for assessing the accuracy of LC product derived from remote sensor data were determined to be
inappropriate for an assessment of our LC change products because these methods require statis-
tically sufficient class representation (
) and relatively homogeneous distribution (CGC, 1994;
Congalton and Mcleod, 1994). In our study LC change was a relatively rare event and tended to
be concentrated in relatively few areas. Reference data to support a traditional accuracy assessment
approach were not available.
The area-based method developed by Lowell (2001) was implemented to provide an independent
estimate of change for which confidence intervals were calculated. State estimates of change were
then compared against these confidence intervals to provided a means of evaluating the accuracy
of the state-produced LC change products. State estimates were within the established 95% confi-
dence intervals for 60 of the 67 scenes initially tested. The seven scenes that did not meet the
acceptance criteria were reprocessed and retested, and five were subsequently accepted. LC change
rates were underestimated by the analysts or overestimated by the states for the two scenes not
accepted. The method overcame the difficulties caused by the lack of suitable reference data. This
is likely to be a common difficulty in large-area studies of LC change. Suitable reference data will
rarely be available to match the multiple dates for LC change studies. Even when multiple-date
data are available, obtaining a “true” change map will be difficult, since the overlay of multiple-
date LC is likely to introduce error (Lowell, 2001).
We used an area-based sampling unit rather than a discrete sample-based approach because of
the relative rarity of woody vegetation change and the difficulty (and cost) of sampling enough
points across a change map to support a rigorous statistical assessment. Based on extensive testing
of the sample unit size, Lowell (2001) determined that the 500-
500-pixel sample unit provided
stable estimates of the confidence intervals after relatively few sample units had been examined.
The present study demonstrated that a 500-pixel sample unit was a practical size for evaluation.
When sample unit location had been automated, one operator could evaluate a change map with
33 samples in approximately 10 h. The area-based reliability method provided a cost-effective
evaluation of the results of the ALCC project and represented only 3.5% of the total project budget.
Overall, the assessment demonstrated that the process of detecting LC change from TM data
provided repeatable and reliable results. Different change techniques and approaches to radiometric
calibration among individual states did not negatively affect results. Because LC change was a
relatively rare event, the area-based methodology had a considerable advantage over more traditional
point-based evaluation methods that require a large number of points (
) to support a rigorous
statistical analysis. The method is particularly useful when suitable reference data for testing the
change estimates are unavailable.
Digital data sets and the final report are available on CD-ROM. Copies can be obtained from
the first two authors or downloaded (http://adl.brs.gov.au/ADLsearch/).
Australia's first NGGI identified that land clearing could be contributing as much as 25% of
Australia's total greenhouse gas emissions. These figures were regarded as very uncertain, and a
collaborative project was undertaken with eight state agencies using TM imagery and other data
to document the rates of change in woody vegetation from 1990/1991-1995. The reliability of this
project's results was assessed using a method developed by Lowell (2001) for this purpose.
Traditional methods of accuracy assessment were impractical given the large size of the study area,
the relative rarity of the change detected, and the lack of an appropriate reference data set. Lowell's
method was implemented to provide an independent estimate of change against which state agency
estimates were compared. The reliability assessment demonstrated that the process of detecting
land-cover change from TM imagery was repeatable and provided consistent results across states.