Application of Automated Crater Detection for Mars Crater GIS Database Production
J. I. Simpson *, J. R. Kim **, J-P. Muller ** * Dept. of Civil, Environmental and Geomatic Engineering, UCL, London WC1E 6BT, UK.** MSSL, Dept. of Space and Climate Physics, UCL, Surrey, RH5 6NT, UK.
MotivationApproachPerformance Assessment
Detection Percent = (100 * TP) / (TP + FN)Quality Percent= (100 * TP) / (TP + FN + FP)Branching Factor = FP / TP
Assessment Results
•
Craters from automated detectionconverted into polygon shapefiles
•
Results merged
•
False positives removed
•
False negatives added
•
Duplicates in overlap regions resolvedby identifying closest matching pair ineach set via distance measurement factor(incorporating radii and distance betweencentres)
•
Georeferenced by defining projectedcoordinate system in projection file(optional part of shapefile format)
GIS Database Production
8,857 Craters in Iani VallisGreen = TPRed = FPBlue = FNYellow = Duplicate
Conclusions
Image TP FP FN TotalDetection %
FN TPTP
+=
*100
Quality %
FPFN TPTP
++=
*100
BranchingFactor
TPFP
=
Elysium 1196 142 17 21 180 87.12 78.89 0.12Elysium 2066 168 11 30 209 84.85 80.38 0.07Elysium 2099 369 8 33 410 91.79 90.00 0.02Elysium 2110 105 2 4 111 96.33 94.59 0.02Elysium 2121 65 10 32 107 67.01 60.75 0.15Elysium 2143 366 26 47 439 88.62 83.37 0.07Elysium 2154 346 9 47 402 88.04 86.07 0.03Elysium 2165 143 28 25 196 85.12 72.96 0.20Elysium 2176 237 30 54 321 81.44 73.83 0.13
Subtotal 1,941 141 293 2,375 85.59 80.09 0.09
Iani 0912 2,592 159 163 2,927 94.08 88.55 0.06Iani 0923 2,131 229 710 3,070 75.01 69.41 0.11Iani 0934 3,313 305 785 4,403 80.84 75.24 0.09
Subtotal 8,036 693 1,658 10,400 83.31 77.74 0.09Total 9,977 834 1,951 12,775 84.45 78.92 0.09
MSSL/DEPARTMENT OF SPACE AND CLIMATE PHYSICS
GIS Files
•
Best case: 157 craters detected in overlap
•
Worst case: 45 craters detected in overlap2,543 Craters inElysium PlanitiaGreen = TPRed = FPBlue = FNYellow = Duplicate
2 regions on Mars were chosen for their geological diversity,Elysium Planitia and Iani Vallis.
©USGS
ElysiumPlanitiaIaniVallis
•
Demand from planetary geologists forimpact crater databases
•
As spatial resolution of imageryincreases, volumes become too large formanual digitisation
•
Accuracy of automated algorithms mustbe quantified and improved
•
Fully automated crater detectionsystems not yet available
•
Aim is to emulate a fully automatedsystem to demonstrate feasibility andperform a quantitative assessment ofexisting automated system
•
Process HRSC images using Kim-Muller** automated crater detectionalgorithm
•
Calculate the detection and qualityrates for each image
•
Automatically merge results into single,georeferenced GIS format shapefiles
•
As part of the merging process,automatically resolve duplicatedetections for side overlapping orbits
•
A software tool was developed to aid aquantitative assessment, optimised forsimplicity and speed
•
Two regions on Mars were selected andthe HRSC images were processed usingthe Kim-Muller algorithm:
•
Using the tool, craters from theautomated detection were rapidly taggedas true positives (TP), false positives (FP)or false negatives (FN)
•
For each set of detection results,Shufelt’s metrics, originally designed forbuilding extraction from digital imagerywere calculated as follows
•
Image quality significantly affects detectionresults
•
Detections were reduced by a factor of 3 for theworst case compared with the best case
•
It is possible to perform a quantifiableassessment of automated crater detectionalgorithms in the absence of existing groundtruth databases
•
The construction of a fully automated craterdetection system is achievable
•
It is conceivable that automated crater detectionalgorithms will be improved sufficiently to thepoint where they become a useful tool especiallyif DTMs are included
Summary
Detection %84.45Quality %78.92Branching Factor0.09for crater diameters >= 8 pixels when averaged across11,400 craters by visual inspection, in 2 different geologicalenvironments.
Detection Percent = (100 * TP) / (TP + FN)Quality Percent= (100 * TP) / (TP + FN + FP)
Duplicates between setsresolved using distance &radius measure
Leave a Comment