A comparison of automated crater detection methods

This work presents early results of a comparison between some common methodologies for automated crater detection. The three procedures considered were applied to images of the surface of Mars, thus illustrating some pros and cons of their use. We aim to establish the clear advantages in using this type of methods in the study of planetary surfaces. Introduction The usefulness of an accurate method for automated crater detection does not cease to be highlighted by the continuous improvement in spatial resolution shown by new image acquisition systems currently in the course of their missions (HiRISE/MRO) or planned for the near future (LROC/LRO). This will most certainly lead to the analysis of small craters in limited areas, in fact creating a new set of (until now) unknown surfaces, and the need for improvement in existing catalogues of impact craters on several planetary objects, namely Mars and the Moon. The recent flyby of Mercury by the Messenger probe, and its future insertion into orbit, also point to the fact that other unknown terrains still exist in the Solar System, and that studies in crater density continue to occupy a high priority when it comes to analysing new vistas of planetary surfaces. In fact, through the years there have been several different approaches to the issue of automated crater detection. Generally, each team involved in the research of this theme adopted a methodology and stuck with it, producing results that are not easy to compare, also because the type of image data analysed is not always the same [1-4]. Some workers have tried to change this situation [5], and it is hoped that their efforts will lead to improvements in the evaluation of automated crater detection. In this work, we adopt a somewhat different approach: we compare our methodology, whose development and testing is documented in a number of papers [6-8], with two other widely used methods. Dataset and methodologies Our work on automated crater detection has been centred on the use of a template matching approach, developed and applied to a large set of MOC/MGS wide-angle images of the surface of Mars, and that produced results that can be seen as very good, within a given crater size range. Its application to other images of higher resolution is currently in execution. However, we have also researched other approaches, and we present in here some results of this work which, in a way, confirm the correctness of our original methodology. At this point our focus is on a general comparison of methodologies; thus, the images employed in this work were not selected subject to any particular considerations on location or age of terrain. The dataset used can be simply said to be a number of wide angle MOC images of the surface of Mars. The methods used in this work, and very briefly described below, were: Template matching Hough transform Boosting procedure The methodology that we developed originally comprises a binarization phase followed by a template matching procedure based on FFT (Fast Fourier Transform), which allows for the pinpointing of the centre of a crater and of its diameter. The Hough transform is a classic methodology for the detection of circular features, which is of wide application and commonly achieves good results. Its use in this field faces the issue of the non-regularity of crater shapes, and the possible question of being too demanding in computational resources. The boosting procedure [9] is a machine learning tool which calls for the definition of training features by the operator and also demands a substantial period of time for the algorithm to get acquainted with the type of images and the objects it is designed to identify. This can lead, however, to interesting results in the identification of several types of features. Results To illustrate the comparison made, we present in Fig. 1A a wide-angle MOC/MGS image. This is centred on 78.03W-19.91S, which is to say that it shows a scene from the plains to the south of Valles Marineris. Its spatial resolution is 236.95 m/pixel. This wide area does not have a high crater density, and thus seemed like a good starting point for this comparing study. The lower limit in crater size considered was a diameter of 10 pixels, corresponding to craters with a little more than 2.5 km in diameter. There are a number of smaller craters visually discernible in the scene, but these were not targeted for automated detection by any of the methods. Hence, there were six craters in the range of detection. Note that there are other features present, such as ridges and pits which could easily result in problems for automated methodologies, namely false detections. A: Original MOC/MGS wide-angle image E01-00976 [NASA/JPL/MSSS] B: Results obtained by the template matching method C: Results obtained by the Hough Transform method D: Results obtained by the boosting procedure Fig. 1 Original image and results for automated crater detection with different methodologies (blue: correct identifications; green: misses; red: false positives). Our methodology detected five of the craters targeted (see Fig. 1B), and missed one, with no false positives. The Hough transform, also with no false positive detections, located only the two largest craters, and missed the other four (Fig. 1C). Finally, the boosting procedure, though correctly identifying five craters and missing one of the smallest, also produced a false positive, even in such an uncluttered scene (Fig. 1D). Conclusions and future work Though limited in scope, these results show some of the advantages and problems associated with each of the methodologies tested. Our methodology (template matching) produces very good results, especially in images with an almost noiseless background; note that it is not affected by the other features present, even if they have rounded edges. The Hough transform has a reputation for large consumption of computational resources, but this can be circumvented by a correct implementation; it seems to deal less well with small craters, though producing better results for the diameters of those that are identified. Finally, the boosting procedure does show a tendency for false positives and artificial enlargement of the detected craters. Note, however, that these are preliminary results, and that these tendencies can probably be controlled by careful selection of the parameters involved in the application of each methodology. This work will continue, using more images corresponding to different geomorphological settings, in order to achieve a clear understanding of the strong points of each method and thus lead to the fruitful application of automated methods for crater detection, which can result in saving a lot of time and attention on the part of human experts. AcknowledgementsProject MAGIC (PDCTE/CTA/49724/2003), funded byFCT, Portugal, was the framework of this work. Financialsupport for JS (SFRH/BD/37735/2007) and LB(SFRH/BD/40395/2007) was provided by FCT, Portugal. References[1] Flores-Méndez, A. (2003) LNCS, 2905, 79-86.[2] Kim., J. et al. (2005) Photogr. Eng. Rem. Sens., 71,1205-1217.[3] Plesko, C. et al. (2006) LPS XXXVII, 2012.[4] Bue, B. and Stepinski, T. (2007) Trans. Geosci.Rem. Sens., 45, 265-274.[5] Salamuniccar, G. and Loncaric, S. (2008) Adv.Space Res., 42, 6-19.[6] Barata, T. et al. (2004) LNCS, 3212, 489-496.[7] Bandeira, L. et al. (2007a) LNCS, 4477, 193-200.[8] Bandeira, L. et al. (2007b) Trans. Geosci. Rem.Sens., 45, 4008-4015.[9] Martins, R. et al. (subm.) Geosci. Rem. Sens. Lett.