**5. Uncertainty**

Remote sensing science involves the inference of *in situ* physical characteristics based on electromagnetic energy received at the sensors. The sensor and inferences made have a degree of error, which propagates through processing and into any results produced.

Giglio & Kendall (2001) explained that while sensors such as AVHRR have effectively generated baseline fire products for fire distribution and basic qualitative information for parameterization of biomass burning, there are fundamental limitations that need to be addressed and could not be improved upon until recently. A limitation prior to sensors such as the Moderate Resolution Imaging Spectroradiometer (MODIS) and Spinning Enhanced Visible and Infrared Imager (SEVIRI) was that satellite sensors did not have dedicated fire channels or did not possess optimal sensor characteristics for fire detection (Kaufman et al., 1998). AVHRR's mid-infrared channel, for example, is subject to greater atmospheric perturbation and, due to sensor capabilities, increased frequency of saturation. Advancements of new generations of sensors for fire detection and monitoring have included refinements to specific wavelength selection in order to optimize spectral characterization (Giglio & Justice, 2003). Giglio & Kendall (2001) state that in order to reliably determine instantaneous fire temperature and area over a wide range of active fire sizes using Dozier's method, sensors with higher spatial resolution (~100m) and very high (~1000K) middle infrared band saturation would be necessary. However, replacement of the mid-IR channel aboard AVHRR (3.7μm) with a wavelength less sensitive to solar radiation (i.e. 3.9μm) would reduce the pixel saturation by half (Giglio & Justice, 2003; Kaufman et al., 1998). Defining higher pixel saturation temperatures and including wavelengths that can be used for false alarm detection on MODIS were improvements based on experiences with older systems (Justice et al., 2002).

Earlier work by Schroeder et al. (2008) showed that cloud obscuration in the Brazilian Amazon could lead to fire detection omission errors of roughly 11%. However, commission errors may occur as result of cloud shadows and semi-transparent clouds influencing surface thermal characteristics. Another source of commission errors may also result from the very efforts to correct for cloud obscuration by leading to an overestimate in the number of (assumed) detections. Schroeder et al. (2010) recently provided a thorough analysis of FRP, temperature estimates, and fire area estimates from moderate resolution sensors. Their results showed that location of fires within a pixel can be biased because of the sensor's point spread function leading to as much as a 75% underestimation in FRP. On the other hand, improper characterization of the ambient background surrounding fire pixel(s) can result in an overestimation of FRP up to 80%. This particular situation is mostly prevalent in areas of tropical deforestation.

The accuracy of the empirical formula for computing FRP was taken from the evaluation performed by Kaufman et al. (1998), who showed a potential error of 16% using 150 simulated mixed-energy fire pixels. Wooster et al. (2003) found a theoretical accuracy (RMSD) of 65 x 106 J over a range of 0 to 2000 x 106 J (or 6.5% for the average) using their MIR FRE approach.

MODIS omission rates of small active fires were also observed by Hawbaker et al*.* (2008). In their study, 73% of Aqua and 66% of Terra active fire detections were missed, primarily because of fast moving fire fronts, cloud cover, or spatial resolution. Hawbaker et al*.* (2008) was clear though that these small fires likely may have little impact in terms of total emissions as had been stated previously by Kaufman et al. (1998).

