This paper proposed to study urban land cover and land use mapping in a chain approach. More specifically, we consider urban mapping process as a unification of training set selection, feature extraction, learning strategy, and validation. We empirically study the interactions among these factors in urban mapping using a 48 bands and 1m spatial resolution CASI image in Zhangye city. We firstly analyze the separativity of various land cover and land use in a large variety of projected spectral feature space using mannually collected training data set. Then, we compare the capability of discrimination of these classes in the image using three diverse learning methods. Finally we discuss these results together with the aim of improving urban mapping. Our preliminary results imply of multifold ways to improve urban mapping: adopting an ensemble and hierarchical mapping strategy; building an informative training data library; exploiting discriminative low-dimensional spectral features; using sample but adequate learning functions.