An application of statistical methods to determine the appropriate size and location of classification reference area

Image classification techniques are usually applied to land-use and land-cover researches. One of the most common means of expressing classification accuracy is the preparation of a classification error matrix. Error matrixes compare the relationship between known reference area data and the corresponding results of a classification. Both the representativeness of reference area data and the classification technique influence the overall accuracy. Representativeness of reference area data means the similar component proportion to study area data. Therefore, the purpose of this paper is to find out statistic methods to assist in determining the most representative reference area. In order to assess different methods to determine reference area we tried to make a simulated urban ground truth data which includes 6 categories: road, building, bare land, vegetation, water, and shadow by supervisors Maximum-likelihood method. We tried to apply the Chi-square goodness of fit test to assess the representativeness of reference areas with different sizes and locations, and calculated the indexes x (Chi-square), P _G (the probability of Chi-square cumulative distribution function), and AA (absolute error of classification overall accuracy). The results indicated that the lower x or the higher P _G presents higher representative of reference area, and there is a negative functional relationship between x and P _G. These results led to the conclusion that the appropriate size and location of the reference area could be determined exactly and efficiently by the index P _G.