In this letter, an algorithm is proposed that robustly extracts urban areas from polarimetric synthetic aperture radar images. Polarization orientation angle (POA), volume scattering power (Pv) derived by four-component decomposition, and total power (TP) are utilized in the proposed algorithm. The dependence of the four decomposition components on POA can be lessened by rotating the elements of the coherency matrix by the POA. However, a level of POA dependence remains even after the correction. The proposed algorithm utilizes POA-corrected components, but pixels are grouped into several categories according to POA. First, urban and farmland training data are selected for each category in a study area. Then, urban and mountain areas are separated from farmland, bare ground, and sea by utilizing the Pv-TP scattergram. Finally, a measure of the POA randomness between neighboring pixels is used to discriminate between urban areas with nearly homogeneous POA and mountain areas with randomly distributed POAs. When performing classification on more than one study area, thresholds manually selected for one of the study areas are used to automatically estimate thresholds for the other areas. An accuracy assessment demonstrates that POA-based categorization and utilization of POA randomness contribute to improving classification accuracy.