Impact of training and validation sample selection on classification accuracy and accuracy assessment when using reference polygons in object-based classification

Reference polygons are homogenous areas that aim to provide the best available assessment of ground condition that the user can identify. Delineation of such polygons provides a convenient and efficient approach for researchers to identify training and validation data for supervised classification. However, the spatial dependence of training and validation data should be taken into account when the two data sets are obtained from a common set of reference polygons. We investigate the effect on classification accuracy and the accuracy estimates derived from the validation data when training and validation data are obtained from four selection schemes. The four schemes are composed of two sampling designs (simple random and systematic) and two methods for splitting sample points between training and validation (validation points in separate polygons from training points and validation points and training points split within the same polygons). A supervised object-based classification of the study region was repeated 30 times for each selection scheme. The selection scheme did not impact classification accuracy, but estimates of overall (OA), user's (UA), and producer's (PA) accuracies produced from the validation data overestimated accuracy for the study region by about 10%. The degree of overestimation was slightly greater when the validation sample points were allowed to be in the same polygons as the training data points. These results suggest that accuracy estimates derived from splitting training and validation within a limited set of reference polygons should be regarded with suspicion. To be fully confident in the validity of the accuracy estimates, additional validation sample points selected from the region outside the reference polygons may be needed to augment the validation sample selected from the reference polygons.