Semantic segmentation of light-toned veins in multimodal ChemCam data

Since the Mars Science Laboratory landed in 2012, the ChemCam instrument aboard the rover has collected in-situ laser-induced breakdown spectroscopy (LIBS) data and context images along more than 35 km of the Gale Crater traverse, providing valuable observations including diagenetic features such as light-toned veins. These veins are of particular scientific interest because they are interpreted as indicators of past fluid circulation on Mars and provide insights into the evolution of habitability on Mars. Their identification, however, currently relies on manual visual inspection of Remote Micro Imager (RMI) images, a process that is time-consuming and sensitive to differences in human interpretation. To address this issue, in this paper we introduce a novel pixel-level labeled, multimodal dataset of ChemCam observations specifically tailored for vein detection, along with customized U-Net models to integrate both textural (RMI) and chemical (LIBS) modalities. To further ensure trustworthy scientific use, we incorporate the Learn-Then-Test (LTT) framework to provide statistical control of the false discovery rate without requiring model retraining. The experimental results demonstrate that the proposed customized U-Net models trained on the developed dataset, combined with risk-controlled prediction, increases the efficiency of pixel-level vein identification through automation and produces statistically reliable predictions for multimodal ChemCam data.