Automatic Recognition of Impact Craters on the Martian Surface from DEM and Images - Planetary Exploration and Science: Recent Results and Advances - page 111

where TD is the number of true detections, FD is the number of false detections,
and GT is the number of ground truth. The ground truth served as the basis for a
quantitative assessment of the methodology.

6.4.1 Performance Evaluation

6.4.1.1 Recognition Results from Images

Four regions are chosen for the test in Bandeira et al. ( 2007 ). The average value of
the TDR is 86.57 %, which is an amazing result comparing to the ever published
results with automatic detection method. In region C, the method shows the best
performance in detecting craters maybe because of the relatively low number of
craters. However, four regions all show good consistency between each other even
though the number of GT is substantially different. The difference between them
may result from the slightly different geomorphological settings in regions.
Because of the good performances in face recognition field, the Adaboosting
approach has been introduced to detecting craters. Adaboosting is a machine
learning method, which is different from the unsupervised method, like template
matching. By using the series of weak classifiers from the training set, we can finally
build a strong classifier for the classification. The detections will be automatically
evaluated as true or false craters by comparison with the manually built truth. It is
clearly that the performance varies with the change of the threshold. When a lower
threshold is determined in the classification, higher true detection rate (TDR) and
false detection rate (FDR) are obtained at the same time.
In Tables 6.1 and 6.2 , we have shown the performances of two automatic crater-
detection methods within the planetary images. When the threshold is set to be
0.75, the modified Adaboosting method performs better than the template-matching
method.

6.4.1.2 Recognition Results from DEM

Since the resolution of DEM data on Mars cannot be obtained as high as the
planetary images, we can only detect some relatively large craters based on DEM

Tabl e 6. 1 Detection rates

Original method Improved method

Threshold TDR (%) FDR (%) TDR (%) FDR (%)

0.55 92.3 40.1 97.3 40.3

0.65 88.2 25.6 94.2 28.3

0.75 83.9 17.6 91.2 18.6

0.85 75.5 8.8 85.2 10.1