Optimized Martian Dust Displacement Detection Using Explainable Machine Learning

Abstract

The ChemCam instrument on the Curiosity rover performs geochemical analyses of rocks on Mars using Laser-Induced Breakdown Spectroscopy (LIBS). The shockwaves generated during the LIBS measurements sometimes shift dust from the surface of the target. The study of the Martian dust phenomena in the scope of the ChemCam instrument has the potential to provide insight into the planet's geology and aid calibration methods for data processing. In this study, we develop a pipeline, named Dust Displacement Detection (DDD), for automatic detection of dust displacement on LIBS targets based on the image dataset acquired by ChemCam. To this end, we introduce a data pre-processing methodology and test two-stage models with a pretrained model in the first stage for feature extraction and a Random Forest classifier or a Support Vector Machine as a binary classifier in the second stage. The best performing model was found to consist of the first 10 layers of VGG16 and a Random Forest classifier, achieving 92% accuracy. Additionally, we use Explainable AI (XAI) methods such as Shapley values and guided backpropagation for model optimization. The experiments show potential for model optimization, and the application examples presented encourage discussion of machine learning in the field of Martian dust research.