Skip to Main Content
Automatic feature identification from orbital imagery would be of wide use in planetary science. For example, the ability to count craters on homogeneous surfaces would enable relative dating of geological processes. The scaling of crater densities and impact rates with crater size is another important issue which could be addressed by automated crater counting. Geological feature cataloging can practically be achieved by hand-labeled imagery only for restricted numbers of features. To handle massive new data sets and higher resolutions such as those arising from Mars Global Surveyor, automated feature identification will be required. Many pattern recognition algorithms could be applied to this problem, but a systematic validation process will be required to select the best method for each scientific application and to determine its reliability for scientific use. We demonstrate such a validation process applied to a particular trainable feature identification algorithm when used to detect craters in synthetic imagery and in Mars Global Surveyor imagery. The feature identification algorithm is the continuously scalable template matching algorithm of Burl et al. (2001). The validation process involves separate experiments for subpopulations selected from a labeled crater corpus. The subpopulations are defined by crater density. For the selected subpopulations, the validation process includes training the algorithm on some craters and testing its identification accuracy on others. These results can be summarized in terms of statistical efficiency measures. Efficiency results depend on the subpopulation tested. We illustrate algorithm performance on data from Martian regions of high scientific interest.