QUANTITATIVE ASSESSMENT OF AUTOMATED CRATER DETECTION ON MARS 
Jung Rack Kim *, Jan-Peter Muller, Jeremy G Morley 
Dept. of Geomatic Engineering, University College London, Gower Street, London, WCIE 6BT UK 
$^ ~ a 
jkim@ge.ucl.ac.uk 
WG VI/9 
KEY WORDS: Extra-terrestrial, DEM/DTM, Feature, Geomorphology, Vision, Automated crater detection 
ABSTRACT: 
Crater Size-Frequency Distributions (SFD) on planetary surfaces are crucial to dating the geological age. On the Moon they have 
been employed together with radioactive K-Ar techniques to determine ages of different regions. The launch of the ESA Mars 
Express (MEX) mission on 6 June 2003 with the 9-view camera HRSC (High Resolution Sterco Camera) orbiting instrument and 
subsequent spectacular multi-angle and colour data acquired since January 2004 opens up the possibility of applying the lessons 
learnt on the Moon to Mars. Although there is an on-line web-based cataloguing and mapping system at USGS which shows the 
location and characteristics of some 40,000 craters on Mars (Mars Crater Consortium, MCC) with diameters >5km, these craters 
represent'only a tiny fraction of the millions of craters which are believed to be present on the Martian surface. It is highly unlikely 
that there will ever be sufficient resources to map these smaller craters using existing manually-intensive techniques. An automated 
crater detection algorithm has been developed which exploits both image data and DTMs derived from laser altimetry (MGS-MOC) 
and in future DTMs from HRSC. The algorithm is described and examples of it's application for a variety of different crater types 
are demonstrated. Central to the application of any automated algorithm and prior to systematic application to the Martian planetary 
surface it is crucial to perform a quantitative assessment of any automated algorithm's performance. We show results from three 
different approaches here: (1) inter-comparison of automated crater locations with those in the MCC catalogue; ; (2) inter- 
comparison of automated crater locations with manually-derived crater locations; (3) simulation of crater images using an idealised 
3D model of a Martian crater changing the illumination conditions. 
1. INTRODUCTION with simulated crater features, manually detected craters and the 
MCC (Mars Crater Consortium) catalogue (Barlow, 2003). 
1.1 Aims 
: SIS : 1.2 Previous research work 
Impact crater detection and crater size frequency counting have 
a very high priority in Extra Terrestrial Mapping and planetary 
chronological research. In spite of the increasing demand for 
geological and geodetic control over the last few decades, the 
application of machine vision to address this problem has not 
been very successful. The main reasons can be summarized as 
Several methods to automatically detect craters have been 
developed but are not operational yet. The first data mining 
system for planetary images including the functionality for 
impact crater detection was Diamond Eye (Burl et al. 1999). No 
quantitative evaluation was reported. Another case of impact 
follows:. crater detection on an asteroid, which appears to be relatively 
successful is Leroy et al. (2001)'s work. The primary aim of 
this research was the automated detection of impact crater and 
3D modelling of asteroids. Recently, Michael (2002) developed 
a crater detection algorithm using an elliptical Hough 
a) The “visibility” of impact craters in optical images depends 
not only on the surface scattering behaviour but also on the 
illumination direction, atmospheric state as well as the sensor 
  
incidence direction. 
b) Some geographical features like small valleys and volcanoes 
have similar morphological characteristics as craters in low- 
level feature space. 
c) Impact craters are often concentrated into clusters resulting 
in overlap and for larger structures, multi-ring structures 
frequently occur. This means that the separation of individual 
craters from their background can be very difficult to generalise. 
d) Craters on Mars are frequently eroded due to surface (acolian 
processes such as dust storms as well as the action of water. 
To address these problems, a combinatorial fusion technique has 
been developed to exploit not only image features bur crucially 
3D information. The final objective is to develop a fully 
automated processing system which can accurately detect 
boundary rims of impact craters using raw planetary images and 
3D data with sufficient accuracy for practical planetary research. 
Here we report on an evaluation of the crater detected products 
816 
transformation applied to a global MOLA DTM. His result was 
apparently very successful so that the correct shift vector 
between the MDIM and MOLA DEM could be extracted using 
this information. Magee et al. (2003) showed an automated 
impact carter detection by edge processing and template 
matching. Kim and Muller (2003) suggested a similar crater 
detection method but employing ellipse fitting on the DTM and 
optical image. 
2. ALGORITHMS 
The overall processing steps are shown in Figure 1. The 
procedure consists of 3 stages. Firstly, target edge segments in 
so-called ROIs (Region of Interests) are defined in a focusing 
stage using a GLCM (Grey Level Co-occurrence Matrix) 
texture classifier and edge direction analysis. These 
“Preliminary crater edges” are then organized to find optimal 
ellipses in a second processing step. Optimal ellipses for impact 
craters are evaluated using a fitness function and refined using 2 
Int« 
Ho 
ten 
In 
opt 
  
2.1 
WI 
cor 
hur 
rel: 
alg 
stre 
her 
At 
thr 
edge 
but 
gen 
mo 
bot 
Thé 
illu 
Fig