How do training sets influence crater and boulder detection in machine learning?

The surfaces of planetary bodies reflect their evolution through primary surface shaping via their continuous evolvement over time. Surface formation and degradation processes need to be understood in detail to infer the timescales over which these processes operate.

Planetary surfaces which are heavily cratered offer the opportunity to investigate various aspects of the cratering processes which are initiated when an impactor strikes their surface and ejects rock fragments from the impact point upon the newly-formed crater cavity and its surroundings (e.g. Hörz, F. and Cintala, M., 1997). Among the ejecta material from the impact are boulders covering a wide range of sizes (e.g. Nagori,R. et al., 2024). Dependent on the planet’s environment and the size of the impact fragments, these boulders can form either secondary craters or simply become subject to the various environmental forces which ultimately add through different degradation processes to the formation of planetary regolith. To understand many of the aspects of the above processes, size distributions of both the impact-generated boulders and secondary craters need to be understood (e.g. Cadogan, P., 2024).

As many of the techniques to identify boulders and small craters on albedo images are using shadow-based identification methods one has to be aware that ambiguities can arise through complex topographies and overlapping surface features. These factors can modify the shape of the shadow and make the identification of its borders difficult, thereby preventing a precise determination of both it’s location and it’s radius.

To obtaining high-quality statistics for boulders and craters over large and varied planetary surfaces, machine learning and deep learning methods have been applied to automate the tedious human based detection work (e.g. DeLatte, D. et al, 2019). However, little attention has been paid to investigate the influence of the training sets on the success rates of these efforts (Mall, U. et al., 2023). We are investigating in this study the influence of crater training sets, originating from specifically chosen lunar areas on the resulting confusion matrices produced by specific convolution neural networks and compare these with the results found from traditional imaging methods.

Cadogan, P., (2024), Automated precision counting of small lunar craters - A broader view, Icarus, Volume 408, 2024,115796.

DeLatte, D.M., Crites, S.T., Guttenberg, N., Yairi, T. (2019), Automated crater detection algorithms from a machine learning perspective in the convolutional neural network era, Advances in Space Research, Volume 64, Issue 8, Pages 1615-1628.

Hörz, F. and Cintala, M. (1997), The Barringer Award Address Presented 1996 July 25, Berlin, Germany: Impact experiments related to the evolution of planetary regoliths. Meteoritics & Planetary Science, 32: 179-209. https://doi.org/10.1111/j.1945-5100.1997.tb01259.x.

Mall, U., Kloskowski, D., Laserstein, P., (2023), Artificial intelligence in remote sensing geomorphology—a critical study, Front. Astron. Space Sci., 30 November 2023, Sec. Planetary Science , Volume 10 – 2. https://doi.org/10.3389/fspas.2023.1176325.

Nagori,R., Dagar, A. K., Rajasekhar, R.P., (2024),  Age estimation and boulder population analysis of the West crater at Apollo 11 landing site using Orbiter High Resolution Camera on board Chandrayaan-2 mission, Planetary and Space Science, Volume 240, 2024, 105828.