Development of indicators of fire severity based on time series of SPOT VGT data

1
Development of indicators of fire severity
based on time series of SPOT VGT data

• Stefaan Lhermitte, Jan van Aardt, Pol Coppin
• Department Biosystems
• Modeling, monitoring, and management of
  bioresponse
• Geomatics group KU Leuven Belgium

2
Outline

• Global burn datasets
• GBA2000 (SPOT VGT data)
• Globscar (AATSR)
• Fire detection vs. quantification of impacts
• General constants/biome
• Essential
• Global and regional carbon models
• Understanding of vegetation recovery

3
Objective

• Development of indicators to quantify
  spatio-temporal variation of fire impacts
• Fire Severity (FS)
• Percentage of the biomass per pixel that is
  burned

4
Data

• Study area South Africa
• Satellite Data SPOT Vegetation S10
• Year 2000
• 10-daily Maximum Value Composites (B, R, NIR,
  SWIR, NDVI)
• 1x1 km²
• Fires GBA2000 Burnt Areas
• Year 2000
• Monthly detected fire scars (no exact date, only
  month)
• 1x1 km²

5
Fires

• Indication of fire frequency by area
• Very large fires exist, indicating a possible
  exaggeration

Multitemp
6
Fires byvegetation type
Large areas in (i) forest and woodland, (ii)
thicket and bushland, (iii) shrubland and fynbos,
(iv) unimproved grassland, and (v) cultivated
commercial dryland
Multitemp
7
Techniques

• Spectral mixture analysis (SMA)
• Bare soil, charcoal, vegetation
• Hypothesis
• FSi ?(vegetation fraction)i
• where i fire pixel
• --gt Absolute values
• Changes in vegetation indices (?VI)
• Hypothesis
• FSi ?(vegetation index)i
• where i fire pixel
• --gt Relative values

8
Spectral Mixture Analyis

• Assumes that the reflectance spectrum can be
  deconvolved into a linear mixture of the spectra
  of endmembers (pure pixels)

9
Spectral Mixture Analyis

• Assumes that the reflectance spectrum can be
  deconvolved into a linear mixture of the spectra
  of endmembers (pure pixels)
• Result
• relative abundance (fractions) of different
  endmembers for every pixel
• when only 1 vegetation endmember is chosen, the
  fractions reflect an absolute measure of
• FSi be expressed by ?(VFi)

10
Spectral Mixture Analyis Procedure

• Endmember selection
• Iterative Error Analysis (IEA)
• (Neville et al., 1999)
• An automated selection procedure
• Grouping of all burnt pixels
• 3 observations before fire
• 3 observations afterwards
• Selection of desired endmembers
• Assumption that endmember spectra are time
  invariant
• Only correct estimation for the
• Forest and Woodland Type

11
Spectral Mixture Analyis Procedure

• Typical reflectance of 3 endmembers Vegetation,
  dark wet soil (or charcoal), and light or dry
  soil
• Problem IEA could only retrieve meaningfull
  endmembers for the Forest and Woodland landcover
  type

Multitemp
12
Spectral Mixture Analyis Procedure

• Endmember selection
• Fraction images

13
Spectral Mixture Analyis Procedure

• Vegetation Component3 decades before fire

Multitemp
14
Spectral Mixture Analyis Procedure

• Vegetation Component3 decades before fire

Multitemp
15
?Vegetation Index

• Assumes that the FS can be expressed by ?(VI)
• VI
• no absolute measure of vegetation quantity
• related to vegetation but have phenological
  fluctuations
• cannot be used for FS without normalization

16
?Vegetation Index

• Normalization
• use relative index (RI) to reduce phenological
  influences
• reference areas areas located adjacent or close
  to the burned sites, but not affected by the
  disturbance. They should have similar
  environmental conditions and vegetation

17
Analysis

• Look at a fire as a complete entity
• Analysis of mean(FSi)j where i fire
  pixel j fire id
• Look at spatial variability for every fire
• Analysis of FSi where i fire pixel

18
Spectral Mixture Analyis(fire.id)

• Change curves of fractions for every fire scar
• (Example 1)

Multitemp
19
Spectral Mixture Analyis(fire.id)

• Change curves of fractions for every fire scar
• (Example 2)

Multitemp
20
?Vegetation Index(fire.id)

• Change
• curves of ?VI
• for every fire
• Scar
• (Example 1)

Multitemp
21
?Vegetation Index(fire.id)

• Change
• curves
• of ?VI for
• every
• fire scar
• (Example 2)

Multitemp
22
SMA(fire.id) and?VI(fire.id) (Example 1)
23
SMA(fire.id) and?VI(fire.id) (Example 2)
24
SMA Spatial variability of every fire
Dark soil
Vegetation
Light soil
?VI
25
?VI Spatial variability of every fire
26
Actual Fire Severity

• FS can now be derived from change detection of
  the derived data sets
• Change detection on the RI-images before and
  after fire
• Change detection on the fraction images of the
  vegetation component before and after fire
• E.g. Image differencing was performed and the FS
  was calculated for both techniques

27
Validation

• Fire records of Kruger National Park (KNP)
• Validation of FS with field data containing burn
  severity
• Statistical regression techniques to assess the
  performance of both techniques and the resulting
  quantitative indicators of burning efficiency
• Results were unsatisfactory
• Possible errors endmembers, reference areas
• KNP fire records are very subjective
• Additional validation is necessary
• severity indices Landsat imagery

28
Conclusion

• Two techniques to quantify spatio-temporal
  variation of the impact of fire were presented
• Additional validation is necessary

29
Acknowledgements

• Funding provided by the Belgium Science Policy
  Office (BELSPO) as part of the GLOVEG project
• Jan Verbesselt for scientific inputs

30
stefaan.lhermitte_at_biw.kuleuven.be Laboratory of
Geomatics KU LeuvenVital Decosterstraat 102,
3000 Leuven Belgium