1017 
AUTOMATED ROCK SEGMENTATION 
FOR MARS EXPLORATION ROVER IMAGERY 
Yonghak Song 
School of Civil Engineering, Purdue University 
West Lafayette, IN 47906, USA - songlO@purdue.edu 
Commission IV WG 7 
KEY WORDS: Active Contour, Edge Flow, Feature Extraction, Level Set Method, Mars, Multi-resolution, Segmentation, Texture 
ABSTRACT: 
Rock segmentation is important for the success of the Mars Exploration Rover mission and its scientific studies. In this paper, a 
framework for automated rock segmentation using texture-based image segmentation and edge-flow driven active contour is 
developed and implemented. Three schemes: wavelet based local transform, multi-resolution histograms, and inter-scale decision 
fusion are combined and applied for texture-based image segmentation. The result is refined by active contour based on level set 
method, which is propagated in the edge flow vector field. Test images taken by the panorama and navigation cameras on the rover 
Spirit at the Gusev Crater landing site are used in this study. This paper presents the theory, implementation, and the test results 
along with discussions on the performance of the proposed method. 
1. INTRODUCTION 
The Mars Exploration Rover (MER), Spirit and Opportunity 
have collected a large amount of Mars surface imagery since 
their arrival on Mars in 2004. As the most important tasks of the 
MER mission, route planning and geologic analysis demand the 
identification of observed rocks. For route planning, rocks must 
be detected before producing rock maps at the landing sites. In 
terms of geologic and planetary science, rocks might hold the 
clues to past water activity and carry important information 
about environmental characteristics and processes. 
Rock segmentation in an image is essential for rock mapping. 
Currently, rock segmentation in MER imagery is mostly 
accomplished by manual labelling which is extreme time 
consuming and tedious. Further more, the increasing amount of 
data being collected by the rovers or similar missions makes 
manual operation impractical and automated solution demanded. 
In addition, automated rock segmentation is also needed as part 
of the on-board processing. Improvement in the mobility and 
lifespan of MER allows for more images to be collected than 
the capability of the outer space communication bandwidth to 
transmit to the Earth. This fact highlights a crucial demand for 
effective data compression schemes that can prioritize regions 
in an image based on their scientific values. Automated rock 
segmentation will benefit such on-board data compress schemes 
(Roush et al., 1999). 
To meet these needs on automated rock segmentation, this 
study presents an automated solution consisting of two stages: 
texture-based image segmentation as initials and active contours 
based boundary refinement. For the texture-based image 
segmentation, three texture analysis approaches are used: multi 
channel approach, multi-resolution histogram, and inter-scale 
decision fusion. These three approaches are integrated and 
embedded into a framework for rock detection using discrete 
wavelet transforms. This texture-based image segmentation can 
roughly segment the rocks in the MER images, but can not 
yield satisfactory rock boundaries. To resolve this problem, the 
initial boundaries are refined by means of active contours based 
on the level set method. This boundary refinement allows us to 
achieve not only finer boundaries but also topologically correct 
rock segmentation results. Finally, the suggested framework for 
automated rock detection is applied to Mars surface images 
collected by MER PANCAM using various filters and 
NAVCAM, all at the Gusev Crater landing site. 
The rest of this paper is organized as follows. Section 2 briefs 
the previous work for automatic rock extraction, while Section 
3 explains the proposed methodology and describes the detail 
process. Presented in Section 4 are our implementation and its 
results on MER images. The paper concludes in Section 5 with 
the evaluation about the properties and performance of the 
proposed method with perspectives on future efforts. . 
2. RELATED WORK 
There have been a number of efforts towards automatic rock 
extraction from imagery. For mining studies, Crida and Jager 
(1994) propose a knowledge-based approach for rock 
recognition from imagery, which consists of two parts. The first 
part includes three stages: blob edge detection, boundary 
completion, and blob extent calculation. The second part 
involves testing the hypothesis, where the classification of 
interesting regions detected as blobs in the first part is 
performed according to twelve rule-based features. Initially, the 
feature vectors are classified by thresholding and then the 
remaining vectors initially rejected as non-rocks are reclassified 
using a supervised k-nearest neighbour classification. Although 
it is one of the good initial efforts for rock detection, it suffers 
from heavy computation and the difficulty of threshold 
determination. Gilmore et al. (2000) show that rock is texturally 
distinctive features and can be detected successfully in Mars- 
like desert pavement environment since rocks differ 
significantly from soils in terms of texture. They use Gabor- 
filter for texture feature extraction and maximum-likelihood 
method for classification. They focus on general strategy rather