at least three months ahead of the peak fire months, thus allowing for fire forecasts with a length of several months—the forecast lengths were location dependent (section 2). These models were then used to evaluate contributions of individual CIs to interannual variations in fire activities and fire carbon emissions and to track the relationships between various MPs/CIs and fires on a global scale.

2 Materials and Methods

2.1 The 21-Year Fire Activity Records From Satellite Observations

We used a gridded satellite-based index, the fire activity index (FAI), as an indicator of fire activity. FAI is calculated as follows:
urn:x-wiley:08866236:media:gbc20906:gbc20906-math-0001(1)
where the subscripts i and y represent a specific spatial grid cell and year, respectively; FCi,y is the satellite-detected fire count during the fire season of year y; mean (FCi) is the mean annual fire season fire counts over the entire satellite detection period; and Std (FCi) is the standard deviation of the fire season fire counts over the detection period. Globally, the land area is gridded into 5° × 5° resolution cells. Grid cells with less than 10% vegetation cover (European Commission, 2017) are excluded. FAI is designed to represent interannual variations in fire activities at a given location rather than spatial variations in fire activities between different locations. Equation 1 ensures that the time series of FAIs for a given grid cell has an arithmetic mean of 0 and a standard deviation of 1. FAIs derived from different fire datasets (including satellite fire detections and their derivatives, that is, burned areas and carbon emissions; Arino et al. 2012; Giglio, Descloitres, et al., 2003; Giglio, Kendall, et al., 2003; Giglio et al., 2013; van der Werf et al., 2017) showed strong interannual correlations, with the caveat that some correlations are caused by a lack of independence between datasets (Text S1, Figure S1S3, and Table S1 in the supporting information). We constructed a 21-year record of FAIs covering the period from 1996 to 2016 based on FAIs derived from two satellite fire detection products that together provided the longest temporal coverage—ATSR Algorithm 1 derived from the ATSR fire detection (Arino et al., 2012) for the period 1996–2001 and MOD14CHM derived from the Moderate Resolution Imaging Spectroradiometer (MODIS) Terra fire detection (Giglio, Descloitres, et al., 2003) for the period 2002–2016. We calculated the weighted average values of FAI to represent its regional variations using standard deviations of annual MODIS-detected fire counts as weight factors. The FAI data set is provided in supporting information Data Set S1.

2.2 Model Variables

We introduced multiple predictors, composed of 37 CIs (Table 1 and Figure S4) and 33 MPs (Table S2), and developed four model clusters (M1, M2, M3, and M4) to simulate the FAI variations. The MP variables include temperature, humidity, wind speed, precipitation, and seven Canadian FWIs (Schroeder & Buck, 1970), and average values for MP variables were calculated over each of the following three periods of a year—the fuel accumulation season, the fire season, and the peak fire month (supporting information Text S2 and Figure 1). Specifically, the fire season was determined as the fewest number of months containing 90% or more of the fire counts detected by MODIS in a year (supporting information Text S2). We used reanalysis and forecast MPs reported by the National Centers for Environmental Prediction Climate Forecast System version 2 (Saha et al., 2014), which provides up to 10 months of MP forecasts, to ensure that the models with MP variables can provide fire forecasts (i.e., FWI values) prior to the onset of the peak fire month. The MPs of the fire season (both for reanalysis and forecast MPs) were calculated as the MP means of the first half of the fire season ending in the peak fire month. For example, if the third month of the fire season was the peak fire month in a given grid cell, the MP means for the fire season would be the means of the daily MPs during the first three months of the fire season. Reanalysis MPs (or 0-month forecast MPs) were used for both model training and evaluation, whereas 1 to 3 months forecasts of MPs were used exclusively to test the model predictive capacity at different forecast lengths (see section 5 for details). Yearly MP means were calculated for each of the three periods and the 11 indices, yielding 33 MP predictors as the model variables.

