Machine learning approaches for classifying lunar soils

Global Mapping of Fragmented Rocks on the Moon with a Neural Network: Implications for the Failure Mode of Rocks on Airless Surfaces

“…Similar (global-scale) machine-learning-driven searches were previously performed by Bickel et al (2020a) and Bickel et al (2022). In addition to mapping applications, earlier work used machine learning for global-scale analyses of the physical surface properties of the Moon and asteroids, such as performed by Cambioni et al (2019), Kodikara & McHenry (2020), and Moseley et al (2020). Our neural network was trained on images that cover the full range of known block degradation states, including boulders with single fracture, heavily fractured boulders, as well as catastrophically shattered boulders, specifically excluding large, intact boulders.…”

Using machine learning to reduce observational biases when detecting new impacts on Mars

“…Machine learning methods are increasingly employed to assist in the analysis of remote sensing data for planetary science and exploration. Examples include the classification of lunar soils using reflectance spectra (Kodikara and McHenry, 2020), the classification of terrain types in Mars orbital images to inform landing site selection (Ono et al, 2016;Barrett et al, 2022), the classification of rover and orbital images to enable content-based search of the Planetary Data System image archives (Wagstaff et al, 2018(Wagstaff et al, , 2021 or map features such as volcanic rootless cones and transverse aeolian ridges (Palafox et al, 2017) and rockfalls on Mars (Bickel et al, 2020b) and the Moon (Bickel et al, 2020a), and the identification of novel features in rover images to accelerate discovery (Kerner et al, 2020). 2021) adapted a CNN that was trained on large craters in THEMIS data to apply to smaller craters in CTX images and used it to determine crater populations and infer the age of the ejecta blankets of ten large craters on Mars.…”

Application of deep learning and spectral deconvolution for estimating mineral abundances of zeolite, Mg-sulfate and montmorillonite mixtures and its implications for Mars

“…A review of deep learning and its evolution can be found in LeCun et al (2015) and Schmidhuber (2015), respectively. Kodikara and McHenry (2020) examined the ability of ML algorithms, adopting 9 ML algorithms representing linear, non-linear, classification trees, and rule-based models, to determine the physical and mineralogical properties of lunar soil using the reflectance spectra of lunar soils.…”

“…Feature selection is an effective way to identify the important features in a dataset and discard others as irrelevant and redundant. Irrelevant and redundant features in the training dataset can result in highly unstable models and poor performance (e.g., Kodikara and McHenry, 2020).…”