An area-based accuracy assessment methodology for digital change maps

A methodology was developed to evaluate the accuracy of maps of land-cover change produced from multi-temporal Thematic Mapper images for Australia. The methodology is based on 500-pixel-by-500-pixel areal sample units. Samples units were selected in a two-stage procedure that first ensured that a certain number were selected for which change was apparent on the change map being evaluated. The remaining sample units were then selected according to a spatially distributed random sample without replacement. An independent estimate of change was obtained for each sample unit by a human operator using one or more image enhancement techniques to determine subjectively how much change occurred. Confidence intervals for an increase and a decrease in woody vegetation were calculated on the independent estimate of change and the amount of change on the original map compared against them. To obtain stable confidence intervals, the method requires that 33 samples per change map/TM image are examined for each map of land-cover change, something that requires approximately two person-days per change map. The methodology was found to be robust for change maps on which change was relatively rare (about 0.13% of the total change map) and also for those on which change was relatively common (about 2.5% of the total change map).