**3.1 Burned area**

In theory, remotely sensed data should offer the capability to directly quantify atmospheric emissions from fire events, but in practice this requires determining the source of emissions which involves complex, computationally demanding inversion and geochemical transport modeling. Therefore most current approaches, referred to as the "bottom up" method in this paper, involve multiplying the fuel consumed by an emission factor for the atmospheric

The Science and Application of Satellite Based Fire Radiative Energy 181

Merlet, 2001; van der Werf et al., 2003; French et al., 2004; Korontzi et al., 2004). Although datasets used for this application are always improving (Roy et al., 2005; van der Werf et al., 2006), due to the uncertainty in current estimates it is worthwhile to explore other

Vegetation fires can be thought of as the obverse of photosynthesis in which energy stored

 (C6H10O5)n + O2 + *ignition temperature* → CO2 + H2O + *heat* (3) The cascade of chain of reactions starts with the pre-heating of fuels ahead of the fire front and partial pyrolytic decomposition. Ignition signifies the transfer from pre-heating to combustion in which exothermic reactions start and the next phase, encompassing a combination of flaming and smoldering combustion, begins. Flaming combustion occurs when flammable hydrocarbon gases released during pyrolysis are ignited with wildfire flaming combustion temperatures in the range of 800 – 1400 K (Lobert & Warnatz*,* 1993). Pyrolytic action involves the thermal decomposition of fuel resulting in the release of water, CO2, and other combustible gases (e.g. CH4) and particulate matter. The heat produced, often measured as heat yield (MJ/kg), is thermal energy transferred via conduction, convection, vaporization, and radiation and provides a metric of the total potential energy released if complete combustion of the fuel occurs. Although other factors, including slope, fuel arrangement, and wind speed influence the actual heat yield in a fire event, the theoretical value varies very little between fuel types (Stott, 2000; Whelan, 1995). As described by Stefan-Boltzmann's Law, the radiant component is emitted as electromagnetic waves traveling at the speed of light in all directions and is proportional to the absolute temperature of the fire (assumed to be a black body) raised to the fourth power. The relationship between fire temperature and spectral radiance was shown to closely match the Stefan-Boltzmann law (Radiance = σT4) and thus a simple equation incorporating the sample size, emissivity of the fire (with some assumptions needed), and Stefan's constant could provide the rate of fire radiative energy, or fire radiative power (FRP), emitted as

where *A*sample is the total area of the satellite pixel (m2), *ε* is the fire emissivity, σ is the Stefan Boltzmann constant (5.67x10-8J-1m-2K-4), *A*n is the fractional area of the *i*th thermal

The foundation for using measurements of FRP is based on the fact that the rate of biomass consumed is proportional to the rate of FRE. Kaufman et al. (1996, 1998) suggested that estimates of fuel load combustion and emission rates could be made from satellite observations of the radiative energy liberated during fire events. The hypothesis is that the rate of emitted energy (i.e. FRP), and rate of fuel combustion are proportional to the fire size and fuel load (*A* and *B*, respectively) from equation [1]. It follows then that the rate of energy released is directly related to the rate of particulate matter and trace gas emissions. Integrating FRP over the lifespan of the fire event provides the total fire radiative energy (FRE) released, which in turn is directly proportional to the total fire emissions. It is the radiative component that is estimated from Earth observing satellite sensors, offering an

component, and *T*n4 is the temperature of the *i*th thermal component (K).

FRP = *A*sample εσ Σ *A*<sup>n</sup> *T*n4 (4)

approaches.

shown in equation (4).

in biomass is released as heat (equation 3).

