Transcript An Accuracy Assessment of a Digital Elevation Model Derived From

Assessing Annual Forest Ecological Change
in Western Canada
Using Temporal Mixture Analysis
of Regional Scale AVHRR Imagery
Over a 14 Year Period
Joseph M. Piwowar
Derek R. Peddle
Diedre P. Davidson
Waterloo Laboratory for Earth Observations
University of Waterloo
Waterloo, Ontario, Canada
Department of Geography
University of Lethbridge
Lethbridge, Alberta, Canada
Objectives


to model the inter-annual characteristics of peak
vegetation vigour in Central North America
to assess the effectiveness of Temporal Mixture Analysis
for identifying and modelling the temporal characteristics of
vegetation vigour
Rationale

the North American boreal forest is an important
component in the global carbon cycle
–

it has been the focus of recent international attention through large
global climate change research projects such as the Boreal
Ecosystem Atmosphere Study (BOREAS) in western Canada.
the spatial extent of this vast tract of land coupled with an
emphasis on assessing regional scale environmental
change has created a distinct need for comprehensive
temporal analyses of ecological change over various time
periods using regional scale remote sensing imagery
Data



14 years (1981-1994) of NOAA AVHRR imagery from the
NOAA/NASA Pathfinder program were used
coincident subscenes of the central Canadian boreal forest
region acquired in late July, near the peak of the growing
season, were extracted from each of the 14 years
NDVI images were used as estimates of plant vigour
–
different cover types have been shown to have characteristic
profiles corresponding with their phenology on NDVI sequences
Temporal Mixture Analysis (TMA)

based on Spectral Mixture Analysis (SMA)
–



a procedure that attempts to extract the fractional radiance
components from the pixels in an image
endmembers are defined which characterize the most
extreme, or "pure" spectra present in the data
endmembers are then entered into a mixing model which
defines the spectral mixing of the scene components from
which a set of fraction images is produced
in this study, the SMA technique is extended to analyze
temporal spectra of NDVI values of the boreal forest
Endmember Definition



spectral endmembers are defined as spectrally "pure"
features (e.g. vegetation, soil, etc.).
pure spectral endmembers are usually defined under
idealized in situ or laboratory conditions where reflectance
spectra are acquired using a portable spectroradiometer
focussed only on a single surface (e.g. a single leaf from a
maple tree).
when in situ measurements are not possible, spectral
endmembers can also be derived from "pure" features in
the imagery
Purification Formula



since idealized in situ measurements are impractical in
TMA, we used image-based endmembers for our analyses
sampling spectrally pure pixels cannot always be
guaranteed when using image-based endmembers
we have devised a purification formula to "purify" a sample
of image-based spectra:
EMi =
where:
 MAXi if MEAN i > MEDIAN I

 MIN i if MEAN i < MEDIAN I
EMi
- endmember spectral value for the ith spectral interval
(where i is represented by year in the present context)
MAX i
- the maximum of all sampled spectral values at the ith spectral interval
MIN i
- the minimum of all sampled spectral values at the ith spectral interval
MEAN i - the mean of all sampled spectral values at the ith spectral interval
MEDIAN I- the median of all sampled spectral values at all spectral intervals
Procedure
stratify the study region into its 2 principal cover types:
prairie grassland and boreal forest
 sample temporal spectra from locations within these two
land covers and purify them for the creation of the first 2
endmembers
 define additional temporal endmembers at locations of
high residuals, using the resulting Error Image as a guide
 continue this process until there is appropriate endmember
coverage of the study region
Results
a 3 endmember model was defined:
 Prairie
 Boreal 1
 Boreal 2
Prairie Endmember
0.0
Fraction
1.0



characterized by consistently low to moderate NDVI values
the temporal signature shows higher NDVI values between
1981-1985 and 1990-1994, and a lower regime during the
mid- to late-1980s; coincident with increased drought
conditions
regions with the highest endmember fractions are in good
spatial agreement with the known extents of the grassland
prairies of Central North America
Boreal 1 Endmember
0.0
Fraction
1.0




elevated NDVI values suggest that this is an area of
consistently high plant vigour
lower NDVI values are at the beginning and end of the
temporal record, and higher readings occur during the
middle years
spatially, highest fractions from this endmember follow a
line extending southeast from the northern Alberta border
to Lake Winnipegosis
Environment Canada identifies this region as the Boreal
Plains Ecozone, characterized by mixed deciduousconiferous forests
Boreal 2 Endmember
0.0
Fraction
1.0




moderately high NDVI values indicate good plant vigour
this region is defined as the Boreal Shield Ecozone; it is
over 80% forested with dense coniferous stands
highest fractions are in the northern sections of the study
area
a temporal anomaly is evident in 1983 where some NDVI
values dip below +0.1
– there was a particularly strong El Niño event in 1983 which brought
hot, dry weather over this region which may have resulted in a
dramatic decline in plant vigour
Example Interpretation

The NDVI values for a Boreal Shield location can be separated into a temporal mixture
of 76% Boreal 2 and 15% Boreal 1, with 14% of the temporal signal being consistent
with the Prairie temporal. This suggests that much of the biomass at this location can
be expected to remain moderate from year to year (at least during the last weeks of
July), except under more severe drought conditions. There are portions of this location
whose annual NDVI values more closely resemble those of prairie grasslands, while
other fractions appear as denser forest stands similar to those represented by the first
boreal endmember. Thus, based on the temporal characteristics of its annual July
vegetation index we can deduce that the forest canopy is more variable here. Some of
this variability may also be attributed to variations in timing of the peak of a particular
growing season, which may be affected by, for example, a late winter thaw or different
precipitation and temperature regimes in spring and early summer as a result, the peak
of the growing season will not always occur at the same time, and this "peak" therefore
may not necessarily be captured consistently in the timing of these images we
analysed. This is a key point in the seasonal temporal interpretation that may explain
some of the seasonal variability noted
Conclusions


an annual sequence of NDVI images of Central North
America can be temporally unmixed into three
characteristic temporal patterns: one representing the
prairie grasslands, one from the boreal plains ecozone,
and a third from the boreal shield ecozone
TMA is an effective technique for identifying and modelling
the various temporal characteristics of spatial features.
TMA encourages both temporal and spatial interpretations
through the use of graphical and image displays, hence it
is also a useful tool for exploratory image analysis