fire_cci D4.1.3 Product Intercomparison Report (PIR) - Fire-CCI

fire_cci
D4.1.3 Product Intercomparison
Report (PIR)
Project Name
ESA CCI ECV Fire Disturbance (fire_cci)
Contract N°
4000101779/10/I-NB
Project Manager
Arnd Berns-Silva
Last Change Date
22/10/2014
Version
1.4
State
Final
Author
Rubén Ramo Sánchez, Angelika Heil, Emilio Chuvieco
Document Ref:
Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Document Type:
Public
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
Project Partners
Prime Contractor/
Scientific Lead
-
UAH - University of Alcalá de Henares (Spain)
Project Management
-
GAF AG (Germany)
-
ISA - Instituto Superior de Agronomia (Portugal)
UL - University of Leicester (United Kingdom)
DLR - German Aerospace Centre (Germany)
System Engineering Partners
Earth Observation Partners
Climate Modelling Partners
GMV - Aerospace & Defence (Spain)
DLR - German Aerospace Centre (Germany)
-
IRD-CNRS - L’Institut de Recherche pour le Développement - Centre National de la
Recherche Scientifique (France)
JÜLICH - Forschungszentrum Jülich GmbH (Germany)
-
LSCE - Laboratoire des Sciences du Climat et l’Environnement (France)
Distribution
Affiliation
ESA-ECSAT
Project Team
Name
Address
Stephen Plummer (ESA – ECSAT)
Emilio Chuvieco, (UAH)
Patricia Oliva (UAH)
Itziar Alonso (UAH)
Stijn Hantson (UAH)
Marc Padilla Parellada (UAH)
Arnd Berns-Silva(GAF)
Christopher Sandow (GAF)
Stefan Saradeth (GAF)
Federico González Alonso (INIA)
Jose Miguel Pereira (ISA)
Duarte Oom (ISA)
Gerardo López Saldaña
Kevin Tansey (UL)
Luis Gutiérrez (GMV)
Ignacio García (GMV)
Andreas Müller (DLR)
Martin Bachmann (DLR)
Martin Habermeyer (DLR)
Kurt Guenther (DLR)
Thomas Krauß (DLR)
Eric Borg (DLR)
Martin Schultz (JÜLICH)
Angelika Heil (JÜLICH)
Florent Mouillot (IRD)
Julien Ruffault (IRD)
Philippe Ciais (LSCE)
Patricia Cadule (LSCE)
Chao Yue (LSCE)
[email protected]
[email protected]
[email protected]
[email protected]
[email protected]
[email protected]
[email protected]
[email protected]
[email protected]
[email protected]
[email protected]
[email protected]
[email protected]
[email protected]
[email protected]
[email protected]
[email protected]
[email protected]
[email protected]
[email protected]
[email protected]
[email protected]
[email protected]
[email protected]
[email protected]
[email protected]
[email protected]
[email protected]
[email protected]
D4.1.3 Product Intercomparison Report
Copies
electronic copy
electronic copy
Page II
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
Document Status Sheet
Prepared
Reviewed
Affiliation/Function
UAH
UAH, LSCE, IRD, JÜLICH
Authorized
Accepted
UAH/ Prime Contractor
ESA/ Project Manager
Name
Ruben Ramo
Emilio Chuvieco, Chao Yue,
Florent Mouillot, Angelika
Heil
Emilio Chuvieco
Stephen Plummer
Date
04/08/2014
04/08/2014
08/10/2014
Name
Emilio Chuvieco
Stephen Plummer
Signature
15/10/2014
Signatures
Signature of authorisation and overall approval
Signature of acceptance by ESA
Date
Document Status Sheet
Issue
1.1
1.2
1.4
Date
05/08/2014
14/10/2014
20/10/2014
Details
First Document Issue
Addressing comments according to CCI-FIRE-EOPS-MM-14-00xg.pdf
Updating document
Document Change Record
#
1.2
Date
14/10/2014
Request
LSCE,
JÜLICH;
IRD
Location
Abbreviation
Section 2
Section 3
Section 5.4
Section
5.5
Section
5.6
Section
5.9
Section 5.10.1
1.4
20/10/2014
UAH,
JÜLICH
D4.1.3 Product Intercomparison Report
Whole
document
Section 3
Section 4.3
Section 4.4
Section 5.4
Section 5.6
Whole
document
Details
Updated
Updated
Updated and Table 1 amended;
Figure 10 annotation updated;
Updated; Figure 15 annotation updated; Table
7 annotation updated; Including paragraph on
NFG
Updated and rephrased; Figure 16 and 17
annotation updated;
Introducing Matching of MERIS burned area
with different hot spot products
Figure 24, 27, 28 and 29 annotation updated;
Table 11 and 12 annotation updated;
Updated references
Table 1 completed; Rephrased;
Justification why Australia was choosen
Rephrased
Amplifying Figure 11, 12, 13
Rephrased and updated
Typo correction
Page III
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
Table of Contents
1
2
3
4
5
6
7
Executive Summary ...................................................................................................................... 1
Introduction ................................................................................................................................... 1
Description of Products................................................................................................................. 2
Description of Methods ................................................................................................................. 4
4.1
Data Pre-processing ................................................................................................................. 4
4.2
Burned area calculations.......................................................................................................... 7
4.3
Time and Space domains ......................................................................................................... 7
4.4
Source of National Statistics ................................................................................................... 7
Results............................................................................................................................................. 9
5.1
Comparisons between the pixel products ................................................................................ 9
5.2
Global estimation of burned areas (grid products) .................................................................. 9
5.3
Regional estimation of burned areas (grid products)............................................................. 10
5.4
Spatial differences between GFED and Fire_cci BA grid products ...................................... 13
5.5
Percentile comparisons .......................................................................................................... 17
5.6
Comparison of the size distribution of globally gridded burned area ................................... 21
5.7
Fire seasonality ...................................................................................................................... 23
5.8
Statistical trends .................................................................................................................... 24
5.8.1
Fire_cci products ............................................................................................................................ 24
5.8.2
MCD45........................................................................................................................................... 28
5.9
Matching of MERIS_cci burned area with different hotspot products................................. 30
5.10
Comparison of the spatial patterns of mean annual burned area for selected regions ........... 34
5.10.1
Australia ......................................................................................................................................... 34
5.10.2
Europe ............................................................................................................................................ 42
Conclusions .................................................................................................................................. 47
References .................................................................................................................................... 48
D4.1.3 Product Intercomparison Report
Page IV
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
List of Figures
Figure 1: Methodology applied to the intercomparison of BA ............................................................................... 4
Figure 2: Example of the binarization process: MCD45 of June (left) and the original product (right) ................. 5
Figure 3: Annual composite of the MCD45 product (left) and single month binarized product (Right) ................ 5
Figure 4: Representation of the binarized MCD45 mosaic. Burned pixels in red ................................................... 6
Figure 5: 0.5 degree grid over Australia.................................................................................................................. 6
Figure 6: Map of the 14 regions used for the intercomparison analysis (after (Giglio et al. 2006). Description of
the areas is included Table 3. ................................................................................................................. 7
Figure 7: Burned area detected by MERIS, MCD45 and MCD64 in 2008 ............................................................. 9
Figure 8: Estimation of burned area by different BA products ............................................................................. 10
Figure 9: Average BA for each region (2006-2008). For Geoland-2 only 2008 was computed ........................... 11
Figure 10: Mean annual burned area (2006 - 2008) expressed as fraction burned per 0.5-degree grid cell in the
different burned area products. The grid cell area at the Equator is 3090 km2 and 2670 km2 at 60° N
or S. ...................................................................................................................................................... 13
Figure 11: Spatial difference between GFED and MERIS.................................................................................... 14
Figure 12: Spatial difference between GFED and VGT_cci ................................................................................. 15
Figure 13: Spatial difference between GFED and MERGE .................................................................................. 16
Figure 14: Spatial difference between MERIS and VGT ...................................................................................... 18
Figure 15: Field percentiles of mean annual i) burned area (MERIS, GFEDv4, VGT, and MERGED; ii) FRP
(GFAS) and iii) hotspot (HS) counts (MODIS, WFA and TRMM). The total number of valid fire grid
cells and the percentile statistics of each product are given in Table 6. Fire signals in the Southern
Pacific in the WFA-HS product are created by an artefact in the original August 2008 data .............. 19
Figure 16: Cumulative distribution function (CDF) of annual burned area (BA) per fire grid cell (left panel) and
boxplot of monthly BA per fire grid cell (right panel) over 2006 to 2008 in MERIS_cci, GFED4,
GFED3, VGT_cci and MERGED_cci. The topmost panels show the distribution of burned area size in
all fire grid cells, which further sub-divided into intersecting (middle) and non-intersecting (bottom)
fire grid cells. For the intersecting grid cells, burned area is greater than zero in all the five products.
This subset excludes a substantial fraction of small and large fires which fall outside the detection
limit of any of the four products. It thereby also excludes a large fraction of false alarms. The
remaining fire grid cells are categorized as non-intersecting. The number of fire grid cells (N)
contained in the entire population and in the subsets is displayed as well. .......................................... 22
Figure 17: Cumulative distribution function (CDF) (left panel) and boxplot (right panel) of monthly burned area
(BA) per fire grid cell (left panel) over 2006 to 2008 in MERIS_cci and GFED4 (top) and
MERIS_cci and VGT_cci (bottom), sub-divided into intersecting (labelled with '+') and nonintersecting (labelled with '-') fire grid cells. The number of fire grid cells (N) contained in the subsets
is displayed as well. .............................................................................................................................. 23
Figure 18: Monthly BA estimations for the GFED and fire_cci products ............................................................ 24
Figure 19: Correlations between GFED and MERIS_cci for the three study years ............................................. 25
Figure 20: Correlations between GFED and VGT_cci for the three study years .................................................. 26
Figure 21: Correlations between GFED and MERGED_cci for the three study years ........................................ 27
Figure 22: Correlations between GFED and MCD45 in the three study years ..................................................... 29
Figure 23: Mean annual number of active fire pixels per grid cells in the WFA-HS, TRMM-HS and the MODISHS product (2006 - 2008). Shown are a) all fire grid cells (“all”) and b) the fire grid cells with no colocated fire signal in the MERIS_cci product (“no MERIS”). Intersecting with MERIS_cci was
performed with time-integrated data. ................................................................................................... 33
D4.1.3 Product Intercomparison Report
Page V
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
Figure 24: Same as Figure 15 but for Australia only. The field percentile refers to the global field percentile. The
grid cell values corresponding to the different percentiles are listed in Table 6. ................................. 35
Figure 25: Left: map of the different states of Australia (NSW: New South Wales; NT: Northern Territory;
QLD: Queensland; SA: South Australia; TAS: Tasmania; VIC: Victoria; WA: Western Australia);
Middle: MCD12 dominant land cover per 0.5-degree grid cell of the year 2005 using the UMD legend
(1=ENF: evergreen needleleaf forest; 2=EBF: evergreen broadleaf forest; 3=DNF: deciduous
needleleaf forest; 4=DBF: deciduous broadleaf forest; 5=MF: mixed forest; 6=CSH: closed
shrublands; 7=OSH: open shrublands; 8=WSA: woody savannas; 9=SAV: savannas; 10=GRA:
grasslands; 12=CRO: croplands; grey=OTH: others (merging urban and built-up with barren or
sparsely vegetated); Right: Forest cover of Australia according to the Australia’s State of the Forests
Report 2013 (MPIGfA/NFISC, 2013). ................................................................................................ 36
Figure 26: Latitudinal pattern of fractional fire grid coverage over land (top), mean annual burned area (middle),
and monthly burned area per fire grid (bottom) in the Australian domain (40 ~ 10° S; 110 ~ 155° E),
see also Figure 24. To the right, the coefficient of determination of the zonal correlation of the product
pairs is given. ....................................................................................................................................... 36
Figure 27: Monthly total area burned in Australia during January 2006 to December 2008 in the MERIS_cci,
GFED, VGT_cci, and MERGED_cci product. The coefficient of determination of the temporal
correlation is given to the right. ........................................................................................................... 37
Figure 28: Fire affected area (in red) during the Great Divide Fire Complex 2006/07 in Victoria, Australia,
mapped with high-resolution Landsat and SPOT satellite imagery (VdoSE, 2008). The map spans
from approximately 145.2° to 149.5° E and 38.1° to 36.1° S. The official burned area estimates range
from 11,800 to 12,600 km² (Smith, 2007; VdoSE, 2008) .................................................................... 39
Figure 29: Fire affected area in km2 per grid summed over the period August 2006 to March 2007 in Victoria,
Australia, mapped by MERIS_cci, GFED, VGT_cci and MERGED_cci. The maps span from 145.5°
to 149.5° and 38.5° to 36° S. Burned area estimates for this domain calculated from the fire satellite
products over this period are: MERIS_cci : 7,460 km² , GFED4: 8,937 km²; GFED3: 8,100 km²;
VGT_cci: 2,068 km², and MERGED_cci: 9,356 km². The bottom left figure shows the topographic
height in metres at 0.1 degree resolution (derived from the GTOPO30 product) ................................ 40
Figure 30: Monthly burned area per 0.5-degree grid cell in the different burned area products over the fire year
July 2006 to June 2007 across Australia (top to bottom). Monthly total GFAS FRE, AATSR and
TRMM hotspot number is shown as well. Please note that spatiotemporal pattern of NOAA burned
area