**3.2 FRP** 

species of interest. Emission estimates for natural and anthropogenic ignited vegetation fires are generally calculated using spatially explicit measures of pre-fire fuel loads, fuel consumption, and the areal extent of fire impact. The model presented by Seiler & Crutzen (1980) has been used extensively to quantify the mass of fuel consumed:

$$\mathbf{M} = \mathbf{A} \bullet \mathbf{B} \bullet \boldsymbol{\beta} \tag{1}$$

*M* is the total dry biomass consumed (kg); *A* is the burned area (km2); *B* is the biomass or fuel load (kg km-2); and *β* is combustion efficiency (fraction of available fuel burned). Adapting this algorithm to calculate the emission of a particular species requires *a priori*  information about the emission factor for a given species for the type of vegetation being burned, expressed as grams of species *x* per kilogram of dry fuel burned (Andreae & Merlet, 2001). The equation to estimate emission is then rather straightforward given the fuel consumed, as calculated in equation (1):

$$\mathbf{E}\_{\mathbf{x}} = \mathbf{E} \mathbf{F}\_{\mathbf{x}} \bullet \mathbf{M} \tag{2}$$

*Ex* is the emission load of species *x* (g); *EF*x is the emission factor for species *x* for the specific vegetation type or biome (g kg-1); and *M* is the biomass burned in equation [1].

Traditionally, estimates have been made using statistical information such as FAO data on population and land use practices (Robinson, 1989; Seiler & Crutzen, 1980). Statistical information reported at national scales often requires extrapolation due to incomplete information, sporadic reporting, and highly variable estimates, especially within developing countries undergoing rapid land use change (Andreae, 1991). Advances in satellite technology offer the opportunity to make relatively accurate estimates for several of the parameters in equation [2] at synoptic scales (Justice et al., 2002). For example, Michalek et al. (2000) effectively combined Landsat TM data and field measurements to estimate carbon release from Alaskan spruce forest fires because of the high spatial resolution (<30m) and spectral ability to separate between pre-burn biomass and burn severity of Landsat. Page et al. (2002) estimated 0.19-0.23 Gt of carbon were released to the atmosphere from peat combustion during 1997 Indonesian fires. Their estimates were based on peat thickness, pre-fire land cover, and burnt area data collected from ground measurements and Landsat TM/ETM imagery. Satellite imagery proved useful for classifying land cover and determining burn scars, but Page et al. (2000) discovered that due to residual haze after fires and frequent cloud cover, the use of synthetic aperture radar (SAR) was necessary to determine the extent of burnt areas.

Limitations to the "bottom up" approach include fuel loads and burning efficiency, which cannot be directly estimated from satellite observations. In addition, there is a lack of agreement on the proper algorithm to characterize burned area from satellite data (Roy et al., 2005). Korontzi et al. (2004) showed that significant differences between burned area algorithm estimates can lead to differences as large as a factor of two in estimates of biomass consumed. Differences in spatial and temporal estimates of burned area products was demonstrated by Boschetti et al. (2004) who showed that the GLOBSCAR product had a burned area nearly twice as large as the GBA2000. They concluded that such discrepancies have serious implications for accurately quantifying emissions from fires (Boschetti et al., 2004). The difficulty in accurately measuring these variables leads to an uncertainty in emission estimates of at least 50%, and possibly much greater (Robinson, 1989; Andreae and Merlet, 2001; van der Werf et al., 2003; French et al., 2004; Korontzi et al., 2004). Although datasets used for this application are always improving (Roy et al., 2005; van der Werf et al., 2006), due to the uncertainty in current estimates it is worthwhile to explore other approaches.

#### **3.2 FRP**

180 Remote Sensing of Biomass – Principles and Applications

species of interest. Emission estimates for natural and anthropogenic ignited vegetation fires are generally calculated using spatially explicit measures of pre-fire fuel loads, fuel consumption, and the areal extent of fire impact. The model presented by Seiler & Crutzen

*M* is the total dry biomass consumed (kg); *A* is the burned area (km2); *B* is the biomass or fuel load (kg km-2); and *β* is combustion efficiency (fraction of available fuel burned). Adapting this algorithm to calculate the emission of a particular species requires *a priori*  information about the emission factor for a given species for the type of vegetation being burned, expressed as grams of species *x* per kilogram of dry fuel burned (Andreae & Merlet, 2001). The equation to estimate emission is then rather straightforward given the fuel

