Xenarch

Data for Xenarch, an unsupervised anomaly detector for planetary imagery. The model sees only natural terrain (natural-terrain_training/); the places where it should raise a flag are the known human-made hardware on the Moon and Mars: landers, rovers, experiment packages, crash and rocket-stage impact craters (human-hardware_ground-truth/). Keep the two apart: natural-terrain_training/ is training data, human-hardware_ground-truth/ is the test. human-traces_labels/ turns that ground truth into a labelled classification set: every trace humans left (landers, rovers, rover tracks, astronaut footpaths, experiments, impact craters, landing hardware, blast zones) as a chip with its class, plus natural-terrain chips of the same sites before arrival. synthetic-injections_lroc-nac_hirise/ is synthetic test data: natural terrain with one rendered or pasted hardware object, to measure how often a detector finds an object of a given size, sun angle and terrain (never use it as real ground truth or for training Xenarch).

Every folder and file name says what it holds and which camera took it (lroc-nac = LRO Narrow Angle Camera, Moon; hirise = MRO HiRISE, Mars), so a folder copied out of the dataset still explains itself.

natural-terrain_training/                      natural terrain, no hardware (training + validation)
  moon_lroc-nac/
    metadata.csv
    moon_<landform>_lroc-nac/*.png             445 LROC NAC tiles, 1024×1024, 0.4–1.8 m/px
  mars_hirise/
    metadata.csv
    mars_<landform>_hirise/*.png               384 HiRISE tiles, 1024×512, ~0.5 m/px
  moon_chandrayaan2-ohrc/
    metadata.csv
    moon_cratered_highlands_chandrayaan2-ohrc/*.png
                                               300 Chandrayaan-2 OHRC tiles, 1024×1024, 0.23-0.26 m/px: the
                                               sharpest lunar orbital imagery (three calibrated scenes, 62-74 S)
  legacy_mixed-cameras/                        the original mixed set (formerly xenarch-imagery)
    metadata.csv
    <body>_<camera>/*.png                      984 full images: moon_clementine, moon_lroc-nac-wac,
                                               moon_chandrayaan2-tmc2, mars_ctx, mars_hirise
human-hardware_ground-truth/                   known human-made objects (ground truth)
  catalog_moon-mars_lroc-nac_hirise.csv        THE table: 4,633 rows, one per object × orbital frame (100 objects)
  boxes-coco_lroc-nac_hirise.json              bounding boxes around the hardware in the crops (COCO); crops where
                                               the hardware could not be identified carry `unboxed_reason`
  crops_lroc-nac_hirise/                       527 crops of 2048×2048 around 95 objects
    metadata.csv
    moon_lroc-nac/<site>_lroc-nac/<frame>_<site>.png
    mars_hirise/<site>_hirise/<frame>_<site>.png
  derived/                                     views of the catalog, rebuilt with it
    sites_moon-mars.csv                        100 objects: coordinates and their source, arrival, best frame
    frames_lroc-nac_hirise.csv                 every orbital frame covering each object
    rover-frames_mars_hirise.csv               HiRISE frames that caught Curiosity / Perseverance where they were that sol
human-traces_labels/                         every labelled human trace as a 256×256 chip, by class
  labels_moon-mars_lroc-nac_hirise.csv         THE label table: one row per chip, `label` = class
  chips_lroc-nac_hirise/
    metadata.csv
    moon_lroc-nac/moon_<class>_lroc-nac/<object>_<frame>.png
    mars_hirise/mars_<class>_hirise/<object>_<frame>.png
  contact-sheets_lroc-nac_hirise/              one preview sheet per class (box outlined) + overview
synthetic-injections_lroc-nac_hirise/        SYNTHETIC test data: natural terrain + one injected hardware object
  injections_moon-mars_lroc-nac_hirise.csv     THE table: one row per injected image (what was injected, where, how)
  boxes-coco_synthetic_lroc-nac_hirise.json    the same annotations for the whole set in one COCO file
  realism-report_lroc-nac_hirise.csv           every rendered design against every real frame of that hardware
  natural_lroc-nac_hirise/                     the natural terrain alone, by what it shows
    metadata.csv
    moon_lroc-nac/moon_<landform>_lroc-nac/moon_<landform>_natural_<n>.png
    mars_hirise/mars_<landform>_hirise/mars_<landform>_natural_<n>.png
  injected_lroc-nac_hirise/                    1,431 images of 512×512: the same terrain with the object, by what was added
    metadata.csv
    moon_lroc-nac/moon_<class>_lroc-nac/moon_<kind>_<object>_<n>.png     kind: rendered, real-sprite
    mars_hirise/mars_<class>_hirise/mars_<kind>_<object>_<n>.png
  masks_lroc-nac_hirise/                       pixel masks, same sub-paths as injected: 1 object, 2 its cast shadow, 3 context
  annotation-synthetic-injection_lroc-nac_hirise/   one COCO file per image, same sub-paths as injected
  contact-sheets_lroc-nac_hirise/              real-vs-model (start here), zoomed galleries, designs, clean | injected pairs per body and kind
raw/legacy_source-files/                       what the legacy set was made from (not needed for training)
  <body>_<camera>_originals/                   Clementine JPEGs, HiRISE .IMG EDRs as downloaded
  <body>_<camera>_pds-labels/                  one PDS label (JSON / XML) per image, summarised in the legacy metadata.csv
  squares_mixed-cameras/                       1,240 fixed-size chips cut by the first pipeline
  manifest.csv                                 download log of the first pipeline