Table 1. Detailed Information on the 37 CIs Used as FAI Predictors in the Model Clusters of M3 and M4
CI no. CI name Description CI groupa Source
1 AMO Atlantic Multidecadal Oscillation (short) G4 ESRL (2017a)
2 ONI Oceanic Niño Index G1 NOAA (2017)
3 NAO North Atlantic Oscillation G4 CPC (2017a)
4 EA East Atlantic index G4 CPC (2017a)
5 WP Western Pacific index G7 CPC (2017a)
6 EP/NP East Pacific/North Pacific Oscillation G2 CPC (2017a)
7 PNA Pacific North American index G5 CPC (2017a)
8 EA/WR Eastern Asia/Western Russia index G1 CPC (2017a)
9 SCAND Scandinavia Pattern G7 CPC (2017a)
10 TNH Tropical/Northern Hemisphere index G11 CPC (2017a)
11 POL Polar/Eurasia index G6 CPC (2017a)
12 PT Pacific Transition index G12 CPC (2017a)
13 EX_Var Explained variance of the 10 teleconnections patterns G6 CPC (2017a)
14 QBO_U30 Quasi-Biennial Oscillation Indices at 30 mb (Anomaly) G8 CPC (2017b)
15 QBO_U50 Quasi-Biennial Oscillation Indices at 50 mb (Anomaly) G8 CPC (2017b)
16 D_SLP Darwin Sea Level Pressure (Anomaly) G1 CPC (2017b)
17 T_SLP Tahiti Sea Level Pressure (Anomaly) G1 CPC (2017b)
18 SOI Southern Oscillation Index G1 CPC (2017b)
19 I_SLP Indonesia Sea Level Pressure (Standardized Anomalies) G1 CPC (2017b)
20 EQ_EP_ SLP Equatorial Eastern Pacific Sea Level Pressure (Standardized Anomalies) G1 CPC (2017b)
21 EQ_SOI Equatorial Southern Oscillation Index G1 CPC (2017b)
22 NINO1 + 2 Extreme Eastern Tropical Pacific SST (0–10°S, 90–80°W; ANOM) G1 CPC (2017b)
23 NINO3 Eastern Tropical Pacific SST (5°N to 5°S,150–90°W; ANOM) G1 CPC (2017b)
24 NINO4 Central Tropical Pacific SST (5°N to 5°S; 160°E to 150°W; ANOM) G1 CPC (2017b)
25 NATL North Atlantic SST (5–20°N, 60–30°W; ANOM) G4 CPC (2017b)
26 SATL South Atlantic SST (0–20°S, 30°W to10°E; ANOM) G3 CPC (2017b)
27 TROP Global Tropics SST (10°S to 10°N, 0–360; ANOM) G1 CPC (2017b)
28 ZT Zonally Average 500-MB Temperature Anomalies (Anomaly) G1 CPC (2017b)
29 EQ_OLR Outgoing Long-Wave Radiation Equator (160°E to 160°W; Anomaly) G1 CPC (2017b)
30 TNA Tropical Northern Atlantic Index G4 ESRL (2017b)
31 TSA Tropical Southern Atlantic Index G3 ESRL (2017b)
32 MEI Multivariate ENSO Index G1 ESRL (2017b)
33 PDO Pacific Decadal Oscillation G2 ESRL (2017b)
34 TNI Trans-Niño Index G5 ESRL (2017b)
35 AAO Antarctic Oscillation G9 CPC (2017c)
36 SF Solar Flux (10.7 cm) G10 ESRL (2017b)
37 GT Global Mean Lan/Ocean Temperature G4 ESRL (2017b)
  • Note. CI = climate index; FAI = fire activity index; ENSO = El Niño–Southern Oscillation; SST = sea surface temperature.
  • a The numbers of CI groups are consistent with the “Group No.” in Table S3
Details are in the caption following the image
Fire-related periods used in this study. The peak fire month is the month with the highest average Moderate Resolution Imaging Spectroradiometer-detected fire count. The fire season is the shortest period in a year containing at least 90% of the yearly fire count. The fire activity indices were derived based on the fire detections during the fire season. To ensure a feasible forecast, the first half of the fire season, which ends in the peak fire month, was used to calculate the meteorological parameter means during the fire season. The fuel accumulation season is the 13-month period ending in the peak fire month.

2.3 Model Development

The four model clusters were built upon different sets of predictors. M1 was developed based on a single predictor—the Oceanic Niño Index (ONI). Following a previous study (Chen et al., 2011), lead time was defined here as the difference between the month of the CI (ONI or other CIs) and the peak fire month for a given grid cell, which represents the time delay between the CI and fire occurrence. Figure 2a illustrates how a yearly time series of corresponding CI values can be derived from its monthly time series based on a specific lead time length and the peak fire month in a given grid cell. The correlation coefficient between the FAIs and the derived CI time series is a function of lead time length (Figure 2e), where the lead time length corresponding to the highest absolute value of the correlation coefficient can be defined as the optimal lead time, a time point showing the strongest connection between the CI and FAIs. The CI time series derived at the optimal lead time was then used as independent model variables for the grid cell. Figure 2 shows the procedure to determine the optimal lead time for the ONI. The minimum for an optimal lead time was set to be 3 months to ensure that the fire forecast provides adequate time for land managers to prepare resources and carry out fire management strategies prior to the peak fire month. The maximum of an optimal lead time was set to be 10 months so that the correlation will not be affected by fires occurring in previous years. It should be noted that the actual lead time between certain CIs and fires in certain locations could be out of the time range of 3–10 months, but, according to our test, increasing the range did not increase the predictive ability of our models significantly. For each 5° × 5° grid cell, M1 uses ONI time series corresponding to the optimal lead time as the independent variable. The regression analysis is based on univariate linear regression via maximum likelihood estimation. Each grid cell is associated with a regression model as a model unit. As mentioned, the model cluster is the ensemble of the model units for all grid cells.

