49th Lunar and Planetary Science Conference 2018 (LPI Contrib. No. 2083) 2202.pdf

AUTOMATED DETECTION OF CRATERS USING A CONVOLUTIONAL NEURAL NETWORK. G. K. Benedix1, C. J. Norman2, P. A. Bland1, M.C. Towner1, J. Paxman2, and T. Tan2. 1Dept. Applied Geology, Curtin University, GPO Box U1987, Bentley, Perth Western Australia, 6845 Australia. 2Department of Mechanical Engineering, Curtin University, Kent Street, Bentley, Perth Western Australia, 6102. Australia.

Introduction: Impact craters on planetary bodies Next, we generalised the algorithm to higher- provide a wealth of information about the surface and resolution Context Camera (CTX) imagery to detect subsurface history of that body. The age of a surface craters from approximately down to 100m. We chose can be estimated through analysis of crater frequen- to study the vicinity of Mojave crater – suggested to be cies, assuming random impact rates with known long- a source crater of some Martian meteorites [7]. The term averages [1]. Considerable effort has been invest- crater itself has a diameter of 58km. It is located ed into developing automated techniques for the detec- (7.5˚N; 33.0˚W) close to the in the tion and counting of craters [2,3], but none have pro- northern hemisphere. It is at the confluence of major gressed to become a fully automated pipeline and de- related to Simud and Tiu Valles. It is fault tool in the analysis of planetary surfaces thought to be young because it is a rayed crater. This Approaches [3] to Crater Detection Algorithms crater provides a good test of the CDA to compare to (CDAs) include image analysis techniques such as the age determined by [7]. In addition, the crater size- edge detection and the Hough transform, although, in frequency distribution (SFD) is complex due to resur- recent years, some have also incorporated machine facing [e.g., 7, as well as others]. We wanted to test learning, including neural network architectures. whether our CDA could reproduce that morphology. We present here results of an automated machine learning solution to the crater detection and counting problem on , involving a Convolutional Neural Network (CNN), with ground truth (training) data pro- vided by an existing database of Martian craters with diameters over 1km [4,5]. Background: Our approach is detailed in [6], but briefly, we designed a CNN-based CDA using Ten- sorBox, an open-source object detection framework based on Google Tensorflow. USGS Astrogeology Thermal Emission Imaging System (THEMIS) mosa- ics covering the entire equatorial latitude band of N°30 to S°30 were selected for CDA training and analysis. Each mosaic image spans a region of approximately 2700 km by 1800 km. The mosaics were split into a total of 6387 tiles each with width 1280 pixels and height 960 pixels. Craters identified by [4] were then mapped to each tile. Results: Figure 1 compares the crater counts made using the CDA to the manual results of [4] for a small area within the Iapygia quadrangle on Mars (center latitude = -15˚; center longitude = 67.5˚E). The CDA is able to detect a broad range of craters that vary in size and appearance. The CDA picked 80% of the craters identified by [4] for that same region (note that repro- ducibility studies on manual crater identification find that it is only 85% accurate [5]). On closer inspection, it is notable that there are differences between the manual and CDA identification of craters. The CDA picked some craters that the manual identification Figure 1. One tile of the Iapygia quadrangle of Mars with missed and vice versa. Manual identification is 85% craters identified by A) [4]; B) CDA. Field of view is 32km. accurate as shown by [5]. The CDA achieves a level of Original is THEMIS mosaic of the Iapygia quadrangle accuracy comparable to manual identification. generated by the USGS Astrogeology Science Center. 49th Lunar and Planetary Science Conference 2018 (LPI Contrib. No. 2083) 2202.pdf

We created a mosaic of Mojave crater from CTX (2005). Photogrammetric Engineering & Remote Sens- images, which we then tiled into smaller (the same ing, 71(10), 1205-1217. [3] Salamunićcar G. and pixel sizes as for the training set) images on which the Lončarić S. (2012) Ch. 3 in Horizons in Science CDA was applied. These detections are shown in Fig- Research Volume 8. [4] Robbins, S. J. and Hynek, B. ure 2. M. (2012) Journal of Geophysical Research E: Plan- ets, 117, 1–18. doi: 10.1029/2011JE003966. [5] Rob- bins, S. J. et al. (2014) Icarus, 234, pp. 109–131. doi: 10.1016/j.icarus.2014.02.022. [6] Norman, C.J. et al. (2018) Planetary Science Information and Data Ana- lytics Conference Abstr #6002. [7] Werner, S. C., et al. (2014) Science. 343(6177), pp. 1343–1346.

Figure 2. CTX Mosaic of Mojave Crater with CDA detec- tions in red.

Figure 3 shows a crater size frequency distribution isochron generated through an automated workflow for comparison with earlier analysis by Werner et al. [7] based on manual counts. The crater frequencies are consistent with the results of Werner’s analysis in the 0.1-1km range. The shape of the isochron is similar, but there is some variation at the smaller diameter end. Conclusions: We have developed a Crater Detec- tion Algorithm, based on a Convolutional Neural Net- work and trained using THEMIS images and the Rob- bins database [4] for ground truth. The CDA can detect craters in a processed image in less than a second and is as accurate as manual identification. An analysis of crater frequencies around the Moja- ve crater is proof of concept of generalisation to higher resolution data and smaller craters and comparison to manual detections shows a similar isochron for the same size frequency range. Future work will further refine the algorithm, ex- pand the training set to include examples from a broader size range and even higher resolution datasets, and apply the CDA to larger regions on Mars and other planetary bodies.

Figure 3. a) CDA generated crater count statistics for Mo- References: [1] Hartmann, W. K., and Neukum, G. jave Crater only b) Figure 2b from [7]. Blue lines added to (2001). Space Science Reviews, 96(1-4), 165-194. doi: aid comparison between the data sets highlighting the di- 10.1023/A:1011945222010. [2] Kim, J. R., et al. ameters between 100m and 1km.