In the absence of a dynamical fire model that could link the emissions to the weather dynamics and the availability of fuel, atmospheric composition models, such as the European Copernicus Atmosphere Monitoring Services (CAMS), often assume persistence, meaning that constituents produced by the biomass burning process during the first day are assumed constant for the whole length of the forecast integration (5 days for CAMS). While this assumption is simple and practical, it can produce unrealistic predictions of aerosol concentration due to an excessive contribution from biomass burning. This paper introduces a time-dependent factor , which modulates the amount of aerosol emitted from fires during the forecast. The factor is related to the daily change in fire danger conditions and is a function of the fire weather index (FWI). The impact of the new scheme was tested in the atmospheric composition model managed by the CAMS. Experiments from 5 months of daily forecasts in 2015 allowed for both the derivation of global statistics and the analysis of two big fire events in Indonesia and Alaska, with extremely different burning characteristics. The results indicate that time-modulated emissions based on the FWI calculations lead to predictions that are in better agreement with observations.
For some chemical compounds such as carbon dioxide and carbon monoxide, the total annual emission from biomass burning is comparable to what is emitted from anthropogenic sources (Crutzen et al. 1979). Chemical fluxes arising from fires are therefore a non-negligible source of emissions in forecasting system of the atmospheric composition, such as the European Copernicus Atmosphere Monitoring Services (CAMS). Since 2012, the Global Fire Assimilation System (GFAS) has been a CAMS-operated offline analysis system that estimates fire emissions by converting the energy released during fire combustion into gases and aerosol fluxes (Kaiser et al. 2012; Di Giuseppe et al. 2016a).
GFAS uses observations of fire radiative power (FRP) from two MODIS sensors on board the Aqua and Terra satellites (Kaufman et al. 2003). FRP measures the heat power emitted by fires as a result of the combustion process and is directly related to the total biomass combusted (Wooster et al. 2005). Using land-use-dependent conversion factors, GFAS converts FRP into emission estimates of 44 smoke constituents, such as CO, CO2, CH4, black carbon, and organic matter components of the aerosols (Kaiser et al. 2012). These fluxes are then ingested in the model managed by the CAMS to produce daily forecasts of the chemical composition at the global scale. GFAS converts MODIS observations into emissions and does not explicitly forecast fire behavior or fire evolution. Also the model operated by the CAMS does not include a fire model component for predicting the evolution of fire emissions. The emissions estimated at the initial analysis time by the GFAS are kept constant during the 5-day forecast (Flemming et al. 2015).
Weather is the most important factor in modulating fire intensities where fuel is available (Flannigan et al. 2009). Weather also affects the fire emissions by changing the fuel moisture content (French et al. 2011, 2014; Di Giuseppe et al. 2016a). A simple functional link between weather and fire is provided by the class of fire danger models as they describe the impact that atmospheric conditions have on fuel availability. However simple, they tend to perform very well where biomass moisture is a limiting factor to fire sustainability, as in temperate and tropical rain forests, where fuel is abundant. On the other hand, their performance instead deteriorates in fuel-limited ecosystems such as semiarid, tropical, and temperate savannas and in the Mediterranean. In these regions, the vegetation conditions are persistently hot and dry but ignitions and fuel availability are scarce except on rare and difficult to predict occasions (Di Giuseppe et al. 2016a). Among others, the Canadian fire weather index (FWI) is one of the most commonly used fire danger rating system. It was developed as a metric for fire danger in a standard jack pine stand (Pinus banksiana) typical of Canadian forests (Van Wagner 1974) and it only uses meteorological inputs, such as air temperature, wind speed, relative humidity, and 24-h cumulative rainfall (Van Wagner 1987). The relative simplicity of its formulation has contributed to its popularity and has resulted in its extensive use in other climate regions that are very dissimilar to Canada (Taylor and Alexander 2006), such as Australia (Cruz and Plucinski 2007), New Zealand and Malaysia (Taylor and Alexander 2006), and Indonesia (de Groot et al. 2007; Field et al. 2009). The FWI has been shown to provide useful information worldwide (Di Giuseppe et al. 2016a).