is
similar
to
MERIS_cci
(see
also
http://www.firewatch.landgate.wa.gov.au/landgate_firewatch_public.asp). ........................................ 41
Figure 31: Monthly total area burned in the Australian states Victoria (VICT) and New South Wales (NSW)
during January 2006 to December 2008 in the MERIS_cci, GFED, and MERGED_cci product. The
coefficient of determination of the temporal correlation is given to the right ...................................... 42
Figure 32: Same as Figure 15 but for Europe only. The field percentile refers to the global field percentile. The
grid cell values corresponding to the different percentiles are listed in Figure 15: Field percentiles of
mean annual i) burned area (MERIS, GFEDv4, VGT, and MERGED; ii) FRP (GFAS) and iii) hotspot
(HS) counts (MODIS, WFA and TRMM). The total number of valid fire grid cells and the percentile
statistics of each product are given in Table 6. Fire signals in the Southern Pacific in the WFA-HS
product are created by an artefact in the original August 2008 data .................................................... 44
Figure 33: Latitudinal pattern of fractional fire grid coverage over land (top), and mean annual burned area
(2006-2008) (middle), and monthly burned area per fire grid (bottom) in the European domain (30 ~
70° N; 10° W~70° E). To the right, the coefficient of determination of the correlation of the product
pairs is given. ....................................................................................................................................... 45
Figure 34: Time series of monthly burned area in the European domain (30~70° N; 10° W~70° E) from January
2006 to December 2008. The coefficient of determination (R²) of the pairwise correlation of the time
series is given to the right. .................................................................................................................... 45
Figure 35: Burned area in Eastern Europe in April and May 2006 in MERIS_cci, GFED3, GFED4 and
MERGED_cci, complemented by monthly GFAS FRE. MODIS fire detections for April 21 to May 5,
2006, colour-coded by vegetation type is shown at the bottom. .......................................................... 46
D4.1.3 Product Intercomparison Report
Page VI
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
List of Tables
Table 1: Specifications of products included into the comparison .......................................................................... 2
Table 2: Burned area statistics on national level ..................................................................................................... 8
Table 3: BA estimations from different geographical regions (2006) ................................................................... 11
Table 4: BA estimations from different geographical regions (2007) ................................................................... 12
Table 5: BA estimations from different geographical regions (2008) ................................................................... 12
Table 6: Global number of annual fire grid cells (NFG) and absolute values corresponding to the field percentile
ranks in each product. .......................................................................................................................... 20
Table 7: Correlation between MERIS_cci and GFED BA for the different regions. ........................................... 25
Table 8: Results of the VGT_cci correlation analysis along the different basin regions ...................................... 26
Table 9: Results of the MERGED_cci correlation analysis along the different regions. ..................................... 28
Table 10: Results of the MCD45 correlation analysis along the different basin regions. ..................................... 29
Table 11: Total number of active fire pixels per year in the TRMM domain in the WFA-HS, TRMM-HS and
MODIS-HS products, differentiated by land cover class. The contribution of hotspots missed by
MERIS_cci in each land cover class is given as well as the partitioning of the total number of hotspot
missed across the individual land cover classes. Intersecting with MERIS_cci was performed with
time-integrated data (period 2006-2008). The land cover classes are defined in Figure 24. “Vegetated”
includes all land cover classes except predominantly urban and/or barren and sparsely vegetated grid
cells. ..................................................................................................................................................... 31
Table 12: Total number of time-integrated fire grid cells (NFG) fire pixels in the TRMM domain in the WFAHS, TRMM-HS and MODIS-HS products and percentage missed by MERIS_cci. The average
number of hotspots per fire grid, integrated over 2006 to 2008, is shown for the missed and not-missed
subset. Intersecting with MERIS_cci was performed with time-integrated data (period 2006-2008). 32
Table 13: Forest burned area by state and fire year in Australia compiled by the Australian state agencies
(MPIGfA/NFISC, 2013). Burned area in states marked by a star (*) is mapped with high-resolution
Landsat and SPOT satellite imagery, combined with aerial and ground-based survey. For the other
states, burned area is mapped with MODIS burned area imagery (MCD45), only, which are cut by
forest cover. .......................................................................................................................................... 37
Table 14: Total burned area per fire year (July-June) and per Australian state calculated from MERIS_cci,
GFED4, GFED3, VGT_cci and MERGED_cci. The right column shows the forested burned area
estimated by the Australian authorities. ............................................................................................... 38
D4.1.3 Product Intercomparison Report
Page VII
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
List of Abbreviations
BA
CCI
DoY
ECV
FRP
GDAL
GEOLAND2
GFAS
GFED
HS
MCD45
MCD64
MERIS
MODIS
PIR
TRMM
VGT
WFA
Burned Area
Climate Change Initiative
Day of Year
Essential Climate Variable
Fire Radiative Power
Geospatial Data Abstraction Library
Pre-operational EU-funded project
Global Fire Assimilation System
Global Fire Emissions Database
Hotspot (Active fire pixel count)
Modis Burned Area Product
Modis Burned Area Product
Medium Resolution Imaging Spectrometer, on board of ENVISAT
MODerate Resolution Imaging Spectrometer (on board of TERRA and AQUA)
Product Intercomparison Report
Tropical Rainfall Measurement Mission
CNES Earth’s observation sensor onboard SPOT-4
World Fire Atlas (ESA (A)ATSR)
D4.1.3 Product Intercomparison Report
Page VIII
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
1 Executive Summary
This document presents the comparison of fire_cci global BA products for a three year period (2006 to
2008) with existing BA products (MCD45, GFED, Geoland), plus some products based on active fire
detections. Estimated burned area (BA) of these products are compared using different graphical and
statistical methods, including the analysis of spatial (global and GFED regions) and temporal
(monthly) trends. Most analysis were done for the grid product (0.5 degree resolution), as it provided a
more efficient framework for comparing with existing BA products (particularly the Global Fire
Emission Database, GFED), widely used by the climate modellers.
We found that the MERIS burned area product (3.5 to 3.7 million km²) is in close agreement with the
GFED data in terms of total annual area burned, the spatial and temporal distribution of fires, and
regional trends compared with national BA information. Therefore, we can conclude that a great
improvement has been achieved by the MERIS_cci BA product compared with formerly released BA
products by ESA (GLOBCARBON, L3JRC). However, the VGT_cci product is subject to deficiencies
of the omission of small and large actual fires and commission errors due to the confusion of spectral
changes related to agricultural activities (soil preparation/harvesting) with burned area. The total BA
was found much lower than other products (2.23 - 2.37 million km²), but it was found to overestimate
in Boreal regions. For this reason, the merging of VGT and MERIS (MERGED_cci ) provided an
overall higher estimation than other products (4.86 to 5.01 million km²). Considering previous
validation efforts and the modelling exercises of existing BA products, the MERIS-cci product is
recommended for public distribution, as it provides the most consistent and coherent spatial and
temporal trends of the fire_cci output products.
2 Introduction
The growing interest of better understanding fire impacts on vegetation and atmospheric emissions
makes the estimation of global burned area a very relevant task. Since several global BA products
have been produced in the last few years, the climate users need to assess which provide a better
estimation of BA values and trends, and therefore the intercomparison of products becomes a critical
objective to make these BA products more useful for the international community.
This report compares spatial and temporal trends of BA information derived from the fire_cci project
for three years (2006 to 2008), with existing global BA products that are available for the same time
series. We selected GEOLAND-2, MCD45A1, MCD64A1 plus the three products derived from the
fire_cci project: MERIS_cci, VGT_cci, and MERGED_cci. We selected GEOLAND-2 as
representative of BA products based on European sensors, since the precursors (L3JRC and
GlobCarbon) were not available for 2008, they key year for global validation (Padilla and Chuvieco
2014).
In addition to BA products, we have also considered other fire products based on thermal anomalies,
such as hotspots (HS) from WFA (ESA AATSR), TRMM, and MODIS, and fire radiative power
(FRP) from GFAS. While fire_cci BA is not directly comparable to FRP or hotspots, the comparison
can point to potential commission/omission errors in the burned area product. For example, the
frequent absence of positive WFA and TRMM HS signals in positive fire_cci BA grid cells may imply
potential commission errors. And reversely, extreme values in the GFAS FRP to ESA BA ratio can
point to potential omission errors.
The purpose of the Product Intercomparison Report (PIR) is to determine the degree of agreement
between each BA product, and define where and when disagreements occur at global scale and for
different regions.
D4.1.3 Product Intercomparison Report
Page 1
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
3 Description of Products
Table 1 includes the technical description of each BA product compared in this document. The
VGT_cci product is based on SPOT-VEGETATION imagery. The algorithm relies on a time series
analysis of the near infrared band, applying robust filtering techniques, fire seasonality masks and
contextual algorithms (Pereira et al. 2014). The MERIS_cci product is based on a two-phase BA
algorithm: seed identification and region growing. The algorithm analyses changes in spectral
reflectances together with thermal anomalies, extracted from the MODIS sensor (Alonso and
Chuvieco 2014). The MERGED_cci product was designed using a combination of the single sensor
products (VGT and MERIS). Both sensor outputs are combined, without previous screening, which
implies that all pixels detected as burned in both sensors are included (Tansey et al. 2013).
Table 1: Specifications of products included into the comparison
Product
VGT _cci
Spatial resolution
1000 m / 0.5 degree
Period
2006 - 08
Comments
BA Pixel / Grid products
2006 - 08
BA Pixel / Grid products
2006 - 08
BA Pixel / Grid products
2000 - present
2000-present
1996 - present
1995 - present
BA Pixel
BA Pixel / GFED base
BA Grid
BA Grid
1000 m
0.1/0.5
1000 m
Temporal resolution
Date of Detection / 15day
Date of Detection / 15day
Date of Detection / 15day
Date of Detection
Date of Detection
Monthly
Monthly (>1995)
Daily (> 2000)
Date of Detection
Daily
Daily
MERIS_cci
300 m / 0.5 degree
MERGED_cci
300 m / 0.5 degree
MCD45A1
MCD64
GFEDv3
GFEDv4
500 m
500 m
0.5 degrees
0.25 degrees
GEOLAND2
GFAS-FRP
WFA-HS (ESA
AATSR)
TRMM-HS
MODIS-HS
2008
2001 - present
1995 - 2011
BA Pixel
Fire radiative power
Hotspots
2000 m
1000 m
Daily
Daily
1997 - Present
2001 - present
Hotspots
Hotspot
We compared the fire_cci products with the MODIS and other European BA products. The
MCD45A1 product is based on bi-directional reflectance (BRDF) model-based change detection
approach (Roy et al. 2005; Roy et al. 2002; Roy et al. 2006) to map burned areas. The product detects
the date of burning focusing on the changes in the daily MODIS reflectance product, avoiding the fires
that occurred in a previous season or years. The alternative MODIS BA product, named MCD64 is the
basis of the GFED. The product was generated using a hybrid algorithm that takes advantage of both
the thermal anomalies and the post-fire reflectance changes (Giglio et al. 2009).
The Global Fire Emissions Database (GFED) provides globally gridded timeseries of burned area and
fire emissions. It was designed for use in global atmospheric and biogeochemical models. Version 3
provides global monthly burned area aggregated to 0.5º spatial resolution (Giglio et al. 2010). BA
estimates from 2001 onwards, i.e. since the MODIS era, are produced from 500-m MODIS burned
area maps, complemented by indirect mapping. In the latter case, burned area is produced by
calibrating MODIS active fire data with 500-m MODIS burned area maps via geographically weighted
regression analysis. BA estimates for the pre-MODIS era are indirectly estimated from active fire data
from WFA ((A)ATSR) and TRMM (VIRS). For this purpose, fire counts are calibrated to MODIS
burned area using local and regional regression statistics. The calibration factors are then used to
convert WFA and TRMM fire counts into burned area (Giglio et al. 2010; Giglio et al. 2013). The
latest version, version 4, includes daily burned area as well, and the spatial resolution has improved to
0.25º (Giglio et al. 2013). We selected both versions, but most analyses have been done with v3 as the
total accounts are the same, and v3 includes the same spatial resolution as the gridded fire_cci
products. GFED is currently the most commonly used dataset in climate-chemistry modelling studies.
GFEDv3 and GFEDv4 exhibit only relatively small differences in terms of burned area totals or
spatial and temporal patterns. Recently, Randerson et al. (2012) showed that the current publically
available global MODIS burned area products have a tendency to miss the smaller fires. The inclusion
D4.1.3 Product Intercomparison Report
Page 2
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
of these omitted small fires would increase the MODIS global burned area estimates by 35
percentages.
The GEOLAND2 BA product is based on the algorithm developed within the L3JRC project (Tansey
et al. 2008). The V1 version of the GEOLAND2 BA product has been improved removing the need to
use land cover maps, making possible processing the product with a near-real-time 10 day product. It
uses the daily global atmospherically-corrected SPOT/VGT synthesis (S1 products) and the algorithm
detects burned areas using a temporal index in the near infrared band. Since the product is not yet
publicly released, we have been able to process only 2008 data, after a special request to the
GEOLAND2 office.
The GFAS (Global Fire Assimilation System) product provides assimilated MODIS fire radiative
power (FRP) observations. FRP relies on the measurements of the rate of middle-IR radiant energy
which is emitted per pixel area from actively burning fires. FRP is largely proportional to burned area
and area-fire intensity (Alexander, 1982). However, the relation between FRP and burned area remains
ambiguous: very small but intense fires can have the same FRP as very large and less intense fires
(Archibald et al. 2013). Hotspot (active fire pixel) data have been frequently used as a proxy for
burned area assuming that area burned is positively related to the number of fire pixel counts (e.g.
Giglio et al. 2006).