*Ex* is the emission load of species *x* (g); *EF*x is the emission factor for species *x* for the specific

Traditionally, estimates have been made using statistical information such as FAO data on population and land use practices (Robinson, 1989; Seiler & Crutzen, 1980). Statistical information reported at national scales often requires extrapolation due to incomplete information, sporadic reporting, and highly variable estimates, especially within developing countries undergoing rapid land use change (Andreae, 1991). Advances in satellite technology offer the opportunity to make relatively accurate estimates for several of the parameters in equation [2] at synoptic scales (Justice et al., 2002). For example, Michalek et al. (2000) effectively combined Landsat TM data and field measurements to estimate carbon release from Alaskan spruce forest fires because of the high spatial resolution (<30m) and spectral ability to separate between pre-burn biomass and burn severity of Landsat. Page et al. (2002) estimated 0.19-0.23 Gt of carbon were released to the atmosphere from peat combustion during 1997 Indonesian fires. Their estimates were based on peat thickness, pre-fire land cover, and burnt area data collected from ground measurements and Landsat TM/ETM imagery. Satellite imagery proved useful for classifying land cover and determining burn scars, but Page et al. (2000) discovered that due to residual haze after fires and frequent cloud cover, the use of synthetic aperture radar (SAR) was necessary to

Limitations to the "bottom up" approach include fuel loads and burning efficiency, which cannot be directly estimated from satellite observations. In addition, there is a lack of agreement on the proper algorithm to characterize burned area from satellite data (Roy et al., 2005). Korontzi et al. (2004) showed that significant differences between burned area algorithm estimates can lead to differences as large as a factor of two in estimates of biomass consumed. Differences in spatial and temporal estimates of burned area products was demonstrated by Boschetti et al. (2004) who showed that the GLOBSCAR product had a burned area nearly twice as large as the GBA2000. They concluded that such discrepancies have serious implications for accurately quantifying emissions from fires (Boschetti et al., 2004). The difficulty in accurately measuring these variables leads to an uncertainty in emission estimates of at least 50%, and possibly much greater (Robinson, 1989; Andreae and

vegetation type or biome (g kg-1); and *M* is the biomass burned in equation [1].

M A B β (1)

E EF M x x (2)

(1980) has been used extensively to quantify the mass of fuel consumed:

consumed, as calculated in equation (1):

determine the extent of burnt areas.

Vegetation fires can be thought of as the obverse of photosynthesis in which energy stored in biomass is released as heat (equation 3).

$$(\text{C}\_6\text{H}\_{10}\text{O}\_5)\_n + \text{O}\_2 + \text{ ignition temperature} \rightarrow \text{CO}\_2 + \text{H}\_2\text{O} + \text{heat} \tag{3}$$

The cascade of chain of reactions starts with the pre-heating of fuels ahead of the fire front and partial pyrolytic decomposition. Ignition signifies the transfer from pre-heating to combustion in which exothermic reactions start and the next phase, encompassing a combination of flaming and smoldering combustion, begins. Flaming combustion occurs when flammable hydrocarbon gases released during pyrolysis are ignited with wildfire flaming combustion temperatures in the range of 800 – 1400 K (Lobert & Warnatz*,* 1993). Pyrolytic action involves the thermal decomposition of fuel resulting in the release of water, CO2, and other combustible gases (e.g. CH4) and particulate matter. The heat produced, often measured as heat yield (MJ/kg), is thermal energy transferred via conduction, convection, vaporization, and radiation and provides a metric of the total potential energy released if complete combustion of the fuel occurs. Although other factors, including slope, fuel arrangement, and wind speed influence the actual heat yield in a fire event, the theoretical value varies very little between fuel types (Stott, 2000; Whelan, 1995). As described by Stefan-Boltzmann's Law, the radiant component is emitted as electromagnetic waves traveling at the speed of light in all directions and is proportional to the absolute temperature of the fire (assumed to be a black body) raised to the fourth power. The relationship between fire temperature and spectral radiance was shown to closely match the Stefan-Boltzmann law (Radiance = σT4) and thus a simple equation incorporating the sample size, emissivity of the fire (with some assumptions needed), and Stefan's constant could provide the rate of fire radiative energy, or fire radiative power (FRP), emitted as shown in equation (4).