In this paper, we propose the use of the fire weather index to predict changes in fire emissions during the CAMS forecast, building on the idea that there is a strong link between emissions released during fires and the weather. Linear changes in the FWI are translated into linear changes in fire emission; this assumption substitutes the previously adopted persistence assumption in which no evolution takes place from the initial analysis date. The new method is assessed using 5 months of daily forecasts from June to October 2015. As only limited information is available on the mass fluxes released into the atmosphere during fires, the validation of the new scheme is performed indirectly by looking at the overall quality of the forecast from the point of view of aerosol prediction.
FRP observations from MODIS are available at 1-km resolution. GFAS processes these observations by interpolating them onto to a 0.1° regular grid and then converting the average FRP value into dry matter following Wooster et al. (2005). The conversion into emission rates for 44 constituents is performed using the following simple formulation:
where are the land-use dependent conversion coefficients from Heil et al. (2010) expressed in grams of species per kilogram of dry matter. Then, is the mean daily emission for the given species expressed in grams.
The CAMS uses the emissions available at the analysis time t0 to initialize the chemical constituents arising from the biomass burning process (Flemming et al. 2015). The emissions are persistent during the 5-day forecast. Keeping the emission constant is a practical choice as a fire model is not available. Yet this can lead to severe overestimations of the biomass burning contribution, especially in the case of short-lived events. As an improvement, we propose the introduction of a modulation factor to change the fire emissions on the basis of weather conditions. The function chosen is a linear function of daily changes of FWI (Van Wagner 1985, 1987; Stocks et al. 1989). FWI was chosen as it is directly related to the fire intensity, that is, the energy released during the biomass burning process (Van Wagner 1987). Therefore, the assumed link between the FWI and the fire emission is that the higher the FWI, the higher will be the fuel consumption and, hence, the higher the emissions.
The FWI calculation is performed in two steps. In the first step atmospheric temperature, humidity, precipitation, and wind modulate the fuel moisture content of three surface fuel components representative of typical fuel beds in a boreal forest. In the second step the fuel state is used to calculate two fire behavior indices: the rate of fire spread, which mostly depends on the wind, and the total fuel availability for combustion. The combination of these two latter indices provides a unit-less index of general fire intensity potential, generally referred to as the fire weather index (Van Wagner 1985; Stocks et al. 1989). FWI calculations are available daily through the Global ECMWF Fire Forecast (GEFF) system (Di Giuseppe et al. 2016a). To take into account uncertainties in the fire indices due to uncertainties in the weather parameters, GEFF produces an ensemble of values using the 51 weather forecast realizations of the ECMWF ensemble forecasting system (Buizza et al. 1999). Forecasts for temperature, humidity, wind, and precipitation undergo a temporal interpolation to local noon (see Di Giuseppe et al. 2016a); the results are then used as atmospheric forcing for the FWI forecasts. The ensemble mean is assumed to be the best FWI estimation.
The effect of a change in weather conditions on fire emissions in the proposed model is calculated as the daily change in FWI. A modulation factor is defined as
where t is the forecast time. Since FWI is a daily value, t has lead times between 1 and 5 days. The set of initial conditions, t0 = 0, is the starting day of the forecast and is also referred to as the analysis time. This corresponds to the day at which GFAS emissions (t0) are calculated from the FRP observations. For convenience, the daily value of is scaled to provide the relative change of FWI with respect to this initial day, FWI(t0). During the CAMS integration the daily fire emissions change according to
By substituting Eq. (2) into Eq. (3), it is clear that linear increments in the FWI are transformed into linear increments in emissions for each of the constituents. The assumption of linearity between FWI and the emissions has been verified by looking at the relationship between FWI and FRP, where FRP values are themselves linearly related to the emissions. Figure 1 takes into account all FRP > 0.5 W m−2 data points observed by MODIS (Aqua plus Terra) during 5 months from June to October 2015 and the corresponding FWI calculated from the GEFF. Data have been aggregated over a regular 9-km grid and three functions were tested. The linear model has a correlation coefficient that is similar to the quadratic model and was chosen because of its simplicity.
