摘要： A key strength of remote sensing, and one of the main reason for its usage, is the provision of spatially exhaustive (wall-to-wall) coverage of a region of interest. Classification and interpretation of the remote sensing data allow for thematic mapping of features present in the region. However, mapping complex and often spatially continuous surface conditions into a set of discrete map categories is bound to result in some of the map units being erroneous. The magnitude of errors will determine the reliability, usage, and interpretation of the map, which is why map users and producers have a direct interest in communicating and understanding the quality of maps. This is the primary reason for the tradition within the remote sensing community of conducting map accuracy assessments (terms in italic are explained in the Terminology section). The basis of an accuracy assessment is the comparison of the map and a sample of observations of reference conditions at certain selected locations. The sample is selected by probability sampling if the locations of the sampling units are selected such that the likelihood of a unit (a pixel for example) being included in the sample is known and greater than zero (Stehman, 2001). A probability sample allows for inference of the accuracy of the map for the entire population, which in this case is the collection of map units from which the sample is selected. For example, consider the following common scenario: a land-cover map has been constructed over specific region and a set of units have been selected by simple random sampling. If identifying land-cover reference conditions at each location in the sample, the overall accuracy of the map can easily be computed as the ratio of correctly classified units to the total number of units in the sample. The overall accuracy is a measure of the probability that a random map unit is correctly classified—not just a random map unit in the sample, but of all units of the map. This holds true because the sample was selected by probability sampling. As explained in section “Design-Based Inference”, accuracy measures specific to the individual map categories are also easily computed.