ul 2004 
(5) 
pairs by 
ed conic 
n-l 
1 given 
These 
oy other 
between 
sholding 
optimal 
; a very 
d graph, 
  
  
ses, OE 
more. If 
4), it 1s 
Figure 6 
ied with 
step is 
several 
S: 
on stage 
> of the 
yn value 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004 
  
between a predefined template and a detected crater ellipse. At 
first, the detected ellipse is resampled to a comparable size as 
the template and transformed to similar illumination conditions 
(sun azimuth angle is used here). Exact resizing and rotation is 
not feasible so that the Gruen (1991) image matching scheme, 
which has been the best solution for the registration between 
distorted image patches, is introduced to address geometrical 
distortions, due to effects such as foreshortening in the detected 
crater. The correlation value by caters of various sizes and 
shapes is illustrated in. 
  
  
  
Figure 6. First outline of crater rim and refinement by Hough 
transformation and verification with template 
(Black : verification range by hough and template matching, 
White : verified and refined crater MOC WA image 
M1103889) 
  
  
  
Ev=426 
Corr-0.86 
Reject 
  
Ev=142 
Corr-0.68 
  
Ev=345 
Corr=0.62 
Reject 
  
Ev=49 
Corr=0.83 
  
Ev=146 
Corr=0.72 
  
  
  
  
  
  
Template Original Re-sampled Ev :eigenvalue 
image patch | image patch Corr : cross- 
by Gruen correlation 
process 
  
  
Figure 7. Correlation value by verification with template 
819 
2.4 Crater detection based on DTM 
In some cases, the detection results on the optical image are 
poor, because the crater rim arcs are too short, which result 
from erosion, compared with their radii. To compensate for this, 
a DTM based crater detection is introduced. 
It is much simpler than the corresponding optical image case. 
The focusing method is replaced with a high slope area 
extraction. Instead of using the local edge of the optical images, 
ridge points from a gridded DTM (Wood, 1996) are used for the 
ellipse fitting. The big impact craters, which are not detected in 
optical images due to insufficient robustness of the edge linking 
method, can be easily identified here. 
3. RESULTS & ASSESSMENTS 
Final products are evaluated by visual inspection and 
quantitative assessments are made through comparisons with 
MCC and manually detected crater ellipses. 
3.1 Detection result on DEM 
Crater detection is performed with a MOLA DTM, gridded at 
256 m /pixel at the equator, which is shown in Figure 8. One 
characteristic of such DTM crater detection results is a high 
detection ratio for big craters with radii>15 km, even though the 
impact craters with small radii(< 4-5km) are usually not 
detected. Therefore it is highly complementary to the weakness 
of the detection results for optical images. 
     
        
(a) E99.226-101.58°N  (b) E 119.476-121.828° N 
23.375-25.72° 20.828-23.164° 
Figure 8. Crater detection on MOLA DTMs 
3.2 Detection results on optical image 
— 
Several examples of crater detection evaluation are shown in 
Figure 9. The detection ratio of relatively small impact craters 
(8<R<about 60 pixel) is excellent but large or multi-ringed 
structured crater show relatively poor detection accuracy. 
For quantitative assessment, quality asscssment factors (Shufelt 
& McKeown, 1993), originally developed for building 
detection work, are introduced as follows: 
Detection Percentage = 100 TP /(TP+FN) 
Branching Factor = FP / TP (7) 
Quality Percentage = 100TP / (TP + FP + FN)