Since the launch of the first Landsat satellite in 1972, satellite remote sensing has been increasingly recognized as a tool for mapping and area estimation of earth resources. The Landsat MSS records a region on the ground about one acre (0.5 ha) in size. This provides a good spatial resolution for mapping purposes, and Landsat data have been used for mapping such characteristics as general land use and soil type. Estimation of the areal extent of a feature has been a key use of Landsat data. The primary uses for area estimation have been in agriculture with crop and forest area estimation.
Although many researchers and users have analyzed Landsat data, the matter of determining and expressing in a meaningful and useful way the quality of a classification is a difficult problem. In evaluation of classification results, the experimenter may be concerned with two types of accuracy: classification accuracy and proportion estimation accuracy. By classification accuracy, we refer to the pixel-by-pixel count of the percentage of times the decision rule has produced the correct response. By proportion estimation accuracy, we refer to how close an estimate (e.g., of crop proportion) is to the "truth" or to some reference standard.
In the application of remote sensing technology to the problem of area estimation, classification accuracy may not be of prime importance. Compensating classification errors among categories or methods of estimation may enable the researcher to obtain accurate area estimates without attaining a classification accuracy as high as might be needed for mapping purposes.
Proportion estimates of classes of interest can be computed by direct estimation or unbiased estimation methods. The accuracy of these proportions can be assessed with respect to some reference standard or can be compared with results from other data analyses. This paper addresses methods of proportion estimation and qualitative and quantitative methods for evaluation of area or proportion estimates.
Date of this Version