International Archive
minimum area of 0.2 hectares is also imposed (Richards and
Furby, 2003). Given the size of Australia (690 million
hectares) ~370 Landsat Thematic Mapper (TM) scenes were
required for complete coverage for a single date. Where
continent-wide coverage was available for a given date (1989
onwards), TM data were acquired; for dates prior to this,
Landsat Multi-Spectral Scanner (MSS) data was acquired.
For continent-wide image registration, the 2000 imagery was
first geometrically corrected using reference maps and a digital
elevation model (DEM), and mosaic'ed to produce a geographic
reference base for the entire imagery set (Furby, 2002).
Imagery for prior dates was then co-registered to this 2000
geographic rectification base. During this process, TM images
were resampled to 25 m pixels and MSS images were
resampled to 50 m pixels. For radiometric correction over a
multitude of dates, an invariant target correction approach was
adopted (Furby and Campbell 2001). This corrected for sun-
angle and earth-sun distance and employed a bi-directional
reflectance distribution function (BRDF) for surface properties
correction (Danaher et al. 2001).
The continent-wide mosaic was separated into 37 1:1,000,000
map sheets, which are employed by various federal agencies for
mapping Australia. Each map sheet was stratified into areas
that were known to have soils or other characteristics that
would affect the spectral identification of forests. For each
stratum, a series of vegetation indices were formulated and
evaluated to determine their capacity to distinguish between
Forest and Non-Forest. Thresholds for the selected indices
were then calibrated and chosen using training samples to
assign a given pixel to one of three classes: Forest, Non-Forest
and Uncertain (Furby and Woodgate, 2001). The classification
of the Uncertain pixels for all pixels and all dates was resolved
using a continuous probability network (CPN -- Cacetta 1997).
The CPN examined the temporal pattern of the probabilities
derived from the indices while recognising that different change
classes have distinct temporal patterns to their probabilities
(Figure 1). In this manner, Uncertain areas were assigned to
either the Forest or Non-Forest classes to produce the final
maps for each date.
Change maps were derived from this temporal sequence of
classified satellite images. Since there are 12 (multi-temporal)
images associated with any given area, up to 11 change maps
can be produced by comparing sequential pairs of images
covering the period from 1972 to 2002. Temporally, each pixel
can then be classified based on the changes that have occurred
over the entire period 1972 to 2002. Five classes were
identified as follows.
Prababity "brest
|
Tune
s of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
e — Non-Forest throughout - NFT.
e Forest throughout -FT.
e Non-Forest that became Forest and remained Forest -
REG(rowth).
e Forest that became Non-Forest and remained Non-
Forest (deforestation)- DEF.
e Forest or Non-Forest that changed to Non-Forest or
Forest respectively, and then back to Forest or Non-
Forest one or more times - CYC(lic regrowth).
2. METHOD
2.1 Validation
In the context of remote sensing ‘validation’ is the process of
assessing by-independent means the accuracy of data products
(Justice et al, 2000; Privette ef al, 2000). In general,
validation refers to assessing the uncertainty of satellite-derived
products by analytical comparison to reference data (e.g., in
situ, air-craft, and high-resolution satellite sensor data), that are
presumed to represent the true values (Justice et al., 2000). This
is often achieved using a confusion matrix. A confusion matrix
is developed by sampling a sub-set of pixels from the classes
present on a classified image, obtaining better quality
"reference data" for each pixel, and cross-tabulating the sample
pixels. A variety of summary information can then be extracted
from this matrix — e.g., kappa (Cohen 1960), and user's and
producer's accuracy (also known as errors of omission and
commission) (Congalton 1991). The use of a confusion matrix
implicitly assumes that one has reference data that provide a
definitive land classification for each pixel. Gopal and
Woodcock (1994) note that this is rarely the case, and also that
even in the presence of definitive reference data, all
classification errors are not equally incorrect — e.g., confusing a
lake with a swamp is less serious than confusing a lake with a
forest. They therefore propose a fuzzy method of image
classification accuracy assessment that implicitly addresses the
uncertainty inherent in assigning a point to a single class using
the reference data, while also addressing the magnitude of the
difference between the most likely reference data class and the
image class. Foody (1996) notes that while “hard
classifications” are employed in most mapping methodologies
and products, “soft classifications” may be more appropriate
when evaluating the classification of digital imagery. This is
due to the presence of mixed pixels (mixels, particularly at the
interface of two landcovers) and the fuzziness of in situ
observations.
hue dra
terre menm rok: Js
esse ae Lutesd Ooneerx cor
beprtantri
SATS? (NL
Figure |
Typical temporal signatures for
forest and non-forest cover,
after Furby (2002)
M
748
Inter.
2:2
The |
the «
proce
asses
reaso
that
impr
The:
after
some
meth:
(200€
devel
lookii
allow
probl
amen
impro
staten
Initial
envis:
a give
resolu
satelli
altern
arcas
the re
accurz
appro:
quanti
qualifi
the in
point-|
the cu
multi-
Forest
circum
to test
the di
Verifk
accura
suitabi
area. |
were u
1:5000
the sat
covere
each cl
in the
1:2500
colour
Aerial
as the
assesse
from |
photog
agencie
sheets
photog
remote