Details are in the caption following the image
Illustration of the model development of M1. (a) Monthly ONI time series. (b) Monthly fire accounts detected by MODIS and the derived annual FAI time series for a representative grid cell. (c) Monthly fire accounts detected by ATSR and the derived annual FAI time series. (d) Final FAI time series combining MODIS-Terra and ATSR A1. The final FAIs are slightly different from the MODIS- and ATSR-derived FAIs in each of their corresponding periods because the FAI time series are normalized after combination. (e) Correlation coefficients (r) between the FAIs and ONI with different lead times. For each grid cell, the r values between the FAIs and ONI are calculated at lead times from 3 to 10 months (r values for lead times between 0 and 2 are also shown here but not considered in the model development). The lead time with the maximum absolute r value is identified as the optimal lead time. (f) Dependence of the FAIs on ONI at the optimal lead time and the regression equation (as one model unit). (g) Geographic distributions of the r values for individual grid cells at their optimal lead times. This procedure is applied to other CIs for M3 development. A similar flow chart for the analysis of another CI, the Atlantic Multidecadal Oscillation, is provided in Figure S5. ONI = Oceanic Niño Index; FAI = fire activity index; MODIS = Moderate Resolution Imaging Spectroradiometer; CI = climate index.
The development of M2 was conducted using multivariate regression analysis via a forward stepwise way based on the Akaike's Information Criteria (AIC; Akaike, 1974). To reduce overfitting, the maximum number of variables in a model unit was set to 4 (supporting information Text S3). The development of M3 followed M1's variable preparation procedure and M2's regression analysis method and variable selection criteria. Specifically, for each grid cell, the CIs significantly correlated with the FAIs (α = 0.1) were selected for the regression analysis. For the same CI, the derived time series could be different across grid cells because the time series were determined by peak fire months and optimal lead times, which are location-specific variables. M4 combined the variables of M2 and M3 and was developed following M3. The maximum number of variables (MPs + CIs) in a M4 model unit was also set to 4. The four model clusters can be expressed as follows:
urn:x-wiley:08866236:media:gbc20906:gbc20906-math-0002(2)
urn:x-wiley:08866236:media:gbc20906:gbc20906-math-0003(3)
urn:x-wiley:08866236:media:gbc20906:gbc20906-math-0004(4)
urn:x-wiley:08866236:media:gbc20906:gbc20906-math-0005(5)
where the subscripts i, y, j, and k represent a specific grid cell, year, MP, and CI, respectively; FAIi,y is the modeled FAI value in grid cell i for year y; b, d, f, and q are intercepts specified by model clusters and model units; and a, c, e, g, and h are the regression coefficients specified by model clusters, model units, and variables. For the variables excluded from a model unit, the regression coefficients are set to 0. Pi and Li are the peak fire month and the optimal lead time for grid cell i, respectively. CIk,Pi,Li,y is the value of CI k in a specific month of year y. The month can be calculated as (Pi − Li). MPj,i,y is the value of MP j for grid cell i in year y. Model uncertainties of M1 to M4 for individual grid cells were directly derived from regression prediction intervals. The uncertainty is expressed as interquartile ranges throughout the text. The MATLAB source codes, which can be used to develop the predictive model clusters, are provided in supporting information Data Set S1.

To test the MP-CI interaction effects in M4, we added MP·CI terms as independent variables in the regression models. This new set of variables were obtained by multiplying each of the 33 MPs with each of the 37 CIs, thus containing a total of 1,221 variables. A new model cluster, which was only used to address MP-CI interaction, was built upon these variables together with individual MPs and CIs using the same model development procedures as M4.

2.4 Model Evaluation and Performance

The four model clusters were evaluated both spatially and temporally. For spatial evaluation, 20 out of 21 years were randomly selected to develop models with the remaining one year for testing them. To avoid climate-induced simultaneous impacts on all model units, random selections of the 20 years were conducted for each grid cell individually. For temporal evaluation, the FAIs from 1996 to 2010 were used to develop the models with FAIs from 2011 to 2016 for evaluation (Figure S6).

To test the model's predictive capacity at different forecast lengths for M2 and M4, 20 out of 21 years were randomly selected to develop models based on the Climate Forecast System version 2 zero-month MP forecast data, with the remaining year for evaluation using MPs with different forecast lengths (1 to 3 months) as model inputs. The residual sum of squares (RSS) between the projected and observed FAIs were calculated to assess the predictive capacity.

To address the influences of geographic resolution on model performance, high-resolution meteorological data (ERA-Interim) from the European Centre for Medium-Range Weather Forecasts (Dee et al., 2011) were used for model training. Model clusters were built and compared at different spatial resolutions. The resolution levels include 0.5°, 1°, 1.5°, 2°, 2.5°, 3°, 4°, 5°, 6°, 10°, 12°, and 15°. At different resolutions, the model performance was evaluated and compared using R2, AIC, RSS, and relative error (δ, calculated as the absolute error divided by the mean of observations) as indicators. R2 is preferred here because it specifies the proportion of the FAI variation that can be explained by a given model. The regional R2 was calculated as the weighted average R2 of all grid cells within the region using standard deviations of annual MODIS-detected fire counts as a weighting factor.