Since measures changes in fire danger conditions and not fire activities, it is defined for locations where fires are not occurring (i.e., where observed FRP = 0). These locations are not relevant in our application and are therefore masked out. Also, when FWI(t0) = 0, (t) becomes undetermined at all times. In these rare cases, the system reverts to the persistence assumption and (t) = (t0), t. However, Di Giuseppe et al. (2016a) showed that globally observed fires associated with zero FWI value are extremely rare events with a probability of less than 1%.
Since FWI is positive defined and open ended, can in theory vary in the interval [−∞, +∞]. This could lead to unrealistic modifications of the biomass burning emissions during the forecast. However, the FWI formulation has a strong build-up component through the drought code (Van Wagner 1985), which is a smooth function of time that prevents large daily variations in FWI for a given location. In practice, (t) is constrained into a physically meaningful range of values and no rescaling was found to be necessary (see Fig. 2).
In summary, even if (t) cannot be considered a full dynamical fire model with interactive vegetation, it is a direct function of temperature, relative humidity, precipitation, and wind through the FWI formulation. It therefore provides a physical base to simulate the emission evolution, which also is consistent with the meteorology used by the CAMS chemical transport model.
3. Experiment setup
To understand the impact of the fire emission modification in the CAMS some extensive fire events in 2015 were analyzed. The year 2015 was characterized by a strong El Niño that favored extensive fires in Indonesia in the autumn (Benedetti et al. 2016). The fires in Alaska were also particularly severe in June and July. A series of experiments covering both events was performed using an experimental version of the CAMS model that releases biomass burning aerosol emissions at an injection height computed by a plume rise model that has recently been embedded into GFAS (Rémy et al. 2017). There are, however, significant uncertainties associated with these injection height estimates, which has an impact on the forecast plume. Two different runs were carried out to cover the Alaska fires during June–July 2015 and fires in Indonesia during September–October 2015. The control run uses the standard GFAS v1.2 configuration of Kaiser et al. (2012), where persistence is assumed during the 5-day forecast integration. This experiment is simply flagged as the control (CTRL). The second experiment uses the GFAS v1.2 configuration but applies the modification explained in the previous section where the emissions are modulated by the FWI function. This experiment is tagged as FWI.
As a benchmark model simulation, we also use the CAMS interim reanalysis dataset, CAMSiRA, in which, among other observations, fire emissions are constrained to the observed GFAS values (Flemming et al. 2017). CAMSiRA comprises 6-hourly reanalyses of the atmospheric composition for the period 2003–15. It has a horizontal resolution of about 110 km on a reduced Gaussian grid and a vertical discretization of 60 levels from the surface to a model top of 0.1 hPa. The most relevant differences between the CTRL and FWI experiments are that CAMSiRA uses an older model version than the operational model and several observations are assimilated during the reanalysis integration. Total columns of carbon monoxide from the Measurements of Pollution in the Troposphere (MOPITT) instrument (Deeter et al. 2003), MODIS aerosol optical depth (AOD), and several ozone total column and stratospheric profile retrievals constrain the values of the AOD fields. Therefore, differences among the reanalyses, the CTRL, and the FWI cannot be attributed to the different treatments of the fire emissions alone. Still, the comparison helps in sizing the improvement provided by the FWI implementation.
Two horizontal and vertical resolutions were tested. The high-resolution experiment has a TL-1279 horizontal spectral resolution, which corresponds to a grid-box size of about 9 km, and has 137 vertical hybrid sigma-pressure levels. Experiments at the operational CAMS resolution TL-255 (which corresponds to a grid-box size of about 80 km) and with 60 vertical hybrid sigma-pressure levels were also performed. Table 1 summarizes the experiment setup.
a. Modulation factor behavior
FRP changes due to the introduction of the modulation factor are shown as the global probability density function (pdf) of , as a function of forecast time (Fig. 2). Emission changes take place only in the FWI experiment and is zero at all lead times in the control experiment. The pdf comprises data for 154 starting dates and only for observed fire locations, where FRP > 0 at t = t0. Ninety-five percent of the forecast increments are in the range [−0.4, +0.4]. The distribution is centered around zero for all forecast lead times (maximum bias −0.02). This is a very desirable property of the modulation factor as it is not justified by the positive or negative bias into the forecast emissions as a function of lead time. Numerically, could occasionally assume extremely large positive and negative values in the tails of the distribution. The range of is cropped to [−0.4, +0.4] to avoid these cases in the forecast experiments.
