International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
Deforestation/Regrowth event (Cyclic). Qualitative statements,
made by photo-interpreters, about photographic quality or the
general misclassification of an area were also noted.
The photo-interpreter then selected 20% of the photographs for
the map sheet (two photos if 10 were available for a map sheet)
for the purpose of Quality Assurance (QA). The photos
selected were to be typical of the map sheet evaluated and/or be
somewhat difficult to interpret due to topography, sparseness or
height of vegetation, spectral characteristics, etc. A different
photo-interpreter then independently re-interpreted the photos
used for QA using a different set of 40 randomly generated
points. If the general conclusion for both photographs was the
same — i.e., there was/was not a potential problem with the
classification of the area covered by the photograph — results for
the map sheet were communicated to the Australian
Greenhouse Office. If the general conclusion was not the same,
the reason for the differences was determined, and the photos
reinterpreted by both photo-interpreters. In using such a QA
procedure, a reasonable level of confidence was attributed to
the general conclusion for a map sheet.
Results for each photograph were tabulated and then
summarised by map sheet to provide an indication of the quality
of the change map for a given map sheet and results were then
summarised by state and for all of Australia (Jones et al., 2004).
3. UNCERTAINTY IN GREENHOUSE FOREST
ASSESSMENTS
3.1 Results and Discussion
The aim of the continuous improvement protocol is to provide a
methodology that evaluates the general problems in change map
classification, and / or assesses if there are any problems with
specific regions or strata. Presented here are the results for the
state of Tasmania and Australia as a whole (Lowell ef a/., 2003;
Jones ef al, 2004). Tables 1 and 2 describe the photo-
interpretation relative to the change classification. Table 3
provides temporal, or lineage, information. The first point of
note is that few sample points fall into the Regrowth and
Deforestation change map classes. In an assessment of
*change" this may, at first, seem inappropriate. However, it is
in reality of limited concern. Change (Regrowth /
Deforestation) represents a difference in "state" (Forest / Non-
Forest) between sequential images. An ability to determine the
reliability of “state” for a single image will implicitly provide
reliable “change” determination and vice versa. For example, if
it is known that the amount of Forest is overestimated for all
years, then the amount of Deforestation is likely to be
underestimated.
Overall the classification for Australia (Table 1) is reasonably
good. What was classified as Forest was definitely wrong for
only 2% of the total Forest verification points. Non-Forest,
Regrowth and Deforestation returned results with definite error
rates of 4%, 10% and 9% respectively. Probably and definitely
wrong error rates for forest are higher at 6%, whilst the average
probably and definitely wrong error rates for all classes is
around 12%. It is inappropriate to interpret this information as
a determination of change maps being “accurate/inaccurate.” In
an absolute sense, one could argue that the classification is
“good” since 93% of the Forest points are at least probably
correctly classified and only 2% are definitely wrong. It is
more difficult to argue that the classification is “bad” because
only 66% of the Forest pixels are definitely correctly classified,
since the uncertainty of “probably” correct is potentially
attributable to the aerial photograph analysis and not the change
mapping. The indefinite categorization “probably” is used to
highlight areas of doubt in the image classification and possibly
in the photo-interpretation, and where further data should be
sought to determine the correctness of the classification. This
provides an important guide to continuous improvement rather
than only producing a verification statement.
In Tasmania (Table 2) the image classification accuracy is not
Total Points DF PF . U PNF DNF %Def. Wrong %Prob.+ Wrong %Prob.+ Right %Def. Right
Forest 5085 3380 1365 31 209 100 V2 6 93 66
Non-forest 7318 282 847 88 1186 4915 4 15 83 67
Regrowth 105 38 49 1 6. «411 10 16 83 36
Deforestation 56 5 8 0 E32 9 23 77 57
Total 12564 3705 2269 120 1412 5058 3 12 87 67
Table 1 Australia: Tabulation of sample point results by change class (36 map sheets)
Total Points DF PF U PNF DNF %Def. Wrong %Prob.+ Wrong %Prob.+ Right %Def. Right
Forest 277 152 15793 04823 9 3 12 88 55
Non-forest 83 2 2 9.42 . «37 2 5 95 45
Regrowth 3 0 2 0 0 1 33 33 67 0
Deforestation 5 0 0 0 2 3 0 0 100 60
Total 368 548 377 24 229 444 3 10 90 52
Table 2 Tasmania Tabulation of sample point results by change class (one map sheet — nine aerial photos)
Tasmania | Lineage Class!
FT NFT DEF REG CYC
Total Points 368 270 25 54 7 12
% of Total Points 100 73 7 15 2 3
Definite Errors 12 8 2 0 1 1
% of Total Definite Errors 100 67 17 0 8 8
'Lineage classes: FT — Forest Throughout, NFT — Non-Forest Throughout, DEF — Deforestation, REG — Regrowth, CYC — Cyclic.
Table 3 Tasmania Lineage information
750
mean % definite error forest / non-forest classes
Intern
quite
rate i
Forest
The t
(87%
Regro
third |
noted,
small
In the
quality
classif
theref
has no
that, if
not re
Regro|
tenden
photo-
conclu
been
undere
pixels «
as Fore
study p
amount
overest
classifi
misclas
will be
To eva
Table :
class. I
errors,
whilst
lineage
suggest:
Forest r
number
Howeve
potentia
useful ii
by linea
60 ,— — —
50
40 SI
30 |SD52
20
ol x
0 :
200000
Figure 2 C