The ESA World Fire Atlas (WFA) product provides night-time observations of active fire pixels made
with the (A)ATSR ((Advanced) Along Track Scanning Radiometer on board the polar orbiting Envisat
satellite (with AATSR replacing ATSR in 2002). Here, we use the algorithm 2 product, which applies
a threshold of 308 K to the radiance of the 3.7 μm channel in order to detect fires. The product is
available from the World Fire Atlas (WFA) (http://due.esrin.esa.int/wfa/). The (WFA-HS product
shows very good correlation with the TRMM-VIRS and MODIS-Aqua/Terra monthly night-time fire
counts (Arino et al. 2011). While the number of daytime fires is much higher than the night-time
number, the daytime fire size per fire count is much smaller. As a result, the daytime fires generally
attribute little to the large-scale regional burned area (Schultz, 2002). Most of the daytime fires are
relatively small controlled burns, typically set for agricultural purposes (agricultural waste burning,
small clearing fires). As a result, night-time (WFA-HS product is biased towards larger and
predominantly uncontrolled fires.
The TRMM active fire product relies on day- and night-time observations made across the tropics (38°
S to 38° N) with the Visible and Infrared Scanner (VIRS), which is on-board the Tropical Rainfall
Measuring Mission (TRMM) satellite. Pixels are identified as a fire pixel at daytime if the 3.75 μm
channel brightness temperature is above 320 K. At night-time, the threshold is 315 K. The TRMM
VIRS active fire product is publicly available from http://pps.gsfc.nasa.gov/fireintro.html, which are
uncorrected for bare ground pixels and sun glint pixels.
The WFA and TRMM active fire products are widely used by the climate research community since
they are merged into the GFED burned area and emission inventories. As mentioned above, a
combination of monthly WFA and TRMM fire counts were used to extend the GFED burned area time
series into the pre-MODIS era, a combination of monthly WFA and TRMM fire counts were used to
extend the GFED burned area time series into the pre-MODIS era (i.e. prior to 2000).
Finally, we also include the MODIS active fire counts from the NASA FIRMS MODIS Thermal
Anomalies/ Fire Locations Product, available from https://earthdata.nasa.gov/data/near-real-timedata/firms/active-fire-data. This product provides the geographic coordinates of the centre of a ~ 1 km
pixel in which a thermal anomaly is detected, unaccounted for contaminations by spurious signals
such as volcanoes or gas flaring. Furthermore, the product is uncorrected for cloud cover or missing
observations. The MODIS hotspot product is strongly related to the GFAS-FRP product since both
rely on the MODIS fire products MOD14 and MYD14. MODIS active fire pixels are used as seed
pixels for the contextual burned area detection algorithm in both GFED versions and in MERIS.
D4.1.3 Product Intercomparison Report
Page 3
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
4 Description of Methods
This chapter describes the methodology used to carry out the intercomparison between global burned
area products. The comparison has been done using GFEDv3.1 as the reference product, mainly
because GFED is the most currently used source of BA information by climate and atmospheric
modellers.
4.1 Data Pre-processing
Two types of comparisons were performed, at pixel and at grid level. The former required that all
products had the same basic spatial resolution, refer to the same period and be projected in the same
coordinates. Since there was a diversity of pixel sizes from 300 to 1000 m, we selected an intermediate
value of 500 m to perform as the basis for the comparisons. To compare products at grid level, we
selected the 0.5º resolution cell size, which was common to fire_cci and the GFED products.
To obtain these common matrices at 500 m and 0.5º cell size, the following transformations were
required (Figure 1):
-
Reclassification to burned/unburned from date of detection.
Generation of annual composites from the original temporal tiling.
Generation of global BA mosaics from the original spatial tiling.
Reprojection to geographical coordinates.
Compute BA for 0.5 degree grid cell
Python with the GDAL and Arcpy libraries were used to perform the processing.
The following paragraphs describe the different phases. Not all of them had to be applied to all the
products, depending on the original status of the product. We will illustrate the procedures with the
example of MCD45, which required the application of all processes.
Figure 1: Methodology applied to the intercomparison of BA
The first step was the reclassification of the input BA products. Usually the BA products were
downloaded as raster layers, where the pixel value was the Day of the Year (DoY) when the burning
was detected. To sum up all the burned areas during the year, first we convert the DoY to a binary
scale (burned/ unburned, see Figure 2). The objective was reducing the amount of information and
processing time to generate the annual composites.
D4.1.3 Product Intercomparison Report
Page 4
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
Figure 2: Example of the binarization process: MCD45 of June (left) and the original product (right)
Figure 3: Annual composite of the MCD45 product (left) and single month binarized product (Right)
All monthly files were summed up to a single annual layer. When a pixel burned twice the annual
composite took the value of two. To perform the annual global mosaic of the BA products, the GDAL
Merge algorithm was used. It just required a list of the images that were going to be merged and the
output format (TIFF in this case). In the case of overlapping between the different images, the
algorithm used the values from the first image in the list. The results are shown in Figure 4.
D4.1.3 Product Intercomparison Report
Page 5
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
Figure 4: Representation of the binarized MCD45 mosaic. Burned pixels in red
After the generation of the BA mosaics, the product was reprojected to geographical coordinates,
using the WGS84 datum. The MODIS products were referenced to a sinusoidal projection system,
while Geoland and the fire_cci product was originally available in Plate Carré Geographical
projection.
Finally, the total BA for each 0.5º grid cell was computed calculating BA pixels of each product
included in each half degree cell. The result of this operation was a shapefile where each 0.5º polygon
contained the BA sum of all pixels intersecting each grid cell (Figure 5). To reduce the processing
time all grid cells with more than 80 % of water were removed from the analysis. We used the same
shapefile for all products, as it facilitated further statistical analysis.
Figure 5: 0.5 degree grid over Australia
D4.1.3 Product Intercomparison Report
Page 6
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
4.2 Burned area calculations
Since we were worked with geographical coordinates, the cell size in metric units had to be computed
based on the latitude. To simplify the calculations, the total BA of each cell was estimated from the
proportion of burned to total pixels in each grid cell multiplied by the total area of each cell:
𝐵𝐴 = 𝐴𝑟𝑒𝑎 𝑜𝑓 𝑡ℎ𝑒 0.5⁰ 𝐺𝑟𝑖𝑑 𝑥
∑ 𝐵𝑢𝑟𝑛𝑒𝑑 𝑝𝑖𝑥𝑒𝑙𝑠
∑ 𝑇𝑜𝑡𝑎𝑙 𝑝𝑖𝑥𝑒𝑙𝑠 𝑝𝑒𝑟 0.5° 𝐺𝑟𝑖𝑑
Both the fire_cci products and the GFED product included the estimation of burned area at 0.5°
resolution, and therefore, the computation was based on summing up the fifteen-day periods for each
year.
4.3 Time and Space domains
The intercomparison analysis was done with all products for three years 2006, 2007 and 2008. For
pixel products, a few comparisons were made at continental level, since comparing the products at
global scale would have required a very demanding process. As an example of pixel comparisons, we
selected a tile in Northern Australia as it has high occurrence and it is a good example of Tropical
savana fires. For the remaining of the regions, analysis were done at 0.5º grid level, which allows a
great variety of global assessments. We compared global estimations for total burned area of the three
years, as well as for geographical regions proposed by other researchers and currently used by the
GFED product (Figure 6 and Table 2).
Figure 6: Map of the 14 regions used for the intercomparison analysis (after (Giglio et al. 2006).
Description of the areas is included Table 3.
In terms of periods for temporal comparison analysis, we only generate the comparison of GFED and
fire_cci products. As previously described, the other products had to be computed semi-automatically
from the original pixel products, and it was very time consuming to perform the whole process as
many times as monthly periods were available. We selected GFED and fire_cci, as both were already
gridded temporally (GFED monthly and fire_cci in 15-day periods), and the monthly period was used
for trend comparisons.
4.4 Source of National Statistics
Burned area based on regional and national statistics were compiled until 2000 by Mouillot and Field
(2005), and we updated this database to 2010 for comparison with GFED and ESA BA for their
respective overlapping years 2000 - 2010 and 2006 - 2008. The national statistics account for both
large fires as usually detected by global products, and small fires that are rarely detected by global
coarse spatial-resolution sensors (at 1-km or 500-m resolution). We propose here a full analysis for the
countries where statistical data are available, which account for small fires. We tested the relationship
between different BA data at the national level on a yearly time step, and identified global
D4.1.3 Product Intercomparison Report
Page 7
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
discrepancies. We also tested for the inter-product matching for interannual variability for 2000 - 2010
for GFED and 2006 - 2008 for ESA fire_cci data.
We collected burned area statistics at the national level from the forest services or scientific
publications when available. The sources information are presented in Table 2.
Table 2: Burned area statistics on national level
Country
Europe
Number of states / Source
subdivisions
50
http://www.nifc.gov/fireInfo/fireInfo
_statistics.html
11
http://nfdp.ccfm.org/data/compendiu
m/html/comp_31e.html
22
http://forest.jrc.ec.europa.eu/effis/
Chile
1
Japan
1
Algeria
1
Tunisia
1
Russia
1
Mongolia
1
Korea
Ukraine
Total
1
1
91
USA
Canada
D4.1.3 Product Intercomparison Report
http://www.conaf.cl/cms/editorweb/e
stadisticas/inc_fores/nacionaloccurencia_dano-64_11.pdf
Goto and Suzuki, 2013
Brief
description
of
methods for BA estimates
Ground observation from
forest services
Ground observations from
forest services
Ground observation or high
resolution remote sensing
analysis
Ground observation from
forest services
Ground observation from
forest services
Madoui et al. (2002) and pers. Ground observation from
Comm.
forest services
Direction Générale des forêts, Ground observation from
Tunisia
forest services
Dubinin et al. (2010)
AVHRR/landsat
Goldammer et al. (2007)
Ground/aerial/AVHRR
Hayasaka (2011)
Hotspots and ground obs.
Russian federation forest Agency Ground obs. official
(2010)
AVHRR
Shvidenko et al. (2011)
AVHRR
Sukhinin et al. (2004)
SPOT4 VGT
Zhang et al. (2013)
GFED3,
ground
obs,
Shvidenko et al. (2013)
AVRHH
Erdenetuya(2013)
Ground
obs/
Farukh et al. (2009)
AVHRR/MODIS 250m
MODIS hostspots
Lee (2008)
Ground obs. Forest serv.
Zibtsev (2010)
Ground obs. Forest Serv.
Page 8
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
5 Results
5.1 Comparisons between the pixel products
Figure 7 includes an example of intercomparison analysis of pixel products for Northern Australia.
The map shows coincidences between different MCD45, MCD64, and MERIS_cci pixel products.
The original pixels were transformed to a common pixel size (500 m). At pixel level we only
compared these three products, as they provided the highest accuracies in the validation report (see
Padilla and Chuvieco 2014), and had also the most similar BA estimations at global scale. In spite of
this, the common area between the three sensors is not very high, as only 66.93 % of all burned pixels
were detected by two or three products, where the 48.5 % were common between MERIS and MCD45
and 53.72 % between MERIS and MCD64. Many areas were detected just by a single sensor,
particularly in the Western and Southern part of the tile.
Figure 7: Burned area detected by MERIS, MCD45 and MCD64 in 2008
5.2 Global estimation of burned areas (grid products)
Figure 8 shows the total amount of BA estimated by the different products in the three target years
(with the exception of Geoland2, from which only 2008 has been available). The computation is based
on summing up all 0.5º grid cells. Total estimated area ranges from 2.4 to 2.6 million km² for the
VGT_cci product to 4.8 to 5 million km² for the MERGED_cci. The yearly GFED estimations1 were
in the range of 3.2 to 3.5 million km². The closest estimations to these values were obtained from
MCD45 and MERIS _cci products, with 3.31 - 3.52 million km² and 3.62 to 3.77 million km²,
respectively. Geoland2 provides the lowest estimations for 2008 with 2.05 million km². Since we only
had one year available for this product, it will not be included in the more detailed regional analysis.
1
All GFED data provided in this report refers to the BA estimations from this database. These BA estimations
are very close to those provided by the MCD64 product, which is the input for the statistical estimations of BA,
but it may be some minor differences in months with poor MODIS observations.
D4.1.3 Product Intercomparison Report
Page 9
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
Figure 8: Estimation of burned area by different BA products
5.3 Regional estimation of burned areas (grid products)
Figure 9 includes the average BA values for the 2006-2008 periods in each of the geographical regions
used in this analysis. The two regions including the Northern and Southern Tropical Africa are the
most extensively burned by all products, although VGT_cci shows a clear underestimation with
respect to other products. Australia is the third most extensively burned, while the Boreal and Central
regions of Asia and South America follow. The lowest values are found for Equatorial and Temperate
regions. Tables 2 to 4 show the BA estimations for the different geographical regions defined in the
GFED product. Whenever available, official fire statistics from the countries have also been included,
as explained in section 4.4.
The different products show similar trends for the most fire affected areas (Northern and Southern
African hemispheres), which in most products account for 60 to 70 % of all that is burned worldwide.
A notable exception is the VGT_cci and GEOLAND-2 products, which clearly underestimate BA in
Tropical regions with respect to other BA products, as these two regions only account in these
products for less than 35 % of global BA. For most products and periods, Australia was the third most
burned region, with more than 7 % of the global BA, with particular impact in 2006 and 2007.
The estimations from products are very similar between GFED, MCD45 and MERIS_cci, while
VGT_cci and GEOLAND2 estimated much lower BA than other products, with the exception of
Middle East and Boreal regions (particularly in North America) with much higher estimations.
