Effect of spatial resolution on classification error in remote sensing

The effect of spatial resolution on the overall classification error of remotely sensed image data is studied based on simulated images containing two classes of objects. Though containing more detailed ground information, the images of finer spatial resolution do not necessarily achieve higher classification accuracy. This study considered this problem by decomposing the overall classification error into (a) the pure-pixel classification error and (b) the mixed-pixel classification error, for spatial resolution has opposite effects on them. The pixel maximum likelihood classifier was used. Significant parameters considered in this simulation include the ratio of the ground sampling distance to the object field size, and the class variance normalized by class mean difference. Various values of these parameters were used to obtain the relationship between classification errors and spatial resolution. Primary results include: (i) A valley may occur in the relation of the overall classification error versus the ratio of the ground sampling distance to the object field size. (ii) The overall classification error decreases with increasing spatial resolution if the class variance remains constant. (iii) The pure-pixel classification error is a nondecreasing function of spatial resolution