This study aims to compare urban land cover classifications based on the use of two decision trees - Best First Tree (BFT) and Logistic Model Tree (LMT) - for two different legend levels. The legend level 1 presents 11 urban land cover classes and the legend level 2 presents 42 classes. A WorldView-3 image was used, with 16 multispectral bands and a spatial resolution of 0,31 m. The study area corresponds to a transect within the campus of Campinas State University, São Paulo, in the Southeastern Region of Brazil. At the legend level 1, the classifications achieved very similar accuracies, showing global accuracy indices around 76% and 82%. At the legend level 2, on its turn, the accuracies were very different. The results obtained by the BFT algorithm presented a global accuracy index of 64%, while this index reached 78% for the result produced by the LMT algorithm. The comparisons between classifications accomplished at the same legend level were performed by means of hypothesis tests. The comparisons revealed that at the legend level 1 the performance of the algorithm was significantly the same in both scenes. However, at the legend level 2, the LMT algorithm performed better than the BFT algorithm. Thus, one can conclude that the greater spatial and spectral refinement of the WV-3 sensor contributed to the improvement of the classification accuracy in the more detailed legend level.