MERGED_cci regularly produced higher estimations as it includes the detections of both MERIS and
VGT products, and therefore, propagates the overestimations of VGT_cci in Boreal regions.
D4.1.3 Product Intercomparison Report
Page 10
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
Figure 9: Average BA for each region (2006-2008). For Geoland-2 only 2008 was computed
Table 3: BA estimations from different geographical regions (2006)
GFED
MCD45
VGT_cci
MERIS_cci
MERGED_cci
Official stats
AUST
530913
369510
236937
339800
444391
-
BOAS
43288
70383
364737
113926
333515
-
BONA
19117
12904
148677
13133
105345
23362
CEAM
12571
15513
39366
17256
49284
-
CEAS
175395
194598
289138
172419
357372
-
EQAS
26826
3736
15356
15085
27769
-
EURO
4771
17633
53279
9891
45492
3278
MIDE
9015
15283
36416
8637
35569
-
NHAF
1151535
1183013
312973
1227041
1368218
-
NHSA
14925
4199
28729
15591
39018
-
SEAS
59303
72390
102296
87737
164457
-
SHAF
1221849
1263320
495706
1395718
1589177
-
SHSA
124969
123241
109367
204359
281112
-
TENA
24214
27202
141055
29676
131037
3709249446
3418690
3372924
2374032
3650268
4971755
-
WORLD
AUST: Australia, BOAS: Boreal Asia, BONA: Boreal North America, CEAM: Central America, CEAS: Central
Asia, EQAS: Equatorial Asia, EURO Europe, MIDE: Middle East, NHAF: Northern Hemisphere Africa, NHSA:
Northern Hemisphere South America, SEAS: Southeast Asia, SHAF: Southern Hemisphere Africa, SHSA:
Southern Hemisphere South America, TENA: Temperate North America. Official fire statistics come
D4.1.3 Product Intercomparison Report
Page 11
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
Table 4: BA estimations from different geographical regions (2007)
GFED
MCD45
VGT_cci
MERGED_cci Official stats
MERIS_cci
AUST
487099
322263
235144
374883
478750
-
BOAS
32399
50344
247339
43208
196128
-
BONA
15458
9581
169390
9656
114374
18131
CEAM
10611
12434
42309
16085
50511
-
CEAS
124669
164356
240715
151462
306192
-
EQAS
4776
1692
13903
6370
18045
-
EURO
9655
17123
46802
14450
44470
6547
MIDE
11770
13881
32148
8481
32419
-
NHAF
1234422
1279797
330367
1236476
1376471
-
NHSA
25153
26469
66666
27424
79773
-
SEAS
98740
85087
89737
113824
176710
-
SHAF
1242137
1184520
400834
1396252
1564418
-
SHSA
338357
334400
188010
338256
445107
-
TENA
26641
20508
135648
35257
132552
35547-49447
3661888
3522453
2239012
3772086
5015920
-
WORLD
Table 5: BA estimations from different geographical regions (2008)
GFED
MCD45
VGT_cci
MERIS_cci
MERGED_cci
GEOLAND
Official
stats
AUST
266319
240141
169482
223018
308898
180554
-
BOAS
120490
132766
455767
106510
380643
152015
-
BONA
14465
8036
142751
6350
96855
94011
17243
CEAM
11648
14252
33654
16153
42981
41415
-
CEAS
139938
210543
191662
171311
297947
202488
-
EQAS
4239
1505
19830
6068
22622
1073
-
EURO
5367
15727
41641
12624
40135
29513
2114
MIDE
6036
5010
32162
6945
31065
50011
-
NHAF
1176668
1195738
257278
1275054
1378733
459976
-
NHSA
17754
6379
34455
24726
51765
7209
-
SEAS
69744
70929
82553
99912
160901
33836
-
SHAF
1315416
1226237
502814
1473301
1666912
530674
-
SHSA
133827
162230
144701
188187
285709
132714
-
TENA
14524
17897
119860
13118
100896
129384
21152 29785
3296434
3307390
2228610
3623277
4866063
2044873
-
WORLD
D4.1.3 Product Intercomparison Report
Page 12
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
5.4 Spatial differences between GFED and Fire_cci BA grid products
The following section illustrates spatial differences between GFED and fire_cci BA products. We
have selected GFED as it is currently the most standard BA product used by climate and carbon
modellers (Mouillot et al. 2014). Over- and underestimation in the following paragraphs refer to
whether the global BA products estimate above or below the GFED product. They do not necessarily
mean actual over or under estimations over true burned area as GFED is obviously not error free.
Globally, MERIS exhibits largely the same general spatial pattern of mean annual burned area as
GFED (both, GFED3 and GFEDv4) (Figure 10). Conversely, there are drastic differences in both
spatial extent of fire activity, and the locations of regions with most intense fire activity when
comparing the MERGED_cci (respectively the VGT_cci) product to MERIS_cci or GFED.
Figure 10: Mean annual burned area (2006 - 2008) expressed as fraction burned per 0.5-degree grid cell in
the different burned area products. The grid cell area at the Equator is 3090 km2 and 2670 km2 at 60° N
or S.
Figure 11 shows the spatial divergences between the GFED and MERIS_cci product for 2008. Most
grid cells show differences lower than ± 10 % of BA between the two products, with the exception of
some cells in Angola-Zambia-Bostwana, where GFED produces higher estimations than MERIS_cci
and Namibia where the opposite is observed. As expected from our previous analysis, the VGT_cci
product shows more differences with GFED than MERIS_cci, with a clear underestimation in Tropical
regions and an overestimation of temperate forest, particularly in Central Asia and the western regions
of Canada and the USA (Figure 12).
The spatial differences between GFED and the MERGED_cci summarize what was observed in the
two single products (Figure 12). Overestimations are observed in Central Asia, Western North
America, and Tropical Africa, while underestimation is observed in both hemispheres of Tropical
Africa, particularly in Namibia-Bostwana.
D4.1.3 Product Intercomparison Report
Page 13
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
Figure 11: Spatial difference between GFED and MERIS
D4.1.3 Product Intercomparison Report
Page 14
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
Figure 12: Spatial difference between GFED and VGT_cci
D4.1.3 Product Intercomparison Report
Page 15
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
Figure 13: Spatial difference between GFED and MERGE
D4.1.3 Product Intercomparison Report
Page 16
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
To better understand the spatial differences between the MERGED_cci product and the GFED after
analysing the differences with the input product (MERIS and VGT), we have also computed the
differences between the results of these two sensors. In spite of having lower estimations than GFED,
the VGT_cci produces higher estimations in temperate regions, which do not coincide with
MERIS_cci. Therefore, when summing up both sensors in the merging procedure, the overall
estimations of the MERGED_cci are higher than GFED.
In spatial terms, the differences between MERIS_cci and VGT_cci are portrayed in Figure 14. Most
differences are in the range of ± 10 %, with the exception of Tropical savannas, where MERIS
produces much higher estimations (> 30 %) than VGT_cci. Since this is the most burnable area
worldwide, this is the main explanation of the low global BA area estimates of VGT_cci. The opposite
is true for temperature and boreal regions, where VGT_cci offers higher BA estimations than
MERIS_cci, most commonly in the range of 5 - 10 %.
5.5 Percentile comparisons
We also compare the spatial patterns of fire_cci and GFED burned area with those of i) GFAS-FRP, ii)
MODIS-HS, iii) WFA-HSs and iv) TRMM-HS. The latter three are independent active fire count
products. GFAS-FRP and MODIS-HS are strongly interrelated because both products rely on the
MODIS active fire product. Because MODIS-HS information is used in the MERIS BA algorithm as
well as the MCD64 product, which is the basis of GFED, MODIS-HS largely constrains the spatial
fire patterns in MERIS_cci and GFED4. As explained above, the relation between active fire products
and burned area is not straightforward and absolute and relative comparisons of FRP or hotspots with
burned area are hence of limited explanatory power.
In order to maximize the relative comparability of the global spatial gradients in the different products,
global field percentile statistics is used. The field percentile indicates the rank of an individual fire grid
cell relative to the other values in each global fire product. The percentiles are calculated by rankordering the values of the complete set of mean annual global fire grid cells in each product. Rank
ordering was done using the burned fraction (fraction of the grid cell area burned) and not the total
burned area per grid cell. Figure 15 shows the spatial pattern of the different burned area and active
fire products with a colour scaling standardised by intervals of 10 percentiles. The 10th percentile
colour code, for example, contours 10 % of all global fire grid cells within each fire products, and only
those which, by rank, have the lowest grid cell values. The total number of valid of fire grid cells and
the percentile statistics of each product is given in Table 6.
.
D4.1.3 Product Intercomparison Report
Page 17
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
Figure 14: Spatial difference between MERIS and VGT
D4.1.3 Product Intercomparison Report
Page 18
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
Figure 15: Field percentiles of mean annual i) burned area (MERIS, GFEDv4, VGT, and MERGED; ii)
FRP (GFAS) and iii) hotspot (HS) counts (MODIS, WFA and TRMM). The total number of valid fire grid
cells and the percentile statistics of each product are given in Table 6. Fire signals in the Southern Pacific
in the WFA-HS product are created by an artefact in the original August 2008 data
The numbers of Table 6 refer to the mean annual data (years 2006 - 2008) shown in Figure 15. Please
note that TRMM data only cover the domain from 38° S to 38° N while all other products have the
global domain. The NFG for the TRMM domain, complemented by the mean annual number of
monthly fire grid cells (NMFG) for this domain, is given in the rightmost columns.
D4.1.3 Product Intercomparison Report
Page 19
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
Table 6: Global number of annual fire grid cells (NFG) and absolute values corresponding to the field
percentile ranks in each product.
In accordance with Figure 15, MERIS_cci and GFEDv4 show very similar spatial patterns with
respect to the global field percentiles. In both burned area products, grid cells with very high fire level
(90th percentile and greater) are almost exclusively confined to tropical Africa and tropical Australia.
In the VGT_cci product, grid cells with very high fire level have a much patchier pattern than
MERIS_cci and GFED; with high fire level grid cells (70th percentile and greater) also prominent
across the Eurasian taiga belt, Alaska, and the western regions of temperate US.
GFAS-FRP and similarly MODIS-HS show a much larger global spatial extent of fire activity than
MERIS_cci or GFED, notably across Western Europe, the East of China, the westerly half of South
America, the United States, and the southern Canada. These fires typically have low FRP values: they
are generally below the 25th percentile of the global population. It is a common feature in GFAS-FRP,
MERIS_cci and GFED that grid cells with very high fire levels exhibit a strong spatial clustering
across the same regions in tropical Africa and tropical Australia. In contrast to MERIS_cci and
GFED, GFAS-FRP shows relatively high fire levels also across the tropical savanna regions of South
America.
Compared to all other gridded fire products, WFA hotspots show the lowest spatial extent of global
fire occurrence. Furthermore, the spatial clustering of high to very high level fire grid cells across
tropical Africa is much less pronounced than in the burned area or in the GFAS FRP products.
Reversely, these grid cells are relatively more prominent in tropical South America.
TRMM hotspot show very widespread spatial occurrence of fire grid cells within the domain from 38°
S to 38° N, which is the spatial coverage the TRMM-HS product. The spatial extent of fire grid cells
in this domain has large similarities to the extent in the GFAS-FRP, MODIS-HS, VGT_cci or
MERGED_cci products. All products, for example, exhibit fire grid cells in the western regions of
South America and in the central regions of Australia, where fire activity is almost absent in
MERIS_cci, GFED and WFA-HS.
The global number of valid of fire grid cells (NFG), given as mean annual totals in Table 6, is lowest
in WFA-HS, second lowest in MERIS_cci and highest in MERGED_cci. The mean annual number in
the MERGED_cci product is only 2 (global) and 3 (TRMM domain) percent higher than the VGT_cci
product, reflecting that the merging of VGT_cci with MERIS_cci only slightly increases the area
extent of fire grid cells on a time-integrated scale. The merging, however, leads to a pronounced
increase in the total number of monthly fire grid cells (by 24 and 39 %, respectively, globally and in
the TRMM domain). This reflects that the merging primarily leads to an increase of fire duration
and/or of fire re-occurrence.
Yet, the global NFG in the VGT_cci product alone is 73 and 88 %, respectively, higher than in
GFED4 and MERIS_cci. It is still 23 % higher than in MODIS-HS, which has the highest NFG of all
active fire products. The higher number of global NFG in VGT_cci is over-proportionally related to a
higher NFG in the extratropics (i.e. outside the TRMM domain). Here, NFG in VGT_cci is 131 and
149 % higher than in GFED4 and MERIS_cci, respectively, and still 57 % higher than in MODIS-HS,
pointing to substantial overestimation of burned area in the boreal regions by the VGT_cci product.
D4.1.3 Product Intercomparison Report
Page 20
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
5.6 Comparison of the size distribution of globally gridded burned area
Figure 16 shows the cumulative size distribution function (CDF) of gridded monthly burned area in
MERIS_cci, GFED4, GFED3, VGT_cci and MERGED_cci. It highlights that the size distribution
function of MERIS_cci and GFED4 is in general very similar, except for smaller fires in the nonintersecting subset. When fire size is very small, i.e. below 0.5 km² per grid and month, the
distribution function of MERIS_cci for the non-intersecting subset stays above the GFED4 function.