All images are 8-bit grayscale PNG at native resolution (no resampling; only label chips of boxes larger than ~190 px are resized, chip_scale > 1), each stretched from its 0.5–99.5 percentile. Scripts that build every file live in the Project Prometheus repository (xenarch_lroc_nac_dataset.py, xenarch_hirise_dataset.py, artifact-site-catalog/; xenarch_hf.py publishes them under these names).

Columns

The same name means the same thing in every table:

Column	Meaning
file_name	image path relative to the folder of the table (crops: relative to crops_lroc-nac_hirise/)
split	train or validation; fixed per source frame (a hash of frame_id, ~10% validation), so tiles of one frame never straddle the split
body, instrument	Moon / Mars; LROC NAC, HiRISE RED, …
landform	terrain class, = the folder name's middle part
frame_id	archive product (LROC M…LC) or observation (HiRISE ESP_…) the image was cut from
lat, lon_east	degrees, planetocentric, east-positive 0–360
incidence_deg	sun angle from vertical (0° = sun overhead; 90° = sun at the horizon)
emission_deg, phase_deg	viewing angle from vertical; sun–surface–camera angle
resolution_m_px	metres per pixel
utc_start, solar_longitude_ls	acquisition time; Mars season (Ls)

natural-terrain_training/

moon_lroc-nac (LROC NAC CDR, calibrated): six landform folders, moon_fresh_craters_lroc-nac, moon_mass_wasting_lroc-nac, moon_tectonic_lroc-nac, moon_volcanic_lroc-nac, moon_chains_terrain_lroc-nac, moon_polar_lroc-nac. Targets come from the IAU gazetteer; each was imaged under several sun angles; tile 0 is centred on the target and tiles 1–3 are its neighbours across the strip, so lat, lon_east are the target's coordinates (target is its name). Anything within 10 km of a landing or impact site was skipped. Also line0, sample0 (tile origin in the source image) and source_url.

mars_hirise (HiRISE EDR, raw bin-2 RED channels, odd columns repaired): six landform folders, mars_aeolian_hirise, mars_slopes_hirise, mars_periglacial_ice_hirise, mars_polar_hirise, mars_impacts_hirise, mars_layers_volcanic_hirise, chosen from the HiRISE team's own rationale per observation (rationale) and spread over sun angles. File names end with the sub_landform (dunes, gullies, …). lat, lon_east are the observation centre; channel is the CCD channel. Observations mentioning hardware or within 15 km of a lander, rover or crash site were skipped.

Neither set has been checked tile by tile for hardware.

legacy_mixed-cameras is the first-phase training set, formerly xenarch-imagery, one folder per body and camera. Mixed bodies, instruments and resolutions (Clementine ~100–200 m/px, CTX ~6 m/px, HiRISE bin 1–4), no landform labels, no hardware exclusion. metadata.csv is summarised from the PDS labels in raw/: empty where the archive marks a value missing (most Clementine angles); HiRISE position, angles and time come from the HiRISE RDR index, its resolution is nominal (0.25 m × binning). Mk15 trains on it; prefer moon_lroc-nac/ and mars_hirise/ for new work.

human-hardware_ground-truth/

catalog_moon-mars_lroc-nac_hirise.csv is the primary table: one row per object per orbital frame that contains it, with the object (id, name, type, mission, coordinates and their uncertainty, arrival date), the frame (sun and viewing angles, resolution, usable, after_arrival, archive URLs), the crop cut from it (crop_file relative to crops_lroc-nac_hirise/, crop_x, crop_y) and its box if any. Frames taken before the object arrived are kept (after_arrival = False): the "before" half of before/after pairs. usable = after arrival, sunlit (incidence < 90°) and resolution ≤ 2 m/px.

