Speakers
Description
Automated grain size analysis from rover imagery is central to the scientific return of the ExoMars Rosalind Franklin Rover, whose PanCam and CLUPI cameras will image sediment at Oxia Planum within a mission timeline too short for manual measurement. We present the GRAINS project, which adapts a state-of-the-art deep learning segmentation model, originally developed and validated on terrestrial fluvial and glacial sediment, to the imaging conditions of these two instruments. A physical analogue setup was built using commercial cameras configured to match the sensor geometry, working distance and field of view of PanCam HRC and CLUPI, and used to image 21 Antarctic analogue samples under controlled illumination, with further Mars-analogue materials currently being added. Before any retraining, the terrestrial model was tested against more than 400 historical images from past Mars missions, exposing three recurring failure modes linked to grain resolution, size heterogeneity and morphological variability. A first domain-adapted model, fine-tuned to address these failure modes, was benchmarked on a held-out Mars-like test set and outperformed the original terrestrial model across every segmentation and granulometry metric evaluated. We discuss the implications of this transfer learning approach for deploying automated sediment analysis tools on future planetary missions, and outline the ongoing dataset expansion toward a validated, mission-ready GRAINS model.