The MERIS_cci distribution function smoothly decreases to zero towards the lowest detectable burn
size of the MERIS_cci product, which is 0.04 km² per grid and month. The GFED4 distribution
function exhibits more discrete (stepwise) characteristics and reaches zero already at 0.21 km², which
is the lowest detectable burn size of the GFED4 mapping approach (Giglio et al. 2013). As a result,
very small fires in MERIS_cci contribute a larger share to the total number of fire grid cells than
GFED4. Around 5.4 % of all monthly MERIS_cci fire grid cells have burn sizes smaller than the
lowest burn size in GFED4. The ability of MERIS_cci to resolve smaller fires can be explained by the
higher spatial resolution of the sensor, which is 300x300 m for MERIS_cci and 500x500 m in
MODIS (Oliva and Chuvieco, 2013). The size distribution function of GFED3 shows the highest
relative contribution of small fires (below 1 km² per grid and month) among all burned area products.
This representation of smaller fires in GFED3, however, is exclusively related to indirectly mapped
grid cells (Giglio et al. 2010). In GFED3, directly mapped burned area (using MCD64) are
supplemented by burned area indirectly derived from hotspot information which is scaled to burned
area at 0.5º spatial resolution. The indirectly mapped burned area grid cells comprise a larger fraction
of smaller fires than GFED4, which is only directly mapped. When the indirectly mapped grid cells
area is excluded the size distribution function of GEFD3, and GEFD4 are quasi-identical, as reflected
by the CDFs in the intersecting subset in Figure 16.
The box plots of MERIS_cci and GFED show that both products have largely similar variability (as
reflected by the interquartile range) and skewness (as reflected by the distance of the quantiles to the
respective medians).
In the intersecting subset, the CDF of the MERGED_cci product shows a flatter progression than
MERIS_cci, GFED, or VGT_cci when fire size is small. The function is steep for intermediate sized
fires and flattens again towards larger fires. This reflects that size distribution MERGED_cci product
has a stronger dominance of intermediate sized fires and contains relatively less small and large fires
than GFED. The boxplot reflects that the size distribution of burned area in MERGED_cci product,
and even more so in the VGT_cci product, has a much lower variability than GFED or MERIS_cci:
the lower and upper quartiles differ from the median only by a factor of 3 to 4, while they differ by a
factor of 6 ~ 7 in GFED and MERIS_cci data.
D4.1.3 Product Intercomparison Report
Page 21
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
Figure 16: Cumulative distribution function (CDF) of annual burned area (BA) per fire grid cell (left
panel) and boxplot of monthly BA per fire grid cell (right panel) over 2006 to 2008 in MERIS_cci,
GFED4, GFED3, VGT_cci and MERGED_cci. The topmost panels show the distribution of burned area
size in all fire grid cells, which further sub-divided into intersecting (middle) and non-intersecting
(bottom) fire grid cells. For the intersecting grid cells, burned area is greater than zero in all the five
products. This subset excludes a substantial fraction of small and large fires which fall outside the
detection limit of any of the four products. It thereby also excludes a large fraction of false alarms. The
remaining fire grid cells are categorized as non-intersecting. The number of fire grid cells (N) contained in
the entire population and in the subsets is displayed as well.
.
Figure 17 shows in more detail the similarities and differences in the monthly fire size distribution
between MERIS_cci and GFED4 on the one hand, and MERIS_cci, VGT_cci and MERGED_cci on
the other hand. In this more detailed analysis, the division into intersecting and non-intersecting
subsets was done with the individual product pairs, only.
The box plots of MERIS_cci and GFED4 show that both products have largely similar central values
(as reflected by the median), variability (as reflected by the interquartile range) and skewness (as
reflected by the distance of the quantiles to the respective medians) in the intersecting and nonintersecting subsets.
D4.1.3 Product Intercomparison Report
Page 22
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
In contrast, the intersecting subset of the MERIS_cci, VGT_cci and MERGED_cci fire grid cells
shows pronounced differences in terms central values and variability: The VGT_cci median (8.3 km2
per grid and month) is a factor of 2.27 lower than in the MERIS_cci product (18.8 per grid and
month) while the median of the MERGED_cci product (33.5 km2 per grid and month) is 1.8 times
higher than the MERIS_cci product. Accordingly, the MERGED_cci median is three times higher
than in the VGT_cci product. Furthermore, the variability in the MERIS_cci product is 1.4 times
higher than in the VGT_cci or MERGED_cci product. In the non-intersecting grid cells, the
differences in the medians of MERIS_cci, VGT_cci and MERGED_cci are small. The median values
are 3.1, 2.6 and 2.7 is 2.6, 2.7, and 3.1 km2 per grid and month, respectively.
Figure 17: Cumulative distribution function (CDF) (left panel) and boxplot (right panel) of monthly
burned area (BA) per fire grid cell (left panel) over 2006 to 2008 in MERIS_cci and GFED4 (top) and
MERIS_cci and VGT_cci (bottom), sub-divided into intersecting (labelled with '+') and non-intersecting
(labelled with '-') fire grid cells. The number of fire grid cells (N) contained in the subsets is displayed as
well.
To summarise, the MERGED_cci (respectively the VGT_cci) product reproduces well medium-sized
fire grid cells, but tends to miss small and large fires. The size distribution of MERIS_cci is largely
similar to the distribution in GFED, except that MERIS_cci tends to capture more very small fires.
5.7 Fire seasonality
Figure 18 shows the monthly BA for the GFED and fire_cci products. The seasonal trends are similar
for all products, particularly for GFED and MERIS_cci, which show almost parallel lines with very
similar BA values, with the exception of July to September of 2008, when the two estimations are
displaced by approximately one month. The MERGED_cci product shows similar trends as MERIS
or GFED, but the estimations are always higher.
D4.1.3 Product Intercomparison Report
Page 23
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
The product that shows higher diversity is VGT_cci. On one hand, the BA estimations are lower, as
explained above. On the other hand it is noticeable that this product does not include the maximum
found by others in December, in coincidence with the dry season of Northern Tropical regions.
Figure 18: Monthly BA estimations for the GFED and fire_cci products
5.8 Statistical trends
5.8.1 Fire_cci products
Figure 19 shows the cell to cell correlation between MERIS_cci and GFED estimations. Following our
previous comments, the correlation between the two datasets is very significant (p < 0.01), with r²
values higher than 0.68 in all three years.
The equations of the fitted model are:
2008: GFED = 3,90266 + 0,820014*MERIS_cci
2007: GFED = 5,27735 + 0,854073*MERIS_cci
2006: GFED = 5,80323 + 0,803882*MERIS_cci
D4.1.3 Product Intercomparison Report
Page 24
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
Figure 19: Correlations between GFED and MERIS_cci for the three study years
The slopes of the lines are closed to one, indicating again a close estimation between the two datasets,
with an overestimation of MERIS_cci versus the GFED. When this analysis is done at regional level,
larger differences are observed between the two datasets (Table 7). The correlations are higher for
North America (both Boreal regions and Temperate), Australia, the Northern hemisphere of South
America and the Tropical regions of Africa, and are much lower for Central and the Southern
hemisphere of America and, particularly, in Equatorial Asia. The same analysis was done for the
different geographical areas previously presented (table below).
Table 7: Correlation between MERIS_cci and GFED BA for the different regions.
MERIS_cci 2008
R-2
Abbrev.
BONA
TENA
CEAM
NHSA
SHSA
EURO
MIDE
NHAF
SHAF
BOAS
CEAS
SEAS
EQAS
AUST
TOTAL
0,6551
0,6496
0,3188
0,6481
0,3820
0,5979
0,2502
0,7550
0,5818
0,4008
0,5160
0,3857
0,0565
0,7514
0,7069
D4.1.3 Product Intercomparison Report
RMSE
(Km²)
10,2177
13,5714
27,6547
40,0899
57,8599
6,173
7,28673
259,968
343,422
54,5282
47,5924
100,523
11,0038
126,127
114,412
2007
R-2
0,613503
0,477829
0,1406
0,439927
0,594489
0,77254
0,352666
0,758366
0,562401
0,56189
0,321834
0,380533
0,0787253
0,835706
0, 7114
RMSE
(Km²)
10,6852
34,9658
25,7048
65,7767
116,38
8,83588
12,425
264,481
327,213
15,2158
56,5555
116,761
13,3815
155,695
116,537
2006
R-2
0,913615
0,5148
0,0331541
0,323198
0,189168
0,530038
0,283028
0,702468
0,627107
0,130492
0,549457
0,290207
0,184275
0,757259
0,6837
RMSE
(Km²)
6,54719
26,4362
27,9648
43,5874
62,424
6,38217
10,8207
262,065
308,91
30,8448
61,8339
99,4652
68,0172
90,2861
111,698
Page 25
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
For the VGT_cci product the correlations with GFED are lower, as expected from our previous
comments, although still significant (p < 0.01), with a clear tendency towards underestimation (Figure
20).
Figure 20: Correlations between GFED and VGT_cci for the three study years
The regression equations in this case are:
2008: GFED = -1,65903 + 1,53984*VGT_cci
2007: GFED = -5,71494 + 1,84693*VGT_cci
2006: GFED = -4,16732 + 1,58534*VGT_cci
Table 8 shows the regression parameters for the GFED and VGT_cci products for different regions.
Values are much lower than those found for MERIS_cci, with particular divergences in North
America, Central America, Europe, Middle East and Equatorial Asia. The best fits were found with
Australia and the Southern hemisphere of Africa.
Table 8: Results of the VGT_cci correlation analysis along the different basin regions
VGT_cci 2008
2007
2006
Abbrev.
R-squared
RMSE
(Km²)
R-squared
RMSE
(Km²)
R-squared
RMSE
(Km²)
BONA
0,0258655
17,1711
0,0184451
17,0282
0,112796
20,9821
TENA
0,0115998
22,7928
0,0832194
46,3307
0,0622179
36,7528
CEAM
0,0720166
32,2785
0,0495194
27,0327
0,0575048
27,6104
NHSA
0,401407
52,2898
0,306667
73,1848
0,252144
45,8183
SHSA
0,425799
55,7712
0,628762
111,354
0,360262
55,4481
EURO
0,00359904
9,7176
0,0437749
18,1166
0,0291208
9,17317
MIDE
0,0126106
8,3621
0,0470418
15,0755
0,103565
12,0994
D4.1.3 Product Intercomparison Report
Page 26
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
NHAF
0,469644
382,458
0,580157
348,626
0,522962
331,832
SHAF
0,660868
309,275
0,649782
292,726
0,685516
283,687
BOAS
0,0851715
67,3779
0,0813171
22,0337
0,0328813
32,5301
CEAS
0,171835
62,2573
0,261391
59,0221
0,254526
79,5379
SEAS
0,174657
116,521
0,233106
129,914
0,207877
105,076
EQAS
0,00105428
11,3226
0,0007424
13,9363
0,0027906
75,2039
AUST
0,656646
148,223
0,673404
219,518
0,657335
219,016
TOTAL
0,425199
156,425
0,520499
150,211
0,503357
143,359
Figure 21 shows the correlations of GFED and MERGED_cci for the three study years. The
correlations are all significant (p < 0.01), with higher values than both MERIS_cci and VGT_cci,
although the slopes of the regression lines indicate higher overestimation than MERIS_cci.
Figure 21: Correlations between GFED and MERGED_cci for the three study years
The equations of the fitted models are:
2008: GFED = -4,28902 + 1,35969* MERGED_cci
2007: GFED = -2,60686 + 1,37717* MERGED_cci
2006: GFED = -3,26221 + 1,29373* MERGED_cci
Table 9 shows the regression parameters for the GFED and MERGED_cci products for different
regions. The best fits were found for Australia, the Northern hemisphere of Africa and South America,
and the lowest correlations were observed for Equatorial Asia, the Boreal and temperature regions of
North America, the Middle East and Boreal Asia.
D4.1.3 Product Intercomparison Report
Page 27
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
Table 9: Results of the MERGED_cci correlation analysis along the different regions.
MERGE 2008
2007
2006
Abbrev.
R-squared
RMSE
(Km²)
R-squared
RMSE
(Km²)
R-squared
RMSE
(Km²)
BONA
0,145308
16,084
0,180517
15,559
0,446062
16,5793
TENA
0,147168
21,172
0,32903
39,6358
0,233158
33,2301
CEAM
0,250363
29,0114
0,145822
25,6266
0,0691192
27,4398
NHSA
0,626849
41,2852
0,436932
65,9524
0,383773
41,5911
SHSA
0,435794
55,2837
0,630539
111,087
0,240917
60,3991
EURO
0,420705
7,40954
0,537604
12,5981
0,352577
7,49084
MIDE
0,14855
7,76517
0,244263
13,4251
0,256265
11,0208
NHAF
0,769488
252,143
0,783907
250,113
0,736357
246,689
SHAF
0,649521
314,406
0,63939
297,037
0,685217
283,821
BOAS
0,291909
59,2778
0,354429
18,4704
0,134468
30,7742
CEAS
0,513072
47,738
0,336127
55,9564
0,488694
65,8716
SEAS
0,406587
98,8017
0,418195
113,156
0,342167
95,7553
EQAS
0,0191586
11,2196
0,032946
13,7099
0,137539
69,9385
AUST
0,773878
120,287
0,857244
145,132
0,795637
169,139
TOTAL
0,730757
107,058
0,744374
109,676
0,715956
108,416
5.8.2 MCD45
Figure 22 shows the cell to cell correlation between MCD45 BA and GFED estimations. As it may be
expected, the correlation between the two datasets are very significant (p < 0.01), with r² values higher
than 0.83 in all three years.
