Multitemporal and angular information

1
Multitemporal and angular information

• Monitoring of biomass burning
• using NASA MODIS algorithm developed byRoy and
  Lewis
• consideration of spectral, angular, and
  multitemporal information
• summary of talk by Roy Lewis

David P. Roy, University of Maryland, Department
of Geography, NASA Goddard Space Flight Center,
Code 922,Greenbelt, MD 20771, USA
droy_at_kratmos.gsfc.nasa.gov Philip Lewis,
University College London, Remote Sensing Unit,
26 Bedford Way, London, WC1H OAP,
U.K plewis_at_geog.ucl.ac.uk AGU - 2nd July 2000,
Washington DC
2
Burned area mapping rationale

• Global change research
• estimates of trace gas and particulate emissions
  (CO2, CO, NOx, CH4, aerosols) for climate
  modeling and for IPCC/FCCC
• Biomass burned (g) burned area (m2) fuel
  load (g/m2) completeness of combustion
• Emission of gas(g/Kg fuel consumed) biomass
  burned (g) emission factor (g/Kg)
• these vary spatially and temporally as fuel
  mixture changes
• understanding biophysical processes, particularly
  loss of biomass and release of carbon and
  greenhouse gasses to the atmosphere - carbon and
  nitrogen cycles - biogeochemical and ecosystem
  modeling
• fire frequency can be expected to change with
  climate change and variability
• Fire is a proximate cause / indicator of land
  cover change
• fire is an important ecological disturbance
  regime
• fire is a major land management practice in
  tropical systems
• fire frequency will change with population
  dynamics
• Fire can be a natural hazard with large societal
  costs and impacts
• economic damage, air quality, health
• early warning, management and long term
  monitoring of wildfires

3
Overview

• active fire detection at regional to global
  scales using moderate spatial resolution
  satellite data has considerable heritage (e.g.,
  AVHRR, ATSR, GOES, DMSP with MODIS active fire
  products to be released August 2000).
• active fire detection cannot be used as a
  reliable surrogate for burned area as the
  majority of fires are not detected at the time of
  satellite overpass
• Goal develop burned area algorithm applicable to
  moderate spatial resolution polar orbiting
  satellite sensing systems
• Prototyping using 1Km AVHRR
• MODIS implementation under testing
• Evaluation validation using Landsat ETM focused
  on Southern Africa (SAFARI 2000)

4
An active burn near the Okavango Delta, Botswana

• NOAA-11 AVHRR LAC data (1.1km pixels)
• September 1989.
• Red indicates the positions of active fires
• NDVI provides poor burned/unburned discrimination
• Smoke plumes gt500km long

5
An active burn Mongu, Zambia

• Landsat ETM (30m pixels, RGB wavelengths)
• August 1999
• Evident variation of burned scar coloring
  (related to scar age, fire intensity, type of
  material burned etc.)
• Bright white Kalahari sand revealed as vegetation
  is burned off

6
Photograph of burned vegetation

• patch 2m wide
• partial burning differing degrees of combustion
  completeness
• Brown dry unburned vegetation
• Green unburned vegetation
• Black char
• White ash

7
Burned areas

• Characterized by
• the removal of vegetation and alteration of its
  structure
• deposits of charcoal and ash
• exposure of the soil layer
• These vary temporally and spatially because of
• the type of vegetation that burns (landcover)
• the completeness of the burn
• the rate of charcoal and ash dissipation by the
  elements
• the post fire evolution and revegetation of the
  burned area

8
Burned area mapping should be treated as a change
detection problem

• Burning is a temporal phenomena?use
  multi-temporal data
• Select wavelength(s) sensitive to changes in the
  burn signal and insensitive to atmospheric
  contamination (e.g. smoke)
• Threshold temporal changes in reflectance for
  each pixel rather than setting a fixed threshold
  for all pixels in a single data set (threshold
  against the magnitude and direction of change)
• Central Issue
• How to define the the magnitude of spectral
  change associated with vegetation to burned
  vegetation conversion, given sensitivity to
• spatial and temporal variations of burned areas,
• variations in the proportion of the pixels that
  are burned,
• external variations imposed by the remote sensing
  (residual atmospheric and cloud contamination,
  BRDF)
• How to do this globally and independent of the
  time of year ?

