The use of topographic features is proposed for classifying targets in synthetic-aperture radar (SAR) images. These features are based on the curvature properties of a surface model that is used to estimate the underlying image intensity. The stability of these features with respect to aspect, sensor noise, and squint is investigated. Attention is given to ensuring appropriate registration of the SAR images and to the modeling of noise in the SAR imaging process. Sensitivity and stability of features are quantitatively and qualitatively analyzed, based on each pixel’s label and on the relative groupings of features in the corresponding images. Stability results are presented for images obtained from XPATCH simulation software as well from as the Moving and Stationary Target Recognition target data set. The stability of these features is compared with the stability of features obtained directly from the SAR magnitude images.