The legend used to map a given region is most of the time the first characteristic assessed by end users. In GEOBIA, the choice of a classification system is primarily guided by the scale of analysis and the underlying data model. This study focuses on the classification systems related to spatial regions. Depending on the purpose of the GEOBIA analysis, three types of classification systems have been identified: majority-based, LCCS like and feature-based. The validation of the classification results then relies on an appropriate choice of response design, which in turn influences the accuracy results. The impact of the different response designs is highlighted with a real case study. A high accuracy pixel-based classification of the segmented image is used as synthetic reference and classification errors are simulated on the object database. The results shows that all methods are equivalent if all polygons are pure, but that the uncertainty conferred by under-segmented or heterogeneous polygons affects the accuracy estimates in different ways: some methods tolerate delineation or under-segmentation while others are directly affected. This experiment also shows some differences and similarities between the point-based and the object-based validation frameworks, which are both valid depending on the circumstances but could be misleading when they are not applied with the right purpose. The thematic precision of the legend is also discussed: the LCCS authorize mixtures of different spatial entities in the class definition, but the information content differs depending of the purity of the majority of the classes.