A Machine Learning Approach for the Autonomous Identification of Hardness in Extraterrestrial Rocks from Digital Images

Understanding rock hardness on extraterrestrial planets offers valuable insights into planetary geological evolution. Rock hardness correlates with morphological parameters, which can be extracted from navigation images, bypassing the time and cost of rock sampling and return. This research proposes a machine-learning approach to predict extraterrestrial rock hardness using morphological features. A custom dataset of 1496 rock images, including granite, limestone, basalt, and sandstone, was created. Ten features, such as roundness, elongation, convexity, and Lab color values, were extracted for prediction. A foundational model combining Random Forest (RF) and Support Vector Regression (SVR) was trained through cross-validation. The output of this model was used as the input for a meta-model, undergoing linear fitting to predict Mohs hardness, forming the Meta-Random Forest and Support Vector Regression (MRFSVR) model. The model achieved an R2 of 0.8219, an MSE of 0.2514, and a mean absolute error of 0.2431 during validation. Meteorite samples were used to validate the MRFSVR model’s predictions. The model is used to predict the hardness distribution of extraterrestrial rocks using images from the Tianwen-1 Mars Rover Navigation and Terrain Camera (NaTeCam) and a simulated lunar rock dataset from an open-source website. The results demonstrate the method’s potential for enhancing extraterrestrial exploration.