Naslov
Application of machine learning using support vector machines for crater detection from Martian digital topography data

Autori
Salamunićcar, Goran ; Lončarić, Sven

Vrsta, podvrsta i kategorija rada
Sažeci sa skupova, sažetak, znanstveni

Izvornik
38th COSPAR Scientific Assembly / - Bremen : Committee on Space Research (COSPAR), 2010

Skup
38th COSPAR Scientific Assembly

Mjesto i datum
Bremen, Njemačka, 18.07.2010. - 25.07.2010

Vrsta sudjelovanja
Poster

Vrsta recenzije
Međunarodna recenzija

Ključne riječi
Mars; crater detection

Sažetak
In our previous work, in order to extend the GT-57633 catalogue [PSS, 56 (15), 1992-2008] with still uncatalogued impact-craters, the following has been done [GRS, 48 (5), in press, doi:10.1109/TGRS.2009.2037750]: (1) the crater detection algorithm (CDA) based on digital elevation model (DEM) was developed ; (2) using 1/128$^\circ$ MOLA data, this CDA proposed 414631 crater-candidates ; (3) each crater-candidate was analyzed manually ; and (4) 57592 were confirmed as correct detections. The resulting GT-115225 catalog is the significant result of this effort. However, to check such a large number of crater-candidates manually was a demanding task. This was the main motivation for work on improvement of the CDA in order to provide better classification of craters as true and false detections. To achieve this, we extended the CDA with the machine learning capability, using support vector machines (SVM). In the first step, the CDA (re)calculates numerous terrain morphometric attributes from DEM. For this purpose, already existing modules of the CDA from our previous work were reused in order to be capable to prepare these attributes. In addition, new attributes were introduced such as ellipse eccentricity and tilt. For machine learning purpose, the CDA is additionally extended to provide 2-D topography-profile and 3-D shape for each crater-candidate. The latter two are a performance problem because of the large number of crater-candidates in combination with the large number of attributes. As a solution, we developed a CDA architecture wherein it is possible to combine the SVM with a radial basis function (RBF) or any other kernel (for initial set of attributes), with the SVM with linear kernel (for the cases when 2-D and 3-D data are included as well). Another challenge is that, in addition to diversity of possible crater types, there are numerous morphological differences between the smallest (mostly very circular bowl-shaped craters) and the largest (multi-ring) impact craters. As a solution to this problem, the CDA classifies crater-candidates according to their diameter into 7 groups (D smaller/larger then 2km, 4km, 8km, 16km, 32km and 64km), and for each group uses separate SVMs for training and prediction. For implementation of the machine-learning part and integration with the rest of the CDA, we used C.-J. Lin’s et al. [http://www.csie.ntu.edu.tw/$\sim$cjlin/] LIBSVM (A Library for Support Vector Machines) and LIBLINEAR (A Library for Large Linear Classification) libraries. According to the initial evaluation, now the CDA provides much better classification of craters as true and false detections.

Izvorni jezik
Engleski

Znanstvena područja
Fizika, Geologija, Računarstvo

POVEZANOST RADA