Objects (78 Moon, 22 Mars; positions from the LROC team's anthropogenic-objects table, 2026-08-14, unless noted):

• Moon, landed: Apollo 11–17 LMs, ALSEPs, seismometers, retroreflectors, LRVs, Apollo 17 SEP; Surveyor 1/3/5/6/7; Luna 16/17/20/21/23/24, Lunokhod 1/2; Chang'e 3/4/5/6, Yutu, Yutu-2; Chandrayaan-3 Vikram and Pragyan; SLIM; IM-1 Odysseus, IM-2 Athena; Blue Ghost 1.
• Moon, impacts: Apollo 13–17 S-IVB stages, Apollo 12 LM ascent stage, Surveyor 3/5/6 retrorockets, Ranger 6–9, LCROSS Centaur (permanent shadow: no NAC frame shows it), GRAIL Ebb and Flow, LADEE, Beresheet, Longjiang-2, Chandrayaan-2 Vikram, Hakuto-R 1 and 2, Luna 25, the 2022 rocket-body double crater, the 2026 Falcon 9 upper stage (no archived frame after the impact yet).
• Moon, candidate positions only (object says so): Apollo 14 LM ascent stage, SMART-1, Chang'e 1, two rival Luna 9 sites ~28 km apart. Yutu-2's position is in the Chinese mission frame (±450 m).
• Mars: Phoenix; Curiosity and Perseverance (landing site and rover per sol); Perseverance heat shield, backshell, descent stage and Ingenuity; InSight with heat shield and backshell; Viking 1/2, Pathfinder (radio tracking, ±300 m); Spirit and Opportunity final positions; Beagle 2; Tianwen-1 lander. Schiaparelli, Zhurong and the Mars 3 candidate have only the centre of a HiRISE image aimed at them (±1–3 km).

Not in the catalog (no published position): Luna 13; Apollo 15/17 LM ascent stages; Kaguya, Chandrayaan-1 MIP, Lunar Prospector; Chang'e 5/6 ascenders; Curiosity's and Phoenix's heat shields and backshells; Schiaparelli's and Tianwen-1's parachutes and shields; Spirit's and Opportunity's landing hardware.

crops_lroc-nac_hirise/: 2048² crops around each object, label = artifact or before_arrival (same ground before the object arrived). Moon crops are in raw NAC geometry, Mars crops are map-projected. obj_x, obj_y is the catalog's nominal pixel, not the measured one: use the boxes file for that. Objects added to the catalog later may have no crop yet (empty crop_file).

boxes-coco_lroc-nac_hirise.json: COCO file. images lists every crop (file_name relative to crops_lroc-nac_hirise/, plus frame metadata); annotations holds one bbox [x, y, w, h] per object per crop, pixels in the 2048² crop, origin top-left; categories has one entry per object (name = site_id of the catalog). Tracks are categories of their own (<mission>_footpaths, <rover>_tracks, not in the catalog) with one box per visible segment. Boxes include the object's shadow. confidence is high (unambiguous: bright body with hard shadow, tracks, or matches a published labeled image), medium (right size and shadow at the predicted position) or low (could be a rock). note says how the box was found. All six Apollo sites are boxed; most small Soviet and Chinese landers and most Mars rover frames are not (their image record's unboxed_reason says why).

derived/ holds views of the same data, rebuilt with the catalog; edit nothing there by hand. sites_moon-mars.csv (one row per object, with the source of its coordinates), frames_lroc-nac_hirise.csv (every frame per object; obj_sample_frac / obj_line_frac give the object's position in a raw NAC frame only approximately, off by up to ~1,300 px), rover-frames_mars_hirise.csv (object_lat, object_lon_east: where the rover was on that sol, from NASA/JPL waypoints).

human-traces_labels/

One 256×256 chip per labelled trace, cut from the 2048² crops around each box of the COCO file (boxes larger than ~190 px are cut wider and resized, see chip_scale). The same object seen at the same spot through two crops of one frame gives one chip. Classes (label, = the folder name's middle part):

