Speaker
Description
The growing volume of orbital imagery of the Moon has made manual inspection of the surface impractical, while visual interpretation of individual landforms remains subjective and hard to standardise. We present LunaX, a system combining a deep-learning detector with an interactive analytical environment, designed to pre-select regions of interest for an expert rather than to replace the expert.
The detection model is based on the YOLOv8 architecture and was trained on ~1500 image tiles of lunar craters retrieved from the LROC WAC mosaic through the USGS planetary WMS service. Bounding boxes were generated automatically from crater positions and diameters taken from the Lunar Impact Crater Database, then manually refined in Label Studio. Craters were assigned to six chronostratigraphic classes (Pre-Nectarian, Nectarian, Lower Imbrian, Upper Imbrian, Eratosthenian, Copernican), so the model learns morphological degradation as a proxy for relative age - both the strength and the main limitation of the approach.
Around this model we built LunaX, a web application (FastAPI, React, Leaflet) in which the user selects an area on a lunar grid, sets resolution, sampling strategy and confidence threshold, and runs inference on tiles downloaded on demand. Detections and image metadata are stored in Parquet and queried with DuckDB; map, gallery and list views allow results to be confirmed, rejected, tagged and commented on, and stored images can be re-analysed with different weights without repeated downloads. This human-in-the-loop workflow also turns the tool into a generator of curated training data.
The architecture is model-agnostic: the same pipeline can be extended to lava pits, rilles, scarps, boulders, chaotic terrain or pyroclastic deposits - features directly relevant to landing-site and ISRU assessment. The poster presents the data pipeline, system architecture and example detections.