Conservative bias in classification accuracy assessment due to pixel-by-pixel comparison of classified images with reference grids

Abstract

The use of reference grids derived from aerial photography for a pixel-by-pixel comparison with classified images can yield conservative estimales of classification accuracy. Even if the class assignment of each polygon is 100 per cent correct, and there is no change in cover type due to temporal differences between the reference data and the classified image, conservative bias in estimales of classification accuracy are still possible. In this letter, we discuss two major sources of this potential bias: 1. positional errors, and 2. difference between polygon minimum mapping unit (MMU) area and pixel size of the classified image.