The contributions of individual predictors to FAI variations were quantified for each grid cell as the R2 values multiplied by the ratio of ρ2 (squares of the partial correlation coefficient between variables and FAIs) of the specified predictor to the sum of ρ2 of all predictors. The regional contribution of a specified predictor was calculated as the weighted average value of the contributions of all grid cells within the region using standard deviations of annual MODIS-detected fire counts as a weighting factor.

2.5 Estimation of ENSO Contribution to Fire Carbon Emissions

Figure S1 shows the consistency of interannual variation between regional FAIs and regional carbon emissions from fires. The FAIs in each grid cell comprise a normalized time series with the mean equal to 0 and the standard deviation equal to 1. We employed means and standard deviations of gridded carbon emissions reported by fourth-generation global fire emissions database (GFED4s; Van Der Werf et al., 2017) to denormalize the modeled FAI time series predicted by M3. This denormalization procedure transforms the modeled FAIs to carbon emissions based on the precondition that the time series of gridded FAIs and carbon emissions from fires are highly correlated. Hence, variations of carbon emissions from fires contributed by individual CIs can be evaluated. We compared the M3-predicted carbon emissions with those estimated by GFED4s and the Reanalysis of the Tropospheric chemical composition (RETRO) data sets (Schultz et al., 2008). It should be noted that the estimations of carbon emissions in these two data sets rely on satellite fire detections, especially for the estimation of small fire emissions in GFED4s, which amount to one third of the global fire carbon emission estimates (Van Der Werf et al., 2017). A strong correlation between the M3-predicted and GFED4s/RETRO-estimated carbon emissions is thus expected. Our purpose in applying M3 to the estimation of carbon emissions is to decompose the emissions into fractions contributed by individual/groups of CIs so that the contribution of ENSO could be explored.

The case of emissions during a hypothetical El Niño event (ONI > 1 for 12 months) was addressed in M3 by changing the values of CIs in the ENSO group as follows: ONI was set to 1, and the other CIs in the ENSO group were calculated using the following equation:
urn:x-wiley:08866236:media:gbc20906:gbc20906-math-0006(6)
where VCI is the expected value of a specific CI corresponding to the hypothetical El Niño event; μCI is the 21-year mean of the CI time series; σCI and σONI are the standard deviations of the CI and ONI time series, respectively; and r is the correlation coefficient between the CI and ONI time series. If r ≥ 0, sgn(r) is 1, otherwise, sgn(r) is −1. The CIs in groups other than the ENSO group were fixed to their 21-year average values. Using this set of predictors as the input of M3 coupled with the denormalization procedure, we estimated the global carbon emissions from fires during the hypothetical El Niño event. We also estimated the emissions for an ENSO-neutral case by setting all the ENSO group CIs back to their 21-year means. The influence of the hypothetical El Niño event on carbon emissions from fires could be obtained by comparing the total amounts and spatial distributions of emissions between the two cases.

3 Results and Discussion

3.1 The 21-Year Records of Global Fire Activities

Analysis of the regional FAI temporal trends highlights the El Niño-induced fire anomalies in most regions during the period 1997–1998 (Figures 3 and S7). The FAIs in 1998 are associated with the highest global average FAI value (Mean = 0.2) and the largest spatial discrepancy (SD = 1.0) over the 21-year period (Mean = 0.0, SD = 0.8, on average; Figure S8). This finding indicates the strong influence of the ENSO cycle on global fire activities, with fire responses to extreme ENSO phases differing by region. In most regions, El Niño events (or warm phases of the ENSO) correspond to above-average FAI levels. Reverse situations can be found in Europe and Australia, which show higher FAIs following La Niña events (or cold phases of the ENSO). However, ENSO may not fully explain the FAI variations in all regions—there are observed FAI anomalies in ENSO-neutral years (such as in Europe in 2003 and 2007 and Central and East Asia in 2014), suggesting that some other climate circulations are playing a role in these regions.

Details are in the caption following the image
Observed and model-predicted temporal trends in the regional aggregated FAIs for eight regions and the globe. The interquartile ranges of the M4 predictions are shown as shaded areas. The monthly ONI time series is embodied in each panel as a heat map. The FAIs for each year are plotted in the middle of the year. See Figure S7 for Europe and Figure S2 for region classification. The Middle East has hardly any fire and is thus not shown. MP = meteorological parameter; FAI = fire activity index; CI = climate index; ONI = Oceanic Niño Index.