Daily variations in FWI translate into increments, which are in a reasonable range at all lead times (Fig. 2). It is more challenging to verify if these increments are also reducing model errors when compared with observations given the shortage of measurements of fire emissions. However, when MODIS FRP observations are available, they can be compared with the forecast FRP (Fig. 3). FRP spans a vast range of values, with a mean of 0.7 W m−2 and 3 W m−2 standard deviation (not shown). Assuming persistence from the previous day produces an increase in the assimilated fire radiative power that is, on average, 0.14 W m−2, 20% higher than what is observed. This bias is reduced to −0.06 W m−2 at lead time of 1 day and to −0.07 W m−2 at lead time of 5 days with the application of the modulation factor.
One interesting aspect is that positive biases are reduced more than negative ones. This behavior can be explained by noting that by using FWI values only when fire events are observed the FWI distribution is conditionally sampled toward high values (Di Giuseppe et al. 2016a). This is evident from Fig. 4, where the PDF of the normalized FWI at the analysis time (t0) is shown. The vast majority of FWI values are above the upper quartile (>75%), as one would expect. From Fig. 4 it is also clear that FWI reaches an asymptotic high value above which the FWI increase is limited. However, precipitation and humidity rises can produce a large FWI decrease, which will translate into large negative values for the modulation factor. Despite this asymmetric behavior, the error in FRP is highly reduced (mean bias = 0.16 W m−2).
b. Case studies
The Indonesian fires in 2015 extended to vast areas and were long lived. The fires started in early September and only diminished in intensity by the beginning of November, thanks to the onset of seasonal precipitation. Human-caused fires are not unusual in Indonesia, mostly as a result of logging practices (Field et al. 2009; Spessa et al. 2015). Further, in 2015 the effect of human practices was amplified by the extensive and persistent droughts that affected parts of Southeast Asia. These conditions contributed to making the situation far worse than usual. Areas of forest, particularly on peat, that are normally too wet to burn turned into fires, leading to extremely widespread and severe burning activities. The fire season began in August, and by September much of Sumatra, Kalimantan, and Singapore and parts of Malaysia and Thailand were covered in thick smoke, which affected the respiratory health of millions of people (Marlier et al. 2013). At its peak, visibility was reduced to less than 10% of normal in places, similar to conditions in El Niño years (Wang et al. 2004). Large parts of Borneo could not be seen from space. Preliminary estimates suggest that greenhouse gas emissions from the burning exceeded those of Japan’s mean annual fossil fuel emissions (Benedetti et al. 2016). The remnant pollution stretched halfway around the equator, even after the worst of the fire was over.
In summer 2015 intense wild fires also affected Alaska. Because of the unusually warm conditions, fires burned all the way down to the mineral soil. When this occurs, the frozen ground loses its insulation and the permafrost can thaw, sometimes so much that the ground sinks and becomes bumpy and hilly as it loses solid ice mass. The fire characteristics in boreal forests are very much different from the fires in Southeast Asia and those in Indonesia. Since these latter burns were sustained by an almost unlimited peat fuel availability, they tended to be long lasting and of low intensity while forest fires are usually short lived and intense. The boreal forest zone consists of a mixture of conifers (white and black spruce, jack pine, tamarack, and balsam fir), for which the FWI model was specifically calibrated; we therefore expect the FWI to be a better predictor of fire danger for this type of vegetation cover.
To gain an understanding of how the new modulation factor modifies the CAMS emissions, it is useful to consider the Alaska fire event in 2015. This was a collection of intense but localized fires that had their maximum of intensity on 7 July 2015, which was taken as the reference forecast. This study concentrates on the first 48 h of the forecast, but the same results were found to be valid at all lead times. By analyzing the difference between the FRP observed at t0 (7 July 2015) and t0 + 48 h (9 July 2015) (Fig. 5), it is clear that a decrease in fire activity was recorded.