9
Approach BRDF

• Bidirectional Reflectance Distribution Function
  (BRDF) describes how reflectance depends on the
  view and solar angles.
• Non-Lambertian (BRDF) effects significant for
  vegetation monitoring at regional to global
  scales.
• Source of noise if not accounted for
• BRDF effects may be modelled
• invert BRDF model against satellite observations
  and use model parameters to give normalised
  values at fixed view and solar angles.
• BRDF model parameters may be used to predict
  satellite observation values through time.
• Large discrepancies between the predicted and
  measured values may be attributed to
• change or
• residual cloud and atmospheric contamination.

10
Cloud-cleared AVHRR time series - HAPEX-Sahel 1992
11
Kernel-based BRDF models

• can describe BRDF with simple kernels
• linear model
• kernels based on abstraction of physically-based
  models operate on reflectance
• 3x3 matrix inversion (fast)
• implicit modelling of sub-pixel heterogeneity
  i.e., can model different cover types within the
  pixel (e.g., partially burned pixels)
• can predict uncertainty in model parameters or
  linear combinations
• used to examine influence of MODIS/MISR sampling
  patterns (Lucht Lewis, IJRS, 2000)

12
(No Transcript)
13
Multitemporal BRDF-based Change Detection

• Model BRDF over moving window (T-NW to T)
• Predict BRDF of next observation (T1)
• Predict uncertainty in model result
• Produce Z-score between actual and predicted
  observation to detect probability of CHANGE
• I.e., Z-score predicted - observed reflectance
  / model uncertainty
• Threshold Z-score time series

14
Data

• 54 AVHRR LAC (1km) orbits August 1- August 31
  1997
• Pre-processing by state of the art Pathfinder II
  code
• Use the reflective component of the
  middle-infrared (r3)
• sensitive to the presence of liquid water
  (vegetation and soil water content is reduced by
  burning)
• less sensitive to scattering by smoke aerosols
  than shorter wavelengths
• discriminates between burned and unburned
  surfaces
• AVHRR channel 3 (3.55-3.93 ?m)

15
Image ofAll of Southern Africa single day in
August

• Points
• 30013001 1km pixels
• BRDF effects apparent across orbit edges
• Gaps where cloudy and poor R3 retrieval data
  discarded
• Huge burn 200km south of Okavago delta in
  Botswana

16
Algorithm

• 10-day moving window
• RossThick-LiSparseReciprocal model inversion for
  6 or more observations
• Change candidate if at least one subsequent
  observation over the next 3 days pass and none
  fail
• Pass Z-score gt Zthresh
• Fail Z-score lt Zthresh
• Final change candidate is the one with largest
  Z-score gt Zthresh

17
Illustration of a single step of the moving
window Top left R3 predicted from days
226-235 Top Right R3 sensed day
236 Bottom Z-score (rainbow scale blue 1, red
3) Botswana (approximately 300 300km)
18
Zthresh 2.0 Left Maximum Z-score over 30 day
period (rainbow scale blue 1, red
12.0) Right Day when Maximum Z-score
occurred (rainbow scale blue beginning of
month, red towards end of month) Botswana (500
500km)
19
Zthresh 4.0 Left Maximum Z-score over 30 day
period (rainbow scale blue 1, red
12.0) Right Day when Maximum Z-score
occurred (rainbow scale blue beginning of
month, red towards end of month) Botswana (500
500km)
20
Compute heavy algorithm - search for best
Zthresh Left Day when Maximum Z-score
occurred Right Day active fire detected over
same period (rainbow scale blue beginning of
month, red towards end of month)
21
Pixels labeled based on sign of nadir R3
predicted before change date - nadir R3 predicted
after change date green change/burned red
change/not burned black no change blue
unknown (insufficient observations predicted
before and after R3)
22
Conclusions

• This approach
• maps the burned area and the approximate date
  that burning occurred
• removes the need for poorly understood
  reflectance thresholds which are sensitive to the
  spatial and temporal variations of burned areas
• BRDF variations implicit
• MODIS implementation ongoing
• Evaluation validation under auspices of SAFARI
  2000 dry season campaign
• Burn scars interpreted on Landsat ETM images for
  study areas representing different fire regimes
  in Southern Africa
• dry regions large less spatially fragmented
  fires
• humid regions small and fragmented fires