3.2 Model Evaluation

To evaluate model predictive abilities, we randomly selected 20-year FAIs to develop models leaving 1 year for evaluation. The observed FAIs for evaluation are plotted against predicted ones for evaluation (Figures 4a–4d). From M1 to M4, the correlation coefficients r between predicted and observed values increase from 0.15 to 0.43; the RSS decreases by 32%, highlighting a tremendous improvement of M2–M4 over M1, with M4 showing the best agreement with observations. This finding indicates an enhanced predictability of FAIs by introducing multiple predictors in regression models over univariate models. This is further confirmed by the AIC values (54.0, 45.3, 45.0, and 36.7 for M1 to M4, respectively) and δ of the four model clusters (34.3 ± 27.9%, 28.9 ± 24.9%, 27.7 ± 24.2%%, and 22.7 ± 21.0%, representing the medians and semi-interquartile ranges of δ of the FAI-denormalized fire counts for M1 to M4, respectively). The predictive ability of M4 is particularly high in locations with severe fires and potentially strong ecological impacts indicated by large carbon emissions (large circles in Figure 4d). Note that even with the capping of the number of selected predictors (≤4), overfitting in the models is statistically inevitable, which became apparent when we used a subset of the time series to develop the models (Figure S6). The model performance is worse in the validation period than that in the training period. It is expected that with longer records of satellite observations becoming available, the overfitting issue could be diminished gradually.

Details are in the caption following the image
The spatial predictive ability of M1 to M4. (a–d) Spatial evaluation of four model clusters (M1–M4). For each grid cell, 20 out of 21 years was randomly selected to train the models, and the remaining data were used for evaluation; the results are shown by plotting the model predictions against the observations after normalization. The bubble areas are proportional to carbon emissions from fires (Van der Werf et al., 2010). The FAI annual means (0 after normalization) are used for the grid cells that cannot be significantly regressed. The correlation coefficient (r) and the residual sum of squares (RSS) are shown in each panel. (e–h) Geographic distributions of R2 values of the four models. The coefficients of variance are 36%, 36%, 31%, and 21%; the Moran's I values are 0.66, 0.21, 0.16, and 0.14 for M1, M2, M3, and M4, respectively, indicating the increasingly even distributions of R2 from M1 to M4. Distributions of frequency density (FD) are inserted in the bottom left of each panel. (i) The interquartile and inter-90% ranges of R2 values of the four models for various regions and the world. FAI = fire activity index.

3.3 Model Performance

Using ONI as a single predictor, M1 is subject to ENSO impacts on fires (ENSO-fire teleconnections). The model simulation resembles observed FAI temporal trends only in regions highly constrained by ENSO, such as South America, Southeast Asia, and Australia, as well as some regions at high northern latitudes (Figures 2 and 3). In contrast, M2, M3, and M4 show high agreement with observations in all regions (Figure 3). Spatially, the FAIs in most grid cells (>80%) are significantly regressed by the M4 variables (Figures 4e–4h), and the R2 of M4, which represents the percentage of FAI variation being explained by the model, is the highest (0.70 compared with 0.13, 0.47, and 0.55 for M1, M2, and M3) and the most evenly distributed across grid cells (coefficients of variance are 36%, 31%, 36%, and 21% and Moran's I values are 0.66, 0.21, 0.16, and 0.14 for M1 to M4) and major regions (Figure 4i).

Globally, M2 (MPs) and M3 (CIs) can explain 47% and 55%, respectively, of the FAI variations (R2) for the study period, and the combined model can explain as much as 70%. Although the R2 values of CIs and MPs cannot be linearly aggregated given their strong correlation, they are still partially and separately linked with different aspects of FAI variability because the combined model promotes a significant increase in R2. This conclusion is strengthened by the fact that the R2 values of M2 and M3 exhibit disparate tendencies with changing resolutions (Figure 5a). R2 of M3 (as well as M4) remains constant, while R2 of M2 decreases at finer resolutions (finer than 2° × 2°). It is believed that fires are related to CIs on a regional basis, and thus the FAI variation captured by M3 is close to a spatial covariation that will not change with geographic resolution. In contrast, M2 reflects a local-scale fire response to weather conditions, which tends to be disturbed by human or natural ignition randomness and is also subject to accuracy in weather simulations—finer resolutions are associated with higher uncertainties in MP forecasts. These factors hindering the predictive ability of M2 will likely become more critical at finer resolutions. This result also suggests that CIs are efficient fire predictors of interannual FAI variation at multiple spatial scales.

Details are in the caption following the image
Model performance at different spatial resolutions or with different weather forecast lengths. (a) R2 values of the four models with different resolutions. (b) Dependence of residual sum of square (RSS) between the predicted and observed FAIs on the forecast lengths of MPs in M2 and M4. It should be noted that M1 and M3 are independent of the MP forecast lengths and the lead times of CIs are set between 3 and 10 months. Therefore, the predictive abilities of M1 and M3 do not decrease unless the forecast length is longer than 3 months. The RSS values of M1 and M3, which are independent of forecast length within 3 months, are shown as circles for comparison.