D4.1.3 Product Intercomparison Report
Page 28
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
Figure 22: Correlations between GFED and MCD45 in the three study years
The equation of the fitted model is:
2008: GFED_KM = 2,71048 + 0,928353*MCD45
2007: GFED_KM = 3,82852 + 0,948954*MCD45
2006: GFED_KM = 4,21443 + 0,909401*MCD45
Which indicate that both datasets have very similar estimations, with a tendency towards
overestimation by MCD45. Table 10 shows regional trends of this correlation trends. The
determination coefficients are high for the Boreal regions and the two hemispheres of Africa and
South America, but are quite low in Equatorial Asia and the Middle East.
Table 10: Results of the MCD45 correlation analysis along the different basin regions.
MCD45
2008
Abbrev.
R-squared
RMSE
(Km²)
R-squared
RMSE
(Km²)
R-squared
RMSE
(Km²)
BONA
0,77945
8,17037
0,84412
6,78586
0,842291
8,84637
TENA
0,223969
20,1962
0,594321
30,8196
0,621093
23,3618
CEAM
0,629091
20,4069
0,535991
18,8878
0,429432
21,4826
NHSA
0,719212
35,813
0,622035
54,0351
0,598981
33,5515
SHSA
0,630022
44,7678
0,896446
58,8112
0,761704
33,8411
EURO
0,746112
4,90527
0,516373
12,8841
0,272841
7,93874
MIDE
0,325698
6,91033
0,422905
11,7316
0,504102
8,99916
NHAF
0,887452
176,185
0,876733
188,903
0,870612
172,818
SHAF
0,803911
235,172
0,843308
195,801
0,855269
192,451
BOAS
0,726482
36,8417
0,549951
15,4218
0,616352
20,4886
CEAS
0,63305
41,4415
0,638959
41,2653
0,640016
55,2713
D4.1.3 Product Intercomparison Report
2007
2006
Page 29
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
MCD45
2008
2007
2006
SEAS
0,517388
89,1014
0,398631
115,042
0,497107
EQAS
0,0195561 11,2173
0,007419
13,8897
0,0114563 74,8764
AUST
0,673173
144,612
0,829767
158,484
0,644171
223,184
TOTAL
0,848607
80,2787
0,859079
81,4322
0,834447
82,7694
83,7226
The analysis of the parameters of the correlation does not show overestimation in most of the cases;
furthermore, the slope is below one and the constant is near to zero.
5.9 Matching of MERIS_cci burned area with different hotspot products
Both, MERIS_cci and GFED (v3 and v4), use MODIS hotspot pixels as a “seed” in the burned area
detection. As a result, they are both similarly influenced by omission errors in the MODIS HS product.
Here, we make use of two independent active fire products, WFA-HS and TRMM-HS, in order to
analyse the occurrence of active fire pixels not picked up by the MERIS_cci BA product.
When intersecting concurrent and co-located monthly hotspot data with monthly MERIS_cci burned
area, then WFA-HS grid cells not picked up by MERIS_cci BA contribute 41 % to the total number
of global hotspot counts and 44 % to the total number in the TRMM domain. The corresponding
numbers for the MODIS-HS product are 23 and 21 %, respectively. Grid cells not picked up by
MERIS_cci (also denoted as the “missed” subset) contribute even 62 % to the total number of TRMM
hotspot counts. When intersecting annually integrated data in the TRMM domain, the contribution of
“missed” hotspots decreases to 31 % in WFA-HS and to 37 % in TRMM-HS. In MODIS-HS, the
contribution is 10 %, only. Finally, when spatially intersecting time-integrated data, the “missed”
contribution decreases to around 25 % in both, WFA-HS and TRMM-HS, and to 5 % in MODIS-HS.
This strong decrease with increasing temporal integration indicates that the omission by MERIS_cci
could be partially related small temporal displacements in the MERIS_cci product. In these cases,
MERIS_cci actually detects burned area of a given fire, but assigns it to a different date of burning
than the active fire products (e.g. temporal shift by one month).
A large fraction of all WFA-HS missed by MERIS_cci burned are most likely related to false alarms:
18 % of all WFA-HS fire counts are related to fires on barren or sparsely vegetated surfaces (Table
11), i.e. in areas where fuel loads are too low for open vegetation fires to sustain burning. Most of
these hotspots occur in areas where gas flaring is prominent (Figure 23). The contamination by
spurious (i.e. industrial) signals in the TRMM-HS and MODIS-HS products, in contrast, is very small
(1.2 and 0.4 %, respectively).
D4.1.3 Product Intercomparison Report
Page 30
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
Table 11: Total number of active fire pixels per year in the TRMM domain in the WFA-HS, TRMM-HS
and MODIS-HS products, differentiated by land cover class. The contribution of hotspots missed by
MERIS_cci in each land cover class is given as well as the partitioning of the total number of hotspot
missed across the individual land cover classes. Intersecting with MERIS_cci was performed with timeintegrated data (period 2006-2008). The land cover classes are defined in Figure 24. “Vegetated” includes
all land cover classes except predominantly urban and/or barren and sparsely vegetated grid cells.
When only fire signals on vegetated surfaces are taken into account, the partitioning of hotspots
missed by MERIS_cci in the wider tropics (i.e. TRMM domain) then in absolute terms, omission in
WFA-HS and TRMM-HS is highest in open shrublands (3,560 and 16,869 hotspots per year,
respectively), followed by tropical forests (3,257 and 10,381 hotspots per year, respectively) and
agricultural areas (784 and 7,683 hotspots per year, respectively). The total number of MODIS-HS
omitted by MERIS_cci is highest in tropical forests, followed by agriculture and savannah (133,600;
29,201 and 18,078 hotspots per year, respectively).
There is an over-proportional increase in the share of hotspots in open shrublands in the “missed”
subset in both, WFA-HS and TRMM-HS. In this subset, hotspots in open shrublands contribute 38 %
and 40 %, respectively, to the total number of hotspots in vegetated areas whereas the contribution in
the entire population is 13 % and 18 %, respectively. In contrast, the contribution of this land cover
class in the MODIS-HS product remains unchanged. The share also strongly increases for hotspots in
tropical forests, notably in the MODIS-HS product. Here, the share increases from 17 to 60 %.
Reversely, there is an over-proportional decrease in the share of hotspots in savannah (SA and WSA)
in all three products. In all other land cover types, the relative change across all three products is either
ambiguous or small. For example, in the TRMM-HS and the MODIS-HS products, the share of
hotspots in agricultural areas in the “missed” subset is absolutely 4 % and 8 % higher, respectively,
than in the entire population. In contrast, the share in WFA-HS remains almost unchanged (below 0.5
%).
Omission by MERIS_cci is more relevant in terms of the number of fire grid cells [NFG] than in
terms of hotspot counts. Table 12 shows that MODIS-HS fire grid cells omitted by MERIS_cci
contribute 31 % to the total number of fire grid cells in the TRMM domain while they contribute only
5 % to the total number of fire counts (Table 11). Also in the other fire products, the contribution of
the missed subset to the total number of fire grid cells is higher than its contribution to the total
number of hotspot counts.
This higher contribution reflects that omission by MERIS_cci is over-proportionally related to smaller
fires. Accordingly, in all hotspot products, the average number of hotspots per fire grid in the “missed”
subset is substantially lower than in the “not-missed” subset (Table 12). The ratio is 7 times lower in
the MODIS-HS product and still 31 to 36 % lower in the TRMM and WFA hotspot products. In
agreement, there is an over-proportional decrease of savannah fires – which are typically large fires in the missed subset whereas tropical deforestation fires – which are typically small fires - overproportionally increase (Table 11).
D4.1.3 Product Intercomparison Report
Page 31
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
Table 12: Total number of time-integrated fire grid cells (NFG) fire pixels in the TRMM domain in the
WFA-HS, TRMM-HS and MODIS-HS products and percentage missed by MERIS_cci. The average
number of hotspots per fire grid, integrated over 2006 to 2008, is shown for the missed and not-missed
subset. Intersecting with MERIS_cci was performed with time-integrated data (period 2006 - 2008).
A clustering of WFA and TRMM hotspots, which are consistently missed by MERIS_cci burned area,
occurs in the southwestern part of the Amazon forest, e.g. around Rio Branco, and in a belt extending
further southwards towards Paraguay-Boqueron (Figure 23) . In these areas, deforestation fires are
frequent. These missed fires, however, are not related to an omission error in the MODIS hotspot
product which is used as a seed to guide the reflectance-based MERIS_cci burned area detection: the
MODIS-HS product indeed shows fire activity in these regions. The omission by MERIS_cci is
therefore most likely related to restrictions in delineating burned from unburned areas in tropical
deforestation fires. Further work will be dedicated to elucidate the reasons why omission by
MERIS_cci prominently occurs in these particular regions.
This preliminary analysis indicates that omission of burned area in MERIS_cci is primarily related to
limitations in detecting small fires; MERIS_cci burned area misses equally 31 % of all fire grid cells
detected by the MODIS and TRMM active fire products, which, however, exhibit low fire activity.
D4.1.3 Product Intercomparison Report
Page 32
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
Figure 23: Mean annual number of active fire pixels per grid cells in the WFA-HS, TRMM-HS and the
MODIS-HS product (2006 - 2008). Shown are a) all fire grid cells (“all”) and b) the fire grid cells with no
co-located fire signal in the MERIS_cci product (“no MERIS”). Intersecting with MERIS_cci was
performed with time-integrated data.
D4.1.3 Product Intercomparison Report
Page 33
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
5.10 Comparison of the spatial patterns of mean annual burned area for
selected regions
As discussed in section 5.4 fire patterns in the different satellite products (burned area and active fire
count) show pronounced regional distinctions. In the following, we analyse in more detail the level of
agreement and disagreement across Australia and Europe.
5.10.1 Australia
Every year, fires in Australia burn across vast areas of forest, savannah, and grass- and croplands.
Most of the fires burn in tropical Australia, i.e. north of 23°S. The main fire season is August to
November.
Figure 24 illustrates that all fire satellite products capture the main burning regions in the north.
MERIS_cci, GFED, GFAS, and WFA-HS also largely agree in the general spatial fire patterns in
Australia's south. Here, fire activity in MERGED_cci (respectively VGT_cci) and TRMM shows a
much greater spatial extent than the former (Figure 26). The differences are also reflected in the
weaker correlation of the zonal fire grid coverage (Figure 26). The level of agreement (R2) between
MERGED_cci and GFED or MERIS_cci _cci is only ~50 % compared to 94 % between MERIS_cci
and GFED.
Between MERIS_cci, GFED and MERGED_cci, there is a very high agreement (R2 ~ 90% and
higher) in the zonal patterns of burned area and fire size (Figure 26). Furthermore, all three products
agree very well with respect to temporal pattern of monthly burned area, when integrated over entire
Australia (Figure 27).
The bias in the spatial extent of fire occurrence in Australia's south might be related to a clustering of
commission errors (false alarms) in the MERGED_cci burned area product (resp. VGT_cci) and in the
TRMM hotspot product. Ji and Stocker (2002) reported on the distinct clustering of false alarms in
southern Australia in the TRMM-HS product. The bias could also point to the underestimation of fires
by the MODIS platform (MPIGfA, 2008). Compared to burned area information compiled from a
combination of ground-based surveys and high-resolution aerial photography by the Australian state
agencies, the MODIS platform tends to substantially underestimate the extent of fire in closed-canopy
forests, which are common in the southern mesic rainfall zone of New South Wales (NSW) and
Victoria (VIC) (MPIGfA, 2008) (Figure 25). Underestimation is generally much less pronounced in
more open forests of Australia, such as tropical savannas (MPIGfA, 2008). Furthermore, the standard
MODIS burned area products for Australia tend to increasingly underestimate small fires outside the
tropical belt, i.e. south of 23.3°S (Randerson et al. 2012).
D4.1.3 Product Intercomparison Report
Page 34
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
Figure 24: Same as Figure 15 but for Australia only. The field percentile refers to the global field
percentile. The grid cell values corresponding to the different percentiles are listed in Table 6.
.
D4.1.3 Product Intercomparison Report
Page 35
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
Figure 25: Left: map of the different states of Australia (NSW: New South Wales; NT: Northern
Territory; QLD: Queensland; SA: South Australia; TAS: Tasmania; VIC: Victoria; WA: Western
Australia); Middle: MCD12 dominant land cover per 0.5-degree grid cell of the year 2005 using the UMD
legend (1=ENF: evergreen needleleaf forest; 2=EBF: evergreen broadleaf forest; 3=DNF: deciduous
needleleaf forest; 4=DBF: deciduous broadleaf forest; 5=MF: mixed forest; 6=CSH: closed shrublands;
7=OSH: open shrublands; 8=WSA: woody savannas; 9=SAV: savannas; 10=GRA: grasslands; 12=CRO:
croplands; grey=OTH: others (merging urban and built-up with barren or sparsely vegetated); Right:
Forest cover of Australia according to the Australia’s State of the Forests Report 2013 (MPIGfA/NFISC,
2013).
Figure 26: Latitudinal pattern of fractional fire grid coverage over land (top), mean annual burned area
(middle), and monthly burned area per fire grid (bottom) in the Australian domain (40 ~ 10° S; 110 ~ 155°
E), see also Figure 24. To the right, the coefficient of determination of the zonal correlation of the product
pairs is given.
D4.1.3 Product Intercomparison Report
Page 36
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
Figure 27: Monthly total area burned in Australia during January 2006 to December 2008 in the
MERIS_cci, GFED, VGT_cci, and MERGED_cci product. The coefficient of determination of the
temporal correlation is given to the right.