$$\text{FRP} = A\_{\text{sample}} \text{ ɛɔ}{\text{ }} A\_{\text{n}} \ T\_{\text{n}} 4 \tag{4}$$

where *A*sample is the total area of the satellite pixel (m2), *ε* is the fire emissivity, σ is the Stefan Boltzmann constant (5.67x10-8J-1m-2K-4), *A*n is the fractional area of the *i*th thermal component, and *T*n4 is the temperature of the *i*th thermal component (K).

The foundation for using measurements of FRP is based on the fact that the rate of biomass consumed is proportional to the rate of FRE. Kaufman et al. (1996, 1998) suggested that estimates of fuel load combustion and emission rates could be made from satellite observations of the radiative energy liberated during fire events. The hypothesis is that the rate of emitted energy (i.e. FRP), and rate of fuel combustion are proportional to the fire size and fuel load (*A* and *B*, respectively) from equation [1]. It follows then that the rate of energy released is directly related to the rate of particulate matter and trace gas emissions. Integrating FRP over the lifespan of the fire event provides the total fire radiative energy (FRE) released, which in turn is directly proportional to the total fire emissions. It is the radiative component that is estimated from Earth observing satellite sensors, offering an

The Science and Application of Satellite Based Fire Radiative Energy 183

process at 5 to 10 second intervals. Wooster et al. (2003, 2005) expanded on their previous work, providing additional evidence of the effectiveness of using instantaneous and total FRE measurements to estimate biomass consumed from fire. Wooster & Zhang (2004) demonstrated the application of MODIS FRP observations by verifying the often proposed hypothesis that North American boreal fires are generally more intense than Russian boreal fires, while Ichoku & Kaufman (2005) used the MODIS FRP and aerosol products to derive near real time rates of aerosol emissions at regional scales. Research by Roberts et al. (2005) has shown the effectiveness of using geostationary satellite estimates of FRP from The Spinning Enhanced Visible and Infrared Imager (SEVIRI) to quantify rates of fuel consumption and characterize the fire intensity daily cycle. A laboratory investigation of FRE and biomass fuel consumption by Freeborn et al. (2008) supported the accuracy of Wooster et al.'s (2005) findings and lends credence to the application of satellite based measurements of FRE. Ichoku et al. (2008a), in a coordinated effort with research conducted by Freeborn et al. (2008), used laboratory investigations to examine rates and total fire radiative energy emitted and associated aerosol emissions. In both the case of Freeborn et al. (2008) and Ichoku et al. (2008a), the relationship between energy emitted, fuels consumed, and trace gas and aerosol emission demonstrated the efficacy of using FRE. Ichoku et al. (2008b) offered another example of using FRP, but at continental scales while investigating the global distribution of MODIS-based FRP estimates and revealed the regional distributions of fire intensity. Their research also showed significant differences in diurnal cycles and categorized intensities of FRP between regions which could not be explained by ecosystem type alone, suggesting perhaps that land use is a factor. Roberts & Wooster (2008) built upon their previous research (Roberts et al., 2005), showcasing the application of high temporal satellite based FRP measurements from the SEVIRI geostationary sensor to calculate FRE and estimate biomass combusted. Boschetti & Roy (2009) demonstrated a novel fusion approach to derive FRE based on temporal interpolation of MODIS FRP across independently derived burned area estimates. Their work was limited to Australia and the MODIS sensor, but as the authors suggest, the methodology could be expanded to other sensors and "is a fruitful avenue for future research and validation" (Boschetti & Roy, 2009). Freeborn et al. (2009) used frequency density distributions developed from MODIS and SEVIRI fire radiative power to synthesize the two sensors as a means for cross-calibration of their respective estimates. However, until Ellicott et al. (2009) and Vermote et al. (2009) no study had derived FRE at a global scale, in part due

to limitations in temporal or spatial resolution of satellite sensors.

spans from 2001 – 2010 (Figure 1).