To project FAIs ahead of the onset of peak fire months, fire season MP forecasts are required. We evaluate the dependence of the predictive ability on forecast length of MPs and find a decreasing trend in predictive capacities with increased forecast lengths for M2 and M4 (Figure 5b). This is expected because longer forecast length induces higher uncertainty into MPs, which necessarily lowers model skills. If the forecast length is longer than three months, the predictive ability of M4 might be inferior to that of M3 in some model grid cells because the forecast length of M3 is determined by the optimal lead times of CIs, which vary by grid cell. Our analysis generally suggests that within the forecast lengths of 1–3 months, M4 (a combination of CIs and MPs) provides the best fire forecasts.

3.4 CIs as Driving Forces for Regional FAI Variation

Based on M3, we analyze the attribution of FAI variation to variations in multiple CIs and find that contributions of individual CI vary across regions. Given the large number of CIs being considered in the model, we classify CIs into 12 groups based on cluster analysis (Table 1, Figure S4, and Table S3) and show their contributions here by groups (Figure 6). It should be noted that models are built upon individual CI variables instead of CI groups.

Details are in the caption following the image
Relative contributions of 12 CI groups to FAI variations by region based on M3 performance. Red and blue colors represent high and low contributions, respectively. ENSO = El Niño–Southern Oscillation; SST = sea surface temperature; CI = climate index.

Both the ENSO (or the Southern Pacific SST group) and the Northern Atlantic SST groups show dominant impacts on FAI variations. The ENSO group contributes 17% of the FAI variation globally, with the strongest influence in tropical regions, which will be discussed in detail in the subsequent sections. The Northern Atlantic SST group is associated with the Atlantic Multidecadal Oscillation, a long-duration SST variability with near-global climate impacts (Knight et al., 2006). This group contributes 9% of the global FAI variation and represent strong teleconnections (though less than ENSO) with regional FAIs in almost all regions (R2 being consistently higher than 5% in all regions; Figure 6). The two groups together account for half of the M3-explainable FAI variation. Other CI groups, such as the Southern Atlantic SST group (5% of global FAI variation), the Antarctic Oscillation (5%), the Northern Pacific SST group (3%), and the Pacific/North America Pattern (3%, especially pronounced in North America) also factor into FAI variation in specific regions. For example, the Northern Pacific SST group, which is related to the long-lived ENSO-like pattern of the Pacific Decadal Oscillation (Mantua & Hare, 2002), shows notable relations with FAI variations in eastern Australia and the Boreal Asia, which is consistent with the documented teleconnections between the Pacific Decadal Oscillation and climate conditions in these two regions (Mantua & Hare, 2002; Zhang et al., 1997). Other important CI-fire teleconnections include the West Pacific Pattern and the Quasi-Biennial Oscillation in Southeast Asia, the Solar Flux in Australia and New Zealand, and the Antarctic Oscillation in South America.

3.5 ENSO as a Dominant Signal of Global FAI Variation

As revealed by M3, the ENSO group explains 17% of the interannual variation in the FAIs globally. M3 also highlights existing time lags between the occurrence of warm/cold phases of ENSO and regional fire anomalies. As stated before, time lags are measured in the model as the optimal length of lead time (called “optimal lead time”) between the monthly CIs and the peak fire months when the strongest correlation between the FAI and CI time series is detected. The optimal lead time often occurs several months ahead of the peak fire months and varies widely across regions. At the optimal lead time, the correlation coefficients r between the FAI and CI time series will significantly depart from 0. This is illustrated, in an r-lead time line chart, as a ring-like signal (Figure 7, all the CIs shown are related to ENSO). The bigger the ring is, the stronger the ENSO-fire teleconnection will be. Analysis of the changing correlation over different lead time enables a better understanding of the ENSO-fire teleconnections in different regions.

Details are in the caption following the image
ENSO-fire teleconnections. (a) Global distribution of the correlation coefficients between FAIs and Oceanic Niño Index at the optimal lead time. (b–d) Trends in the correlation coefficients between FAIs and ENSO group CIs with lead times (number of months prior to the peak fire month) ranging from 50 to −10 months in Southeast Asia (b), west coast of equatorial South America (c), and the Brazil Highland (d). The CIs shown in the graphs are all ENSO group CIs. Shaded areas show the distributions of optimal lead time of these CIs. The significant level of 0.05 is marked with dashed lines. ENSO = El Niño–Southern Oscillation; SST = sea surface temperature; FAI = fire activity index; CI = climate index.

For regions adjacent to the tropical western/eastern Pacific Ocean, the optimal lead time is approximately zero months (Figure 7b), reflecting an immediate fire response to extreme ENSO events (Dai & Wigley, 2000). Southeast Asia is characterized by a single large ring signal (Figure 7b), indicating a strong ENSO-fire correlation during the fire season but minimum impacts during other times of the year. Drought conditions in this region have been commonly observed following El Niño events (Hendon, 2003). Considering the fuel-abundant local fire regime (Van der Werf, Randerson, et al., 2008), fire season drought induced by El Niño events can enhance fire activities directly by reducing fuel moisture content, which provides an underlying mechanism leading to the single large ring.