The southern Australian states New South Wales (NSW) and Victoria (VICT) suffered from extreme
fires during the 2006/07 (July-June) fire year due to sustained dry conditions. According to
information on the forest burned area compiled by the Australian state agencies, burned area in the
extreme fire year 2006/2007 in NSW and VIC is more than four times larger than in typical fire years
(MPIGfA/NFISC, 2013). The following section is dedicated to a more detailed inter-comparison and
interpretation of the fire products for this extreme fire year.
Table 13: Forest burned area by state and fire year in Australia compiled by the Australian state agencies
(MPIGfA/NFISC, 2013). Burned area in states marked by a star (*) is mapped with high-resolution
Landsat and SPOT satellite imagery, combined with aerial and ground-based survey. For the other states,
burned area is mapped with MODIS burned area imagery (MCD45), only, which are cut by forest cover.
D4.1.3 Product Intercomparison Report
Page 37
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
Table 14: Total burned area per fire year (July-June) and per Australian state calculated from
MERIS_cci, GFED4, GFED3, VGT_cci and MERGED_cci. The right column shows the forested burned
area estimated by the Australian authorities.
Most of the forested areas occur in the tropical or subtropical climate zones of Australia. A direct
comparison of the official burned area in forests and the total burned area calculated from the fire
satellite products (Table 14) is limited because of missing information on how the Australian
authorities discriminate between forested and non-forested burned area (MPIGfA/NFISC, 2013). Fires
in forested areas contribute around 20% to the total area burned in Australia (Table 13). Most of the
fire occur in areas with less dense tree cover (i.e. savanna) or with no tree cover (grass-, crop-, or
shrublands). Around 16 % of Australia's land area is forested according to the official Australian forest
inventory (MPIGfA/NFISC, 2013). Most of the forested areas occur in the tropical or subtropical
climate zones of Australia.
The comparison by states highlights that GFED and MERIS_cci underestimates total burned area in
NSW, VICT, and southern Australia (SA) (Table 14):
During the extreme fire year in VIC 2006/07, GFEDv4, and MERIS_cci burned area across all land
cover types is 30 % and 40 % lower, respectively, than the official statistics for fires in forested land,
only. MERGED_cci total burned area, in contrast, is 6 % higher than the official statistics whereas
VGT_cci is 30 % lower. In NSW, GFEDv4 yields a 10 % lower estimate while MERIS_cci is 10 %
higher. Here, MERGED_cci total burned area (and similarly VGT_cci) is by a factor of ~7 higher than
the official burned area in forests. Finally, in SA, GFED and MERIS_cci are even 40 % and 80 %
lower.
Underestimation by GFED and MERIS_cci in these three states is also present during in the fire year
2007/08, when fires were less extreme. Here, GFED and MERIS_cci total burned area is between 60
% and 20 % lower than the officially estimated area of forest burned. MERGED_cci (resp. VGT_cci),
in contrast, yields distinctively higher burned area than the official estimates (Table 14).
A comparison focusing on the Great Divide Fire Complex in Victoria (Figure 28 and Figure 29)
confirms the underestimation of burned area by GFED and MERIS_cci (by around 30 to 40 %,
respectively) with respect to the official forest burned statistics. The Great Divide Fire Complex
burned between 11,800 to 12,600 km2 (Smith, 2007; VdoSE, 2008) within a period of seven weeks
starting from early December 2006. Also the MERGED_cci product underestimates the area extent of
the Great Divide Fires, but only by ~ 25 %. Underestimation is most pronounced (by a factor of 6) in
the VGT_cci product.
D4.1.3 Product Intercomparison Report
Page 38
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
Underestimation could partially be related to topographic effects: The Great Divide Fires occurred in
mountainous areas (Figure 29) and satellite detection of burned area may have been impaired in valley
regions. Figure 29 illustrates that the VGT_cci product does not capture the large burned areas (with
values well above 600 km² per grid) of the Great Divide Fire Complex. In contrast, the product shows
relatively small fires (burned area of 200 km² per grid and lower) across the entire domain. The
occurrence of fires across the entire domain is not supported by other satellite fire products and points
to substantial commission errors in the VGT_cci product. The MERGED_cci product, due to the
largely additive nature of merging MERIS_cci and VGT_cci, incorporates the commission errors of
the VGT_cci product. This also explains why total burned area in the entire domain displayed in
Figure 28 is only 2 % lower in the MERGED_cci product than in the sum of the MERIS_cci and the
VGT_cci product.
In order to elucidate the discrepancies in the burned area estimates for NSW, VICT, and SA, we plot
the monthly fire seasonality over the July 2006 to June 2007 fire season (Figure 30). The evolution of
the spatial pattern highlights drastic differences between MERIS_cci and GFED on the one hand, and
MERGED_cci (resp. VGT_cci) on the other hand. The figure shows the strong clustering of large fires
in NSW and VICT during March to May 2007 in the MERGED_cci (resp. VGT_cci) product, whereas
there is nearly no fires activity in the same region and period in the MERIS_cci or GFED product or
other independent fire products. In April and May, there is also pronounced fire activity in SA in the
MERGED_cci (resp. VGT_cci) product, which is not supported by other fire products.
Figure 28: Fire affected area (in red) during the Great Divide Fire Complex 2006/07 in Victoria,
Australia, mapped with high-resolution Landsat and SPOT satellite imagery (VdoSE, 2008). The map
spans from approximately 145.2° to 149.5° E and 38.1° to 36.1° S. The official burned area estimates range
from 11,800 to 12,600 km² (Smith, 2007; VdoSE, 2008)
D4.1.3 Product Intercomparison Report
Page 39
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
Figure 29: Fire affected area in km2 per grid summed over the period August 2006 to March 2007 in
Victoria, Australia, mapped by MERIS_cci, GFED, VGT_cci and MERGED_cci. The maps span from
145.5° to 149.5° and 38.5° to 36° S. Burned area estimates for this domain calculated from the fire satellite
products over this period are: MERIS_cci : 7,460 km² , GFED4: 8,937 km²; GFED3: 8,100 km²; VGT_cci:
2,068 km², and MERGED_cci: 9,356 km². The bottom left figure shows the topographic height in metres
at 0.1 degree resolution (derived from the GTOPO30 product)
D4.1.3 Product Intercomparison Report
Page 40
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
Figure 30: Monthly burned area per 0.5-degree grid cell in the different burned area products over the
fire year July 2006 to June 2007 across Australia (top to bottom). Monthly total GFAS FRE, AATSR and
TRMM hotspot number is shown as well. Please note that spatiotemporal pattern of NOAA burned area
is similar to MERIS_cci (see also
http://www.firewatch.landgate.wa.gov.au/landgate_firewatch_public.asp).
D4.1.3 Product Intercomparison Report
Page 41
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
Figure 31: Monthly total area burned in the Australian states Victoria (VICT) and New South Wales
(NSW) during January 2006 to December 2008 in the MERIS_cci, GFED, and MERGED_cci product.
The coefficient of determination of the temporal correlation is given to the right
Visual inspection of aerial photography publicly available via virtual globe machines aids to reveal
potential causes for very high burned area grid cells in May 2007 in NSW. (e.g. by selecting cells with
monthly burned area above 300 km2). These grid cells appear highly agricultural regions characterised
by a sequence of rectangular, around 1 to 5 km2 sized plots. It is unrealistic that fires would actually
spread across more than 300 km2 in a single month in such a fragmented landscape with low fuel
loads. It is rather likely that the MERGED_cci product, more specifically the VGT algorithm, tends to
interpret non-fire related changes in the reflective surface properties of agricultural land as burned
area.
The Australian Grain Belt covers a large portion of southern Australia. In the southeast, it expands
from western Victoria to southern Queensland crossing the center of NSW. In the southwest, it forms
an arc around most of the Perth metropolitan area. Wheat, barley and oat are the major winter crops
grown throughout this belt. Typically, these winter crops are typically planted in during April to July
and harvested in November to January (Sacks et al. 2010).
The intercomparison of monthly burned area intergated over Victoria and NSW (Figure 31) shows that
strong discrepancies between MERGED_cci (resp. VGT_cci) on the one hand and GFED and MERIS
on the other hand occur notably during March to May, which is the planting season of winter crops. In
these months, the MERGED_cci (resp. VGT_cci) product shows unrealistically numerous and large
fires. As mentioned above, probably spectral changes caused by the soil preparation required for crop
planting are confused with burned area in the current VGT_cci version.
5.10.2 Europe
Differences between the individual fire satellite products are also visible across the land areas of
Europe (Figure 32). The most striking is that almost all land grid cells in the MERGED_cci product
(i.e. MERIS + VGT) show fire activity: In the MERGED_cci product, fires occur in around 80 % of
all land grid cells in the domain shown in Figure 33. The area coverage is around 55 % in GFAS and
close to 35 % in both MERIS_cci and GFED. The coverage by night time fires detected by AATSR is
around 30 %.
D4.1.3 Product Intercomparison Report
Page 42
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
The difference in fire grid coverage between the MERGED_cci and the other products is most
pronounced at higher latitudes, notably above 56° N and 66° N (Figure 33), where fire coverage is 90
% and higher in MERGED_cci, but less than 30 % in the other products.
Across all latitudinal bands, not only is the number of fire grid cells in the MERGED_cci product
higher, but also burned area is constantly higher than in the other products. Here again, the highest
discrepancies occur at higher latitudes, notably above 56° N (Figure 33). All products have a zonal
maximum at 49° N with mean annual burned area between 19,000 (MERIS_cci ) and 25,000 km2
(MERGED) and burned area close to zero south of 34° N and north of 66° N. Towards both sides of
the peak at 49° N, burned area decreases much more rapidly in MERIS_cci and GFED than in
MERGED_cci. In fact, the MERGED_cci product exhibits a second distinct zonal maximum of
19,000 km2 at 59° N which is not supported by the very low fire activity in the other products.
The zonal peak in fire size (expressed as mean monthly burned area per fire grid) coincides with the
peak in zonally integrated burned area at 49° N. Here, however, fire size in GFED is highest in GFED
(43 km2) and lowest in MERGED_cci (25 km2), with MERIS_cci laying in between (32 km2). In
contrast to GFED, MERIS_cci and MERGED_cci have a second zonal peak of almost the same
magnitude as their zonal maximum. In MERIS_cci, the peak spikes at 56° N. In MERGED_cci, the
peak is 3° more northward and less spiky than in MERIS_cci. The strong positive bias in zonally
integrated burned area above 56° N in MERGED_cci is therefore the combined result of a much larger
number of fire grid cells and a much larger fire size.
The strong positive bias in zonally integrated burned area in the MERGED_cci product at higher
latitudes (i.e. north of 56° N) is to 85 % attributed to the months March and April, i.e. to months
outside the typical fire season. This period, however, is the typical planting season of spring crops
(spring barley, summer wheat, corn, sunflower, sugar beets) which are widely cultivated in northerly
Europe (Sacks et al. 2010). Hence, a substantial number of the fire grid cells in the MERGED_cci
product, which are detected in addition to MERIS_cci at higher latitudes, are most likely false alarms.
As mentioned above, the current VGT_cci version used in the MERGED_cci product tends to
misclassify surface reflectance changes induced by harvesting and/or soil preparation for planting as
burned area.
The spatial pattern of mean annual burned area in MERIS_cci across Europe largely resembles the
pattern in GFED. MERIS_cci also produces similar patterns as in the active fire products. This good
agreement between MERIS_cci and GFED on the one hand, and MERIS_cci and active fire products
is also reflected in the very high correlation coefficients of the zonal gradients plotted Figure 33. In
addition, GFED and MERIS_cci also agree well in their temporal patterns of domain-integrated
burned area (Figure 34). The zonal, and temporal agreement between MERGED_cci and GFED, in
contrast, is distinctively lower (Figure 33 and Figure 34). Here, only between 32 and 49 % of the
variability in MERGED_cci can be explained by GFED.
A few exceptions in the general good agreement between MERIS_cci and GFED are present in
Western Russia, namely the wider region around Moscow (55° N, 37° E), but most notably in the
south. In this area, there is comparatively little fire occurrence in GFED compared to MERIS_cci,
MERGED_cci or GFAS. Abnormally widespread cropland and forest fires occurred in this region
from mid-April to mid-May 2006 as a result of unprecedented dry and warm spring conditions (Stohl
et al. 2007). The fires resulted in record high air pollution levels in the European Arctic (e.g. Stohl et
al. 2007). While these fires are well represented in MERIS_cci, MERGED_cci, GFAS and MODIS
hotspots, they are not captured by GFED (Figure 33). The MERGED_cci product shows additional,
very intense fires north of around 57° N in May 2006, and to a lesser extent in April 2006. Neither
other fire satellite products nor atmospheric observations of trace species provide evidence for the
occurrence of these fires. As mentioned above, the VGT_cci product, which is the second component
contained in MERGED_cci product besides MERIS_cci, exhibits high commission errors related with
agricultural activities (harvesting, soil preparation for planting). The regions with the highest
discrepancies in fire intensity between MERIS_cci and MERGED_cci are highly agriculturalized
regions where spring crop planting is common during April and May (Sacks et al. 2010).
D4.1.3 Product Intercomparison Report
Page 43
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
Figure 32: Same as Figure 15 but for Europe only. The field percentile refers to the global field percentile.
The grid cell values corresponding to the different percentiles are listed in Figure 15: Field percentiles of
mean annual i) burned area (MERIS, GFEDv4, VGT, and MERGED; ii) FRP (GFAS) and iii) hotspot
(HS) counts (MODIS, WFA and TRMM). The total number of valid fire grid cells and the percentile
statistics of each product are given in Table 6. Fire signals in the Southern Pacific in the WFA-HS product
are created by an artefact in the original August 2008 data
No TRMM fire data are available for the domain north of 38° N.