Remote sensing science involves the inference of *in situ* physical characteristics based on electromagnetic energy received at the sensors. The sensor and inferences made have a degree of error, which propagates through processing and into any results produced. Giglio & Kendall (2001) explained that while sensors such as AVHRR have effectively generated baseline fire products for fire distribution and basic qualitative information for parameterization of biomass burning, there are fundamental limitations that need to be addressed and could not be improved upon until recently. A limitation prior to sensors such as the Moderate Resolution Imaging Spectroradiometer (MODIS) and Spinning Enhanced Visible and Infrared Imager (SEVIRI) was that satellite sensors did not have dedicated fire channels or did not possess optimal sensor characteristics for fire detection (Kaufman et al., 1998). AVHRR's mid-infrared channel, for example, is subject to greater atmospheric perturbation and, due to sensor capabilities, increased frequency of saturation. Advancements of new generations of sensors for fire detection and monitoring have included refinements to specific wavelength selection in order to optimize spectral characterization (Giglio & Justice, 2003). Giglio & Kendall (2001) state that in order to reliably determine instantaneous fire temperature and area over a wide range of active fire sizes using Dozier's method, sensors with higher spatial resolution (~100m) and very high (~1000K) middle infrared band saturation would be necessary. However, replacement of the mid-IR channel aboard AVHRR (3.7μm) with a wavelength less sensitive to solar radiation (i.e. 3.9μm) would reduce the pixel saturation by half (Giglio & Justice, 2003; Kaufman et al., 1998). Defining higher pixel saturation temperatures and including wavelengths that can be used for false alarm detection on MODIS were improvements based

Earlier work by Schroeder et al. (2008) showed that cloud obscuration in the Brazilian Amazon could lead to fire detection omission errors of roughly 11%. However, commission errors may occur as result of cloud shadows and semi-transparent clouds influencing surface thermal characteristics. Another source of commission errors may also result from the very efforts to correct for cloud obscuration by leading to an overestimate in the number of (assumed) detections. Schroeder et al. (2010) recently provided a thorough analysis of FRP, temperature estimates, and fire area estimates from moderate resolution sensors. Their results showed that location of fires within a pixel can be biased because of the sensor's point spread function leading to as much as a 75% underestimation in FRP. On the other hand, improper characterization of the ambient background surrounding fire pixel(s) can result in an overestimation of FRP up to 80%. This particular situation is mostly prevalent

The accuracy of the empirical formula for computing FRP was taken from the evaluation performed by Kaufman et al. (1998), who showed a potential error of 16% using 150 simulated mixed-energy fire pixels. Wooster et al. (2003) found a theoretical accuracy (RMSD) of 65 x 106 J over a range of 0 to 2000 x 106 J (or 6.5% for the average) using their

MODIS omission rates of small active fires were also observed by Hawbaker et al*.* (2008). In their study, 73% of Aqua and 66% of Terra active fire detections were missed, primarily because of fast moving fire fronts, cloud cover, or spatial resolution. Hawbaker et al*.* (2008) was clear though that these small fires likely may have little impact in terms of total

emissions as had been stated previously by Kaufman et al. (1998).

on experiences with older systems (Justice et al., 2002).

in areas of tropical deforestation.

MIR FRE approach.

**5. Uncertainty** 

Freeborn et al. (2011) highlighted an issue with the MODIS Collection 5 FRP product. In the C5 FRP the calculation of the instantaneous energy (MW) derived from the brightness temperature includes a multiplication by the pixel area. Although this is fundamentally correct, since energy is measured per unit time and space, the adjustment leads to an overestimate with increasing scan angle because the pixel area grows as the scan moves off nadir. Interestingly, the opposite effect occurs when examining fire pixel counts (i.e. greater number of detections near nadir and decreasing detections with scan angle).

With regards to the application of FRE-based biomass consumption estimates published by Ellicott et al. (2009) and Roberts & Wooster (2008), there is some degree of uncertainty. Although the assumption that a single combustion factor is applicable for all fuel types and conditions (i.e. moisture content) will incur some bias, in general, heat yield does not vary much between fuels (Stott, 2000) and therefore until more research demonstrates otherwise, the two cited FRE-based combustion factors (Freeborn et al., 2008; Wooster, 2005) seem realistic.

Atmospheric attenuation is another component generally unaccounted for. In simulations conducted by Ellicott (unpublished), the MODIS FRP may be underestimated by as much as 20% (Figure 4). Similarly, Roberts & Wooster (2008) applied a constant correction factor (0.89) to SEVIRI FRP to account for atmospheric transmission loss.

