Geometrically Invariant Feature Descriptors for Terrain Data

In this work, we demonstrate that some feature descriptors, invariant to scaling, rotation, and translation, based on peaks, pits, and saddles are highly effective in registering height data. The feature descriptors are - the single relative height histogram (SHH), multiple relative height histograms (MHH), and Shape Context (SC). These feature descriptors have the advantage over SIFT that they are geometrically meaningful for height data. This allows scientists to decide which features to preserve while compressing large data sets, a task not possible using image features. We compare these feature descriptors with each other and with SIFT (using images). The results indicate that MHH and SC are most effective for data registration. MHH has the advantage that it does not need intergrid spacing information that SC requires.