Object-based image analysis is quickly gaining acceptance among remote sensing community, and object-based image classification methods are increasingly being used for classification of land use/cover units from high-resolution satellite images with results closer to human interpretation compared to per-pixel classifiers. The problem of nonlinear separability of classes in a feature space consisting of spectral/spatial/textural features is addressed by kernel-based nonlinear mapping of the feature vectors. This facilitates use of linear discriminant functions for classification as used in artificial neural networks (ANNs). In this paper, performance of a recently introduced kernel called cloud basis function (CBF) is investigated with some modification for classification. The CBF has demonstrated superior performance to the tune of about 4% higher classification accuracy compared to conventional radial basis function used in ANN. The results are further improved by using probabilistic relaxation labeling as a postprocessing step. This paper has potential applications in urban planning and urban studies.