Fig. 4. Comparison of simulated surface and TOA FRP. Radiances were simulated from randomly generated fire pixel temperature and fractional area components (fire, smoldering, and background). MODIS Aqua profiles were used to provide realistic atmospheric parameters used in the radiative transfer modeling. The 1:1 (dashed) line is plotted for reference.

The Science and Application of Satellite Based Fire Radiative Energy 189

burning emission estimates. Page et al. (2002) estimated 0.19-0.23 Gt of carbon 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. Cloud cover has already been revealed by Schroeder et al. (2008) to limit fire detection capabilities for Brazilian fires. The spatial resolution of MODIS is another limitation to detecting fires in peatlands (and thus FRP estimation) as shown by Siegert et al. (2004). Developing a connection between field estimates of surface and sub-surface organic burning, burned area, and FRP would allow for

I thank Dr. Louis Giglio for his insight to the concepts, science, and application of fire radiative energy and technical assistance and in developing the FRE parametreization method, generously offering data, and providing critical assessment of our ideas and approaches. I thank Dr. Gareth Roberts for helping with SEVIRI data and giving feedback on my research and writing. I thank Dr. Wilfrid Schroeder for his helpful technical insight and fruitful discussions of the limitations and uncertainty in active fire detections and FRP. Finally, I thank Dr. Guido van der Werf for his support through offering insight on the GFED processes and data, providing advice on our research, and feedback on my analysis. Dr. Ellicott would also like to thank NASA's Earth and Space Science Fellowship Program for recognizing the potential benefits of his research and providing financial support during

Aragao, L. E. O. C. & Shimabukuro, Y. E. (2010). The Incidence of Fire in Amazonian

Boles, S.H. & Verbyla, D.L. (2000). Comparison of three AVHRR-based fire detection algorithms for interior Alaska. *Remote Sensing Environment*, v72, p1-16. Boschetti, L., Eva, H. D., Brivio, P. A., & Gregoire, J. M. (2004). Lessons to be learned from

Boschetti, L. & Roy, D. P. (2009). Strategies for the fusion of satellite fire radiative power

Brady, N.C. & Weil, R. R. (1999). In: *The Nature and Properties of Soils*: 12th Edition." (N.C.

Cary, G. J., & Banks, J. C. G. (2000). Fire regime sensitivity to global climate change: An

Chapman, S.J. & Thurlow, M. (1998). Peat respiration at low temperatures, *Soil Biology &* 

Cicerone, R. J. (1994). Fires, atmospheric chemistry, and the ozone layer, *Science*, *263*, 1243-1244. Crutzen, P. J., L. E. Heidt, J. P. Krasnec, W. H. Pollock, and W. Seiler (1979). Biomass burning as a source of atmospheric gases CO, H2, N2O, NO, CH3Cl, and COS. *Nature*, *282*, 253-256. Crutzen, P. J. & Andreae, M. O. (1990). Biomass Burning in the Tropics: Impact on Atmospheric Chemistry and Biogeochemical Cycles, *Science*, *250* (4988), 1669-1678.

the comparison of three satellite-derived biomass burning products, *Geophysical* 

with burned area data for fire radiative energy derivation, *J Geophys Res-Atmos*, *114*.

Australian perspective, in *Biomass burning and its inter-relationships with the climage system*, edited by J. L. Innes, M. Beniston, and M. M. Verstraete, pp. 233-245,

Forests with Implications for REDD, *Science*, *328*(5983), 1275-1278.

*Research Letters*, *31*(21), L21501, doi:10.1029/2004gl021229.

Brady and R.R. Weil, Eds.) Prentice-Hall, London.

Dordrecht, Kluwer Academic Publishers.

*Biochemistry*, *30*, 1013-1021.

parameterization of this component of fire radiative energy.

**7. Acknowledgements** 

his doctoral degree endeavor.

**8. References** 

Finally, an error budget provided by Vermote et al. (2009) suggested that the FREparameterization approach developed by Ellicott et al. (2009) approaches 20% based on comparisons with the SEVIRI sensor.
