Spatial sampling to assess classification accuracy of remotely sensed data

Selecting a proper sample design is important in assessing the classification accuracy of remotely sensed data. However, design guidelines often do not account for spatial autocorrelation of map error or are based on simulations from a few select landscapes and cannot easily be generalized beyond those data sets. This study reviews the effect of spatial autocorrelation and cost on the relative efficiency of three unbiased sample designs within a stratified framework. These designs are simple random sampling, systematic sampling and cluster sampling. To illustrate the results, sampling simulations are performed on possible error patterns in a vegetation cover map of the Great Basin ecoregion. This map was generated from Landsat Thematic Mapper data by the US Fish and Wildlife Service's Gap Analysis Program.