A current limitation of fire energy retrieval from satellites is that observations are of instantaneous energy (power) over some discrete length of time and space. To address this Ellicott et al. (2009) developed a unique approach to parameterize the temporal trajectory of FRP and calculate the integral (i.e. FRE) using MODIS. The parameterization was based on the long term ratio between Terra and Aqua MODIS FRP and diurnal measurements of FRP and fire detections made by satellites with greater temporal resolution. This included the geostationary sensor SEVIRI and the VIRS aboard TRMM. VIRS's low-inclination orbit (35°) provides observation times which precesses through 24 hours of local time every 23-46 days, depending on latitude, thus capturing the general diurnal trend of fire activity. In addition, high latitude (and thus high overpass frequency) daily observations by MODIS were included. The result was a global FRE product from MODIS at 0.5° spatial and monthly temporal resolution which currently

alternative method to quantify the biomass consumed, and assuming an emission factor is known, it also offers the atmospheric emission load.

Unfortunately, sensors are unable to separate the spatially distinct components of the fire, potentially as small as millimeters, and the equation cannot distinguish between fractional areas of the entire fire which often are much smaller than the pixel itself. Thus, different methods have been tested and employed to overcome these limitations. The bi-spectral method, using two distinct channels (usually 4 and 11μm), can provide details about the fractional size and temperature of sub-pixel fire components (Dozier, 1981; Giglio & Kendall, 2001, Wooster et al., 2005), but is plagued by potential errors associated with channel misregistration and point spread function (PSF) differences between channels (Giglio & Kendall, 2001). Wooster et al. (2005) suggested that the bi-spectral method is effective, but primarily for high resolution sensors (<1km). The current method used aboard MODIS employs a single channel approach with fire and background components retrieved solely from the mid-infrared (4μm) channel (Justice et al, 2002). Kaufman et al. (1996, 1998) tested this single channel approach using the MODIS Airborne Simulator (MAS), model simulations of fire mixed-temperature pixels (to realistically mimic the nonhomogeneous behavior of biomass burning temperatures), and *in situ* measurements. Based on the simulated fires, Kaufman et al. (1998) revealed that an empirical relationship exists between instantaneous FRE (i.e. FRP) and pixel brightness temperature measured in the Moderate Resolution Imaging Spectroradiometer (MODIS) middle infrared channel (4 µm). The result was a semi-empirical relationship which forms the basis for the current FRP algorithm (equation 5) used aboard MODIS. The authors also demonstrated the correlation between rates of smoke emission and the observed rate of energy released from airborne observations with the MAS (Kaufman et al., 1996, 1998).

$$\text{FRP} \left[ \text{M} \text{W} \,\text{km} \,\text{:}^2 \right] = 4.34 \times 10^{19} \left( T^{\aleph\_{\text{MIR}}} - T^{\aleph\_{\text{bg}} \text{ MIR}} \right) \tag{5}$$

where FRP is the rate of radiative energy emitted per pixel (the MODIS 4µm channel has IFOV of 1km), 4.34x10-19 [MW km-2 Kelvin-8] is the constant derived from the Kaufman et al. (1998) simulations, *T*MIR [Kelvin] is the radiative brightness temperature of the fire component, *T*bg, MIR [Kelvin] is the neighboring nonfire background component, and MIR refers to middle infrared wavelength, typically 4μm.

Wooster et al. (2003) showed that FRP could also be derived using satellite-based middle infrared radiances and a simple power law to approximate Plank's law. The 'MIR radiance' method is applicable for temperatures covering the range of typical vegetation fires (600 – 1500 K). As with the 'MODIS' method, the MIR method relies on the difference between the fire pixel and background, but uses spectral radiance differences rather than brightness temperature. According to Wooster et al. (2005) the radiance methods allows perturbations, such as atmospheric effects and pixel area variation across the scan angles, to be accounted for after FRP has been derived.
