Accuracy assessment based on the distribution of the classified errors on classified map: Simulation analysis

Accuracy assessment is essential after the classification of remote sensed image. For it is expensive for field survey of every sample, how to select an effective sample that is the unbiased estimation of the population which defined as the all the pixels on the classified map is a very important problem. The classification errors are not randomly distributed on the classified map but distributed by spatial distribution rule. However, the current three sampling methods used do not take this spatial distribution rule into account. So, in this research adjacent index was suggested for sampling which has a strong relationship with the distribution of the classified errors. Five kinds of classified maps were simulated to test the sampling design effect. The result indicate that, with the same sample size, the effect of stratified sampling by adjacent index is similar as simple random sampling when the percent of errors inside the patch is greater, and that the effect of stratified sampling by adjacent index is better than simple random sampling when the percent of errors on the patch edge is greater.