XXIX-B8, 2012
oth considering the
3A product and the
uses of global BA
different fire-related
ng experts, natural
estionnaire with their
spatial and temporal
cations (PSD) were
nitations of the input
its. As a result of this
re cci project would
el, which will be the
VEGETATION and
tion, and another one
following the most
M) size. The BA
lution, with temporal
t and 15 days for the
ts will be properly
confidence level and
n of BA, confidence
fire size distribution
rid product. The BA
:CDF formats, using
luct will be: 85 96 of
sion and commission
ation accuracy, + 3.5
1 15% of temporal
lucts of the fire cci
calibrated radiances
MERIS. To derive
seometrical matching
Landsat GLS2000 as
the three sensors,
ased on the ATCOR
en carried out with
TION were already
er, cloud snow and
oped to reduce the
5, since these covers
s to BA. Particularly
vater bodies may be
nces of water may be
implemented and
tes. Those sites were
include different BA
ire occurrence areas.
| by fires, as well as
ic for burned area
"mporal series (1995-
le in each period.
ce target sensors and
conditions at global
narily aim at the ten
jy in the processing
hms currently tested
ed on multitemporal
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B8, 2012
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia
[1 study sites selection
GFED 3
gC/m2 year
0.0001 - 11,8058
11,8060 - 37 3029
37,3000 . 75,8788
HS 75.0700. 120.1970
TEE 130.1971 - 2500,0000
Figure 2: Study sites used for the fire cci BA project. The estimated emissions based on the GFEDv3 (van der Werf et al., 2010)
database are used as the background
change detection, contextual-regional analysis and fire
seasonality.
A Round-Robin exercise was conducted between October, 2011
and January 2012 to test the most relevant existing BA
algorithms, and rank their performance with the different input
sensors. The goal of this exercise was to select the best
performing algorithm for global production of burned area
maps. The exercise was open to public participation. The same
metrics used for validation were computed for each BA output.
Once the best performing algorithm is selected, a merging
process will be developed to create a synthetic BA product from
the three BA pixel products (A-ATSR, VGT and MERIS).
Finally, the complete processing chain will be applied at global
scale for five selected years (1999, 2000, 2003, 2005 and 2008),
to demonstrate the operational conditions of both the pre-
processing and BA algorithms.
3.5 Validation
Validation of the BA product will be performed by comparing
BA outputs with reference fire perimeters generated from
Landsat-TM/ETM+ multitemporal images. A standard protocol
based on the CEOS LPV recommendations was generated and
agreed between the internal validation teams to extract fire
perimeters from Landsat data, based on a semi-automatic
algorithm (Bastarrika et al., 2011). The validation exercise will
aim to measure both spatial and temporal accuracy and
precision. The spatial assessment will be based on a sample of
110 multitemporal Landsat pairs acquired in 2008, while the
temporal stability will be measured from a temporal series of
one Landsat scene for each of the ten study sites previously
commented.
Validation metrics will be based on computation of the user and
producer accuracy (Congalton and Green, 1999) for each
Landsat scene, as well as the consistency and temporal stability
of those accuracy measurements.
3.6 Use of BA data in models
BA information generated by the fire cci project will be
compared with other global BA products currently available
(GFED3 and MCD45), to check common trends and potential
15
problems. Modellers within the fire_cci consortium will test the
BA information in atmospheric and carbon cycle models to
analyze its advantages and limitations.
4. RESULTS
4.1 Pre-processing
The fire_cci is a three year project that is currently running. The
current development is focusing on the production of corrected
reflectances from the three input sensors (A-ATSR, VGT and
MERIS) and the generation of the BA and merging algorithms.
Currently, the full temporal series of corrected reflectances
(along with the water, snow and cloud masks) for all ten study
sites and sensors is available for three years (2005, 2006 and
2008), and the full time series (1995-2009) for the Australia site
is being processed. This pre-processing of input data is quite
critical to assure a coherent time series for BA production, but it
is also very demanding in terms of computing time, including
the tailoring of the global orbits to the coordinates of the study
sites. The process for all sites is expected to finish in March,
2012.
4.2 Validation datasets
The reference fire perimeters derived from Landsat-TM/ETM+
images are ready for the temporal validation, which includes
sets of Landsat pairs for 10-15 years for each study site (fig. 3).
This validation dataset will be used to estimate the accuracy,
consistency and temporal stability of the BA product.
Another validation dataset is being developed, which will aim to
perform the spatial validation for the global product. A stratified
random sampling has been performed to select representative
Landsat scenes for different biomes according to the estimated
fire occurrence by the MODIS BA product in 2008.
5. CONCLUSIONS
The fire cci project as part of the ESA effort to generate ECV
to meet the needs of the global climate community, is trying to
generate accurate, consistent and stable time series of BA
information based on European sensors. The BA outputs are