The modulation factor is also negative, showing how the decrease in fire activity was correctly predicted by the change in FWI (Fig. 6). The magnitude of the modulation was stronger and more localized in the high-resolution run (Fig. 6b) in comparison with the simulation performed at the operational CAMS resolution (Fig. 6a). The spatial localization of fires was also improved by the high-resolution experiment.
As the modulation was negative at t0 + 48 h, the forecast aerosol emissions were also decreased in the experimental run when compared with the standard CAMS configuration that uses persistence at all lead times (Fig. 7). The effect is remarkable in the high-resolution experiment but becomes negligible at the standard CAMS resolution (not shown) as a consequence of the lower values.
Given the scarcity of the observations, it is challenging to verify if this reduction is also improving CAMS model errors. An encouraging result is nevertheless provided by the comparison with the AERONET station (Holben et al. 1998) at Bonanza Creek (64.997°N, 142.379°W), which is in the middle of the fire event and provides measurements of aerosol optical depth at 0.55 μm that can be compared with the model simulations (Fig. 8).
The analysis of the Alaska fire allows for a detailed understanding of how the introduction of works to modulate emissions during the forecast integration. Being an isolated event, it does not provide strong statistical evidence of the added benefit.
To understand the potential impact of on a global scale, the 2015 Indonesia fires were also analyzed. The fires were very extensive, covering most of Indonesia, and were so persistent that they produced a significant increase in the global aerosol emissions for that year (Benedetti et al. 2016). Model skill scores are calculated over the 65 AERONET stations in Southeast Asia. The small but visible reduction in bias and of root-mean-square errors (RMSEs) at all lead times is encouraging (Figs. 9c,d). As the 2015 Indonesian fire produced a planetary-scale signal in the aerosol emissions, it is worth also analyzing the global skill scores using measurements from all of the 215 AERONET stations over the period 1 September–1 November. The reduction in model biases and root-mean-square error is a very good indication of the improved performance of the model when fire emissions are modulated according to weather conditions. The improvement is less pronounced, even if still significant, in the high-resolution experiments (not shown).
c. Comparison with reanalysis
Verification is performed against the global AERONET network (Fig. 10). Only the first 24 h of forecast were used in the comparison. The reanalysis bias is, as expected, substantially smaller than the operational forecast since AOD observations are assimilated during the 24 h. The remaining bias represents the uncertainty due to the model component. The operational forecast, CTRL, displays a persistent negative bias that is constantly improved by the introduction of the new scheme. Also the RMSE is, in general, improved with the introduction of the new FWI scheme.
There are several studies that show the difficulties of estimating fire emissions starting from fire detection. For example, French et al. (2011, 2014) show how fuel consumption is dependent on fuel moisture and how this is mostly controlled by weather. Not including the influence of atmospheric condition on fire burning produces large errors in the estimation of the associated emission.
This study has explored a simple way to include the effects of weather on fuel combustion by using changes in FWI to predict changes in the aerosol emissions from fires during the forecast. The differences between the operational forecast produced by the atmospheric composition model operated by the CAMS, which uses persistence of the fire contribution observed at the beginning of the forecast, and forecasts made by the new schemes indicated that FWI is a good predictor for fire intensity. FWI also provides a good metric for reducing or enhancing emissions with lead times up to 5 days.
A statistical analysis was performed for the Indonesian fires in 2015, which were extremely long lived (September–November) and of wide geographical extent due to a strong ENSO. The calculation of mean model biases and root-mean-square errors for the Southeast Asia AERONET stations and even the global AERONET network showed a small but positive benefit provided by the introduction of the FWI modulation factor at all lead times. However, even if this new approach is thought to be an improvement over persistence, it has to be stressed that it is still a simplification relative to the use of a fire predictive model, which could include interactive processes between vegetation, fuel, and weather.
This work was funded by EU Horizon 2020 Project ANYWHERE (Contract 700099) and Global Fire Contract 389730 between the Joint Research Centre and ECMWF. An extended version of the material presented in this paper is also available as a non-peer-reviewed ECMWF technical report (Di Giuseppe et al. 2016b). We also thank Luke Jones at ECMWF for the support in the AERONET verification. We are grateful for the comments of three anonymous reviewers.