South America, as a comparison, shows two rings with opposite correlation (Figure 7c). El Niño leads to warmer and wetter conditions in this region (McPhaden et al., 1998). The west coast of equatorial South America is characterized by a mixed species composition along the slope of the Ades, where fires are affected by both fire season drought and growing-season fuel accumulation (Van der Werf, Randerson, et al., 2008, Bradley & Millington, 2006). While wetter fire seasons tend to suppress fire activities in this region, wetter growing seasons facilitate fuel accumulation, which increases fire activities subsequently in the following fire season (Van der Werf, Randerson, et al., 2008). The time of the El Niño occurrence is crucial because it determines if the onset time of wetter conditions occurs in the growing or fire season, which will lead to opposite impacts on fires. This relation between El Niño and the local fire responses is reflected by the two-ring signal.

The Brazil Highland ENSO signal is distinguished from that of the west coast of equatorial South America and Southeast Asia. Characterized by extensive grasslands (a fuel-limited fire regime), fuel load instead of drought is the predominant factor in this area (Archibald et al., 2013; Van der Werf, Randerson, et al., 2008); wet conditions, which could be provided by El Niño events, are essential for fuel accumulation during growing seasons. Given a potential time lag between extreme ENSO phases and MP responses, El Niño events launching 6–12 months ahead of the peak fire months are very likely to enhance precipitation in growing seasons and facilitate fuel accumulation, which subsequently results in higher FAIs in the following fire season. This is demonstrated in Figure 7d as a single-ring signal showing a positive correlation between ONI and FAIs with optimal lead times between 6 and 12 months. Unlike the signal along the west coast, no significant negative correlation is found at lead times <5 months, indicating that fire activities in this region are insensitive to fire season phases of ENSO.

In Africa, La Niña events tend to enhance continental rainfall (Nicholson & Kim, 1997). In tropical forest ecosystems of equatorial Africa, fuel is abundant (Van der Werf, Randerson, et al., 2008). Fires are driven by fire season drought. La Niña-induced enhanced rainfall tends to increase moisture contents of fuel and thus prevents fire spread, resulting in lower FAIs. By contrast, in African subtropical arid ecosystems, where the fuel load limits fires, enhanced rainfall during the preceding wet seasons leads to more rapid vegetation growth and fuel accumulation, resulting in higher FAIs in the fire seasons (Van der Werf, Randerson, et al., 2008; Andela & Van Der Werf, 2014). Hence, the correlation analysis shows spatially contrary correlations between ENSO and FAIs in tropical (positive) and subtropical (negative) Africa (Figure 7a; also refer to Figure S9 showing the spatial distribution of r in Africa at a finer resolution).

It should be noted that due to the global dominance of savanna fires, global mean variations in the FAI tend to reflect fire behaviors in savanna ecosystems. However, as shown in Figure S12, regional variations in FAIs reflect fire variations in multiple ecosystems, which is captured by the models. Generally, our results clearly show that the characteristics of ENSO-fire teleconnections vary by regions and fire regimes, which are likely determined by the onset time of the extreme ENSO events, the lag time of ENSO-MP teleconnections, and the distinguished fire responses to MP changes. The empirical models provide linkages and evidence to improve the understanding of CI-related fire dynamics.

3.6 ENSO-Induced Variation in Global Fire Carbon Emissions

By multiplying the predicted FAIs with the standard deviations of gridded annual carbon emissions reported by GFED4s, M3 can be employed to quantitatively address the interannual variation in fire carbon emissions contributed by ENSO and other CIs (section 2). As expected, the M3-predicted time series of global fire carbon emissions are consistent with the estimates by GFED4s and RETRO (Figure 8) in corresponding periods. We then isolated the contribution of ENSO group CIs and quantitatively evaluated the ENSO impact on global fire carbon emissions.

Details are in the caption following the image
M3 reconstruction of the global fire carbon emissions from 1983 to 2016. The interquartile range of the model estimate is shown as shaded area. The time trends of global carbon emissions reported by GFED4s (1997–2015; Van Der Werf et al., 2017) and RETRO (1983–2000; Schultz et al., 2008) are shown for comparison. The monthly ONI time series is shown as a heat map. The interannual variation in fire carbon emissions attributable to ENSO is shown as a bar chart. RETRO = Reanalysis of the Tropospheric chemical composition; ONI = Oceanic Niño Index; ENSO = El Niño–Southern Oscillation; GFED4s = fourth-generation global fire emissions database.