D4.1.3 Product Intercomparison Report
Page 44
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
Figure 33: Latitudinal pattern of fractional fire grid coverage over land (top), and mean annual burned
area (2006-2008) (middle), and monthly burned area per fire grid (bottom) in the European domain (30 ~
70° N; 10° W~70° E). To the right, the coefficient of determination of the correlation of the product pairs
is given.
Figure 34: Time series of monthly burned area in the European domain (30~70° N; 10° W~70° E) from
January 2006 to December 2008. The coefficient of determination (R²) of the pairwise correlation of the
time series is given to the right.
D4.1.3 Product Intercomparison Report
Page 45
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
Figure 35: Burned area in Eastern Europe in April and May 2006 in MERIS_cci, GFED3, GFED4 and
MERGED_cci, complemented by monthly GFAS FRE. MODIS fire detections for April 21 to May 5,
2006, colour-coded by vegetation type is shown at the bottom.
Figure 35 highlights, as an example, the benefits of the availability of multiple satellite products for
detecting, characterizing and quantifying fire occurrence. Cross-comparing multiple satellite fire
product is a means to validate, assess and possibly improve individual fire products. It is also a means
to facilitate the decision to be made by the climate research groups which fire satellite product to use
as input for a given research question. Each product exhibits different commission and omission
characteristics dependent upon the region and period selected. In particular case of the extreme spring
2006 fire event in Eastern Europe, the MERIS_cci product appears to capture burned area much more
realistically than GFED. Model studies focusing on this particular event could therefore achieve a
better predictability, e.g. of atmospheric trace species with respect to measurements, when making use
of the new MERIS_cci burned area product as model input instead of GFED.
Fire occurrence in the MERGED_cci product also appears unrealistically widespread and of too high
intensity across north western Russia (Vologda Oblast and surroundings), Finland, Sweden, and the
Alps. The widespread occurrence of high intensity fires in these regions is neither supported by other
satellite fire products or atmospheric observations. The timing of these fires is typically too early in
the fire season, but agrees well with typical timing of crop planting or harvesting activities. On
average, 4,100 km2 of burned area per year is detected by MERGED_cci in the Alps. This is 33 and
333 times higher than in MERIS_cci and GFED, respectively. Most of the fires in the MERGED_cci
product occur between July and October. Intense fires even during or at the end of the summer, are
unlikely at high altitudes of the Alps (above 2000 above sea level) because of low fuel loads and low
fuel continuity which impede the spread of fires. It is rather likely that the MERGED_cci burned area here specifically the VGT_cci component - is spectrally confused with the thawing of the seasonal
snow cover at high topography, which predominantly takes place during June to October in the Alps.
An analysis of the pixel level product will provide more concise explanations why and under which
conditions these suspected false alarms occur in the MERGED_cci product.
D4.1.3 Product Intercomparison Report
Page 46
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
6 Conclusions
The general conclusion in terms of the data quality for different fire_cci products is that the
MERIS_cci product is in good agreement with GFED (v3 and v4) burned area (which is currently
considered by most specialists as the most reliable estimation of global BA), both in terms of the total
amount of burned area each year and the spatial and temporal distribution of fires. The BA estimations
of MERIS_cci are approximately 200,000 km2 (< 10 %) higher than GFED estimations in all three
years. Regional differences between the two were more relevant in relative terms. Higher estimations
than GFED are observed in both hemispheres of Africa and Central Asia, Central and South East Asia
and the Southern hemisphere of South America, while lower estimations were found in Australia.
The two other fire_cci products showed much higher divergences with GFED than MERIS_cci, with
much lower estimations for VGT_cci (> 30 % lower than GFED) and much higher estimations for
MERGED_cci (> 30 % over GFED). The VGT_cci product -and hence also the MERGED_cci
product - still have major problems due to a) the omission of small and large actual fires and b)
commission errors due to the confusion of spectral changes related to agricultural activities (soil
preparation/harvesting) with burned area. As the MERGED_cci product adds up both MERIS_cci and
VGT_cci, and there are few spatial coincidences between the two products, the final MERGED_cci
was not yet found suitable for being use by climate modellers. In regional terms, VGT_cci greatly
overestimates the GFED BA in Boreal Asia and North America, and with less relevance in Central
Asia, while greatly underestimates in the Tropical regions. The MERGED_cci BA estimations are in
between MERIS and VGT_cci products, with overestimation in all regions but Australia, with higher
relative increments in the Boreal Regions of Asia and North America.
In terms of seasonal trends, both the MERIS_cci and MERGED_cci products are similar to GFED,
with two seasonal peaks around the dry season of both Tropical belts (December-February for the
Northern Hemisphere and July-September for the Southern Hemisphere). Regional trends of BA
seasonality in Australia and Europe between MERIS_cci and GFED were shown very consistent,
again proving the interest of the MERIS_cci product for estimating BA temporal tendencies. Satellite
observation of fire seasonality is one key parameter climate users are interested. Observations of fire
seasonality are not only required to realistically parameterize the climate – fire dynamics in prognostic
fire models, but also to test model performance. Taking Australia and Europe as an example, we
demonstrated that new MERIS_cci BA product can realistically map fire seasonality, in some regions
even more realistically than GFED. This product appears already now sufficiently sound to be
recommended for uses by climate researchers.
D4.1.3 Product Intercomparison Report
Page 47
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
7 References
Alexander, M. E.: Calculating and interpreting forest fire intensities (1982), Canadian Journal of
Botany 61: 349-357.
Alonso, I., & Chuvieco, E. (2014). Global Burned Area Mapping from ENVISAT-MERIS data
Remote Sensing of Environment, (submitted).
Archibal<d, S., Lehmann, C. E. R., Gómez-Dans, J. L. and Bradstock, R. A (2013): Defining pyromes
and global syndromes of fire regimes, P. Natl. Acad. Sci. USA, 110, 6442–6447, 2013.
Arino, O. and Melinotte, J.: The 1993 Africa fire map, Int. J. Remote Sens., 19, 2019–2023, 1998.
Arino, O., Bicheron, P., Achard, F., Latham, J.,Witt, R., and Weber, J. L.: GLOBCOVER The most
detailed portrait of Earth, ESA Bulletin-European Space Agency, 136, 24–31, 2008.
Arino, O., Casadio, S., and Serpe, D.: Global night-time fire season timing and fire count trends using
the ATSR instrument series, Remote Sens. Environ., 116, 226–238, doi:10.1016/j.rse.2011.05.025,
2011.
Giglio, L., Csiszar, I., & Justice, C.O. (2006). Global distribution and seasonality of active fires as
observed with the Terra and Aqua Moderate Resolution Imaging Spectroradiometer (MODIS) sensors.
Journal of Geophysical Research-Biogeosciences, 111, doi:10.1029/2005JG000142.
Giglio, L., Loboda, T., Roy, D.P., Quayle, B., & Justice, C.O. (2009). An active-fire based burned area
mapping algorithm for the MODIS sensor. Remote Sensing of Environment, 113, 408-420.
Giglio, L., Randerson, J.T., van der Werf, G.R., Kasibhatla, P.S., Collatz, G.J., Morton, D.C., &
DeFries, R.S. (2010). Assessing variability and long-term trends in burned area by merging multiple
satellite fire products. Biogeosciences Discuss., 7, 1171-1186, doi:1110.5194/bg-1177-1171-2010,.
Giglio, L., Randerson, J.T., & Werf, G.R. (2013). Analysis of daily, monthly, and annual burned area
using the fourth‐generation global fire emissions database (GFED4). Journal of Geophysical
Research: Biogeosciences, 118, 317-328.
Ji, Y., & Stocker, E. (2002): Seasonal, intraseasonal, and interannual variability of global land fires
and their effects on atmospheric aerosol distribution, J. Geophys. Res., 107(D23), 4697,
doi:10.1029/2002JD002331.
Kaiser, J. W., Heil, A., Andreae, M. O., Benedetti, A., Chubarova, N., Jones, L., Morcrette, J.-J.,
Razinger, M., Schultz, M. G., Suttie, M., & van der Werf, G. R. (2012): Biomass burning emissions
estimated with a global fire assimilation system based on observed fire radiative power,
Biogeosciences, 9, 527–554, doi:10.5194/bg-9-527-2012.
MPIGfA (Montreal Process Implementation Group for Australia) (2008): Australia’s State of the
Forests
Report
2008,
Bureau
of
Rural
Sciences,
Canberra,.
http://data.daff.gov.au/data/warehouse/pe_brs90000003841/sofr2008Full_11a.pdf.
MPIGfA/NFISC (Montreal Process Implementation Group for Australia and National Forest Inventory
Steering Committee) (2013): Australia’s State of the Forests Report 2013, ABARES, Canberra,
http://www.daff.gov.au/ABARES/forestsaustralia/Documents/sofr2013-web2.pdf.
Mouillot, F., Schultz, M.G., Yue, C., Cadule, P., Tansey, K., Ciais, P., & Chuvieco, E. (2014). Ten
years of global burned area products from spaceborne remote sensing—A review: Analysis of user
needs and recommendations for future developments. International Journal of Applied Earth
Observation and Geoinformation, 26, 64-79.
Oliva Pavón, P. & Chuvieco, E. (2013): Assessment of the discrimination ability of MERIS spectral
data for burned area mapping using ROC curves, GeoFocus 13-2, 41-65.
Padilla, M., & Chuvieco, E. (2014). ESA CCI ECV Fire Disturbance, Product Validation Report II,
Fire_cci_Ph3_UAH_D4.1.2_PVRII_v1_3.pdf, (available on https://www.esa-fire-cci.org/)
D4.1.3 Product Intercomparison Report
Page 48
fire_cci
Doc. No.:Fire_cci_Ph3_UAH_D4_1_3_PIR_v1_4
Issue/Rev-No.: 1.4
Pereira, J.M., Mota, B., Calado, T., Alonso, I.J., Oliva, P., & González-Alonso, F. (2014). ESA CCI
ECV Fire Disturbance, Algorithm Theoretical Basis Document – Volume II,
Fire_cci_Ph3_ISA_D3_6_2_ATBD_II_v2_2.pdf, (available on https://www.esa-fire-cci.org/)
Randerson, J. T., Chen, Y., Werf, G. R. van der, Rogers, B. M. & Morton, D. C. (2012): Global
burned area and biomass burning emissions from small fires, J. Geophys. Res. Biogeosci., 117,
G04012, doi:10.1029/2012JG002128.
Roy, D., Jin, Y., Lewis, P., & Justice, C. (2005). Prototyping a global algorithm for systematic fireaffected area mapping using MODIS time series data. Remote sensing of environment, 97, 137-162.
Roy, D., Lewis, P., & Justice, C. (2002). Burned area mapping using multi-temporal moderate spatial
resolution data—A bi-directional reflectance model-based expectation approach. Remote sensing of
environment, 83, 263-286.
Roy, D.P., Boschetti, L., & Justice, C. (2006). Global mapping of fire-affected areas using
multitemporal MODIS data: The MCD45 product. In, 2006 IEEE International Geoscience and
Remote Sensing Symposium, Vols. 1 (pp. 4165-4168).
Sacks, W.J., Deryng, D., Foley, J.A. & Ramankutty, N. (2010): Crop planting dates: An analysis of
global patterns. Global Ecology and Biogeography, 19: 607-620.
Schultz, M. G.: On the use of ATSR fire count data to estimate the seasonal and interannual variability
of vegetation fire emissions, Atmos. Chem. Phys., 2, 387-395, doi:10.5194/acp-2-387-2002, 2002.
Smith, R. (2007): Key issues identified from operational reviews of major fires in Victoria 2006/07,
Dept of Sustainability & Environment Victoria and Country Fire Authority, Victoria, Australia,
http://www.depi.vic.gov.au/__data/assets/pdf_file/0004/198535/Report_Ross_Smith_Op_Review
_2006-07_fire_seasonv2.pdf, ISBN9781741529654.
Stohl, A., Berg, T., Burkhart, J. F., Fjǽraa, A. M., Forster, C., Herber, A., Hov, Ø., Lunder, C.,
McMillan, W. W., Oltmans, S., Shiobara, M., Simpson, D., Solberg, S., Stebel, K., Ström, J., Tørseth,
K., Treffeisen, R., Virkkunen, K., & Yttri, K. E.(2007): Arctic smoke – record high air pollution levels
in the European Arctic due to agricultural fires in Eastern Europe in spring 2006, Atmos. Chem. Phys.,
7, 511-534, doi:10.5194/acp-7-511-2007.
Tansey, K., Bradley, A., & Padilla, M. (2014). ESA CCI ECV Fire Disturbance, Algorithm
Theoretical
Basis
Document
Volume
III
BA
Merging:
Fire_cci_Ph3_UL_D3_6_3_ATBD_III_v2_3.pdf, (available on https://www.esa-fire-cci.org/)
Tansey, K., Grégoire, J.M., Defourny, P., Leigh, R., Peckel, J.F., Bogaert, E.V., & Bartholome, J.E.
(2008). A new, global, multi-annual (2000–2007) burnt area product at 1 km resolution. Geophysical
Research Letters, 35, L01401, doi:10.1029/2007GL03156.
VDoSE (Victoria. Dept. of Sustainability and Environment) (2008): Great divide fire recovery plan,
Victorian Government Department of Sustainability and Environment & Parks Victoria, Melbourne,
Australia,
http://www.depi.vic.gov.au/__data/assets/pdf_file/0004/192946/Great-Divide-firerecovery-plan.pdf, ISBN9781742082738.
D4.1.3 Product Intercomparison Report
Page 49