2004
hoto-
ble 3
int of
| and
nt of
soit ds
th /
Non-
ne the
ovide
ple, if
or all
o be
nably
1g for
orest,
error
nitely
'erage
ses is
lon as
2" dn
jon is
bably
It is
cause
sified,
itially
hange
sed to
ssibly
ld be
. This
rather
is not
mean % definite error forest / non-forest classes
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
quite as good as nationally. The Forest definitely wrong error
rate is similar at ~3% whereas the definitely incorrect Non-
Forest class, is slightly better than the national average (2%).
The total percent probably or definitely correct is comparable
(87% for Australia and 90% for Tasmania). Results for
Regrowth and Deforestation are very different however (up to a
third of Regrowth being erroneously classified). It should be
noted, however, that this information is derived from a very
small sample.
In the method developed, there is no explicit evaluation of the
quality of change maps. In effect, one evaluates the
classification of a change map for one time period. One is
therefore not in a position to say that landcover change has or
has not been correctly identified. Instead, one can only reason
that, if in Tasmania there is a tendency to identify Forest that is
not really there, then there is a possibility that too much
Regrowth and/or too little Deforestation has been identified. A
tendency to identify Non-Forest for areas that are judged by
photo-interpreters to be Forest would lead to the opposite
conclusion — it may be that the amount of Deforestation has
been overestimated and/or the amount of Regrowth
underestimated. This depends, of course, on the nature of the
pixels examined. If pixels that have been erroneously identified
as Forest remain erroneously classified as Forest over the entire
study period, then they will not lead to an overestimation of the
amount of Regrowth although the amount of Forest will be
overestimated. In contrast, if such pixels were initially
classified as Forest and then from some date onward were
misclassified as Non-Forest, then the amount of Deforestation
will be overestimated.
To evaluate this, the lineage data are employed. To produce
Table 3, definite errors have been tabulated by their lineage
class. It would also be possible, of course, to tabulate probable
errors, definitely correct classifications, etc. For Tasmania,
whilst 2% of the total sample points were in the Regrowth
lineage class, 8% of definite errors were in this class. This
suggests that the incorrect classification of Non-Forest pixels as
Forest may be influencing the Regrowth class. Given that the
number of definite errors is small (12), this might be ignored.
However, the previous table for Tasmania (Table 2) indicated a
potential problem in probable errors. Hence, it might be more
useful in the case of Tasmania to also tabulate probable errors
by lineage class — a relatively simple undertaking.
Definite Errors vs. Quantum of Change
60 ;
i
5 si50
40 Sb3
30 [S052 SG56
20
|
| SF55
10 |
|
| SH55
| SH56 SH50 SG55
0 | i A :
0 400000 800000 1200000 1600000 2000000 2400000
200000 600000 1000000 1400000 1800000 2200000
scaled sum of change in deforestation and regrow th areas (ha)
Figure 2 Graph showing Definite Errors vs. Quantum of Change
751
Using the two types of tables presented and interpreting them in
tandem identifies areas of potential problems, even though a
definitive statement of Correct/Incorrect classification is not
presented, and interpretation of the tables requires a
knowledgeable user. The methodology cannot be employed,
however, without consideration of a number of issues.
One issue is the sampling scheme employed for both aerial
photo selection and individual sample points on each image.
Photos were selected for use based on their geographic and
ecological distribution. However, no attempt was made a priori
to obtain either a completely representative sample or a
completely random sample. Therefore, it is not appropriate to
say, for example, that the amount of Forest has been
overestimated across all of Tasmania. While there is no known
sampling bias related to photo selection, the sampling scheme
employed is probably not statistically robust enough to make
reliable inferences for all of Tasmania. This is not a problem,
however, if one limits the interpretation of results to their
original purpose — to identify any potential problems in
classification to improve the change map methodology. As for
the point samples on the photographs, points were randomly
selected from the grid that was overlayed on a photograph.
However, because the same grid was employed regardless of
photo scale and landscape conditions, the effects of spatial
autocorrelation were probably present in the sample points
extracted from a single photograph and this effect would vary
from photo to photo. For example, for a standard size (23 cm
by 23 cm) 1:20000 photograph, grid points are spaced
approximately 115 m (ground distance) apart whereas on a
1:80000 photograph they are spaced 460 m apart. In the
method developed, no control was placed on the geographic
distribution of points selected from a single photograph, nor
was any attempt made to quantify the effects of spatial
autocorrelation. It remains, nonetheless, that 1:80000
photographs were probably more representative of general
landscape conditions than 1:20000 photographs that cover a
smaller area.
3.2 Prioritisation
It is useful to consider the confirmed errors, in deforestation and
regrowth, for each map tile against the quantum of change. It is
possible to then determine the "performance" in error rate
against the "importance" in the quantum of change.
The error rate is taken as the average percentage definitely
wrong for the Forest and Non-Forest classification. The
quantum of change is the scaled sum of deforestation and
regrowth reported in Jones ef al., (2004). Figure 2 plots each
mapsheet definite error rate against its corresponding quantum
of change. Decision lines are then used to divide the graph into
four regions: High Error / Low Change -Medium Update
High Error / Low Change High Error / High Change
Medium Update Priority High Update Priority
Decision —^
boundary Low Error / Low Change
Low Error / High Change
Low Update Priority Medium Update Priority
Figure 3 Explanation of update priority graph