Our assessment shows that the warm phase of ENSO is associated with enhanced fire carbon emissions in most regions, ultimately leading to an overall increase in global carbon emissions during El Niño events (see Figure 8 for the global trend and Table S4 for a regional summary). Tropical regions show the highest emission increase because of both a strong positive ENSO-fire correlation and intensive carbon emissions from fires. A hypothetical El Niño event with ONI exceeding 1 °C and lasting for 12 months would lead to an increase of 340 (200–480 as an interquartile range) Tg carbon emission globally. Southeast Asia accounts for 58% of the increase, followed by South America (13%). We find that the severe El Niño event occurring in 1997–1998 led to an increase of 590 (310–880) Tg in carbon emissions, equivalent to the current annual anthropogenic carbon emission in India (World Bank, 2017). This finding indicates an important pathway for ENSO to modulate atmospheric carbon dioxide (CO2) and the global carbon cycle (Zeng et al., 2005) and may partially explain the observed correlation between atmospheric CO2 and ENSO (Jones et al., 2001). It is also found that the continuous warm phase of ENSO during the early 1990s released nearly 1,000-Tg carbon through fires (Figure 8), mainly because the longer duration enhanced the probability that the warm phase of ENSO occurs at the optimal lead time in multiple regions. In contrast, the short duration of warm phases of ENSO is a likely factor for the recent decreasing trend in global fire carbon emissions alongside the documented human-related drivers (Andela et al., 2017).

3.7 Rethinking the CI-MP-Fire Nexus

Our study reveals that MPs can explain 47% of the interannual variation of global fire activities, and incorporation of CIs yields a significant increase in the model explanation (R2 of M4 is 70%). If CI-MP teleconnections were the only pathway for CIs to affect fire activities, the model explanation would not be markedly increased—CIs may contain additional information in explaining fire variations. One possible explanation is that CIs can better reflect interannual variations of ecological processes. For example, CIs can modify soil water recharge, which constrains evapotranspiration and affects fire activities (Chen et al., 2013). CIs also regulate insect outbreaks (such as spruce budworm in the American Southwest), which limit mesoscale fuel accumulation (Swetnam & Betancourt, 1998). In the long term, CIs are found to alter age structures, species composition, and thus forest vulnerability to fires (Behrenfeld et al., 2001; Swetnam & Betancourt, 1998). The responses of ecological processes to CI variations provide extra connections between CIs and fire activities. As is revealed by Hallett et al. (2004), these connections, though depending on CI-MP teleconnections, are so complex that short-term monthly means of MPs fail to capture the dynamics in an empirical model. This may explain why the incorporation of CIs substantially increases the model predictability and warrants further investigation.

In addition, abnormal phases of climate circulations represented by CIs can alter ecological processes periodically or even permanently, which consequently changes the regional fire responses to MPs (Behrenfeld et al., 2001; Hessl et al., 2004; Swetnam & Betancourt, 1998). Therefore, FAIs may be subject to different MP functions at different CI levels in the models. This can be tested by adding MP·CI products as interaction terms in M4. It was found that these MP·CI product terms are significant in 79% of the model grid cells. Consequently, the model R2 is further enhanced (70% to 76%, the relative contributions of these products together with those of individual MPs and CIs are shown in Figure S10). Globally, 57% of the FAI variations are associated with MP·CI interactions, with a higher proportion in tropical rainforest ecosystems (72%). One of the MPs, the Seasonal Severity Rating, exhibits the strongest interaction with CIs, suggesting that the representativeness of such fire danger rating indices for actual fire severities may depend highly on the CI levels. We interpret the model interactions as the CI-induced functional change in fire responses to MPs. The high applicability of interaction terms in the model indicates that current fire danger rating systems and fire management may lose their suitability in the future due to climate change, which is likely associated with alteration of many CIs. It should be noted that MP·CI is one of many terms to introduce and analyze interactions. Further investigation is needed to explicitly address the dimensions, directions, and magnitudes of the changing fire responses to MPs under climate change.

4 Conclusions

Previous studies have confirmed the adoptability of MPs in early fire warnings on hourly to daily time scales and of CIs in fire projections on yearly to decadal time scales (Hessl et al., 2004; Schroeder & Buck, 1970; Swetnam & Betancourt, 1990). We show that, on monthly to seasonal time scales, combining MPs with CIs allows for increased fire predictability. Such an approach may enable a better understanding of the pathways of CI-fire teleconnections and their impacts on ecological systems and human health. For fire forecasts, our predictive models fill the gap in forecast lengths between weather-based fire early warnings (hours to days) and climate-driven fire projections (years to decades) and should be particularly beneficial to fire management in less-developed countries where early-warning systems for fires are sparse. With longer FAI records becoming available, model skill is expected to be further improved and self-adapted to a changing climate, which is key to future fire management.

Acknowledgments

This study was supported by the National Natural Science Foundation of China (Grants 41390240 and 41571130010), the U.S. Environmental Protection Agency (EPA Grant R835880), the NASA Applied Sciences Program (Grants NNX11AI55G and NNX16AQ29G), and the China Postdoctoral Science Foundation (Grant 2015M580914). Its contents are solely the responsibility of the grantee and do not necessarily represent the official views of the supporting agencies. Further, the U.S. Government does not endorse the purchase of any commercial products or services mentioned in the publication. Supporting information for this article is available online. The data used in this study including the 21-year FAI data, CIs, and MPs data sets, MATLAB codes, and other data for model development are provided in the supporting information Data Set S1. The authors declare no conflict of interest.