Class	What
lander	crewed and robotic landers, Moon and Mars
rover	rovers and the Apollo LRVs
rover-tracks	wheel tracks (Apollo 15/16/17 LRV, Curiosity)
astronaut-footpaths	trails of footprints (Apollo 12/14/16/17). Single footprints (~0.3 m) are below the resolution of every camera here
surface-experiment	ALSEP, seismometers, laser retroreflectors, Apollo 17 SEP
impact-crater	crash, rocket-stage, ascent-stage and retrorocket impact craters
landing-hardware	heat shields, backshells with parachutes, the Perseverance descent stage
landing-blast-zone	ground scoured by descent engines (Curiosity, Perseverance)
natural-terrain	negatives: the same sites in frames taken before arrival (centre + four quadrants of each crop)

Columns besides the common ones: object_id (catalog site_id, or the track category), object, object_type (catalog type), mission, location_id (objects within 10 km share one), crop_file (the source crop, path in this dataset), chip_x0, chip_y0, chip_side_px (where the chip sits in the crop, crop pixels), chip_scale (crop px per chip px), box_x_min … box_y_max (the labelled box in chip pixels; empty for natural terrain), box_confidence, includes_shadow, note (how the box was found), source_url. resolution_m_px is the chip's own pixel scale. split is fixed per location, so no terrain is in both halves: the hash split of the other tables, plus one validation location for every class seen in two or more locations. Tracks are few (one to three segments per site where they are unmistakable); treat medium boxes as likely, not certain.

synthetic-injections_lroc-nac_hirise/

Injection-recovery test set, built by xenarch_inject.py. Every row of the table is a pair of 512×512 images: natural_file, a crop of a natural-terrain tile (background_file, a path in this dataset), filed by body and landform; and file_name, the same crop with one object added, filed by body and object class. The two end in the same number. The natural image has the same gain, gamma and resampling as its injected twin, so the only difference between them is the object. mask_file marks the object (1), its cast shadow (2) and everything else that changed (3: blast zone, tracks, ejecta, faint fringe); no pixel outside the mask differs by more than 2 DN. annotation_file is a complete COCO file for that image: category, bbox of the object with its shadow in pixels [x, y, w, h], bbox_object_only, segmentation polygons, plus the object's name and size. Category ids are the same in every file and in boxes-coco_synthetic_lroc-nac_hirise.json. Two kinds:

kind	What	Images
rendered	past and new hardware built from parts at real dimensions (raised bodies, tapering upper structures, masts, panels, wheels), lit by the tile's own sun	949
real-sprite	real hardware cut out of the LROC / HiRISE crops (high-confidence boxes): its bright body and the shadow touching it, without the halo, turned by its own body-to-shadow direction onto the tile's sun, only shrunk, only into tiles whose sun incidence is within 12° of the source frame's	482

Rendered designs (hardware_id, era): Moon past: Apollo LM descent stage, Surveyor, Luna 16-24, Lunokhod, Apollo LRV, Chang'e 3/4 lander, Yutu, a spacecraft impact crater; Moon new: Nova-C, Blue Ghost, Blue Moon MK1, Starship HLS, a surface habitat, the Lunar Terrain Vehicle, a vertical solar array, a fission power unit; Mars past: Viking, Pathfinder, Phoenix / InSight, Spirit / Opportunity, Curiosity / Perseverance, a heat shield, a backshell with its parachute; Mars new: a habitat dome, the Sample Retrieval Lander, a pressurised rover.

The renderer is calibrated on the real hardware: realism-report_lroc-nac_hirise.csv and contact-sheets_lroc-nac_hirise/real-vs-model_lroc-nac_hirise.jpg compare every design with every real frame of that hardware at the same sun and m/px, measured by the same code. The tiles are stretched 8-bit, so one unit of ground radiance is far more than the DN of the ground (radiance_scale_dn, estimated from the tile's contrast and sun); composited in radiance, then blurred, then clipped, with bright parts rolling off instead of turning into flat white shapes (body_contrast_dn). Shadows are swept from each part at its own height, so they taper and detach, and bend over the terrain (relief from the tile's shading); their edge width is tuned on real hardware and natural shadows, and they are as dark as the natural shadows of their own terrain (shadow_soft_px, shadow_floor_dn). Rendered objects are softer than real hardware (no saturated bodies, lighter shadows, shorter shadows), so recall measured on them is a cautious figure. Heights match the real shadows (Apollo LM: 3.5 m). Objects stand on flat-ish lit ground, not on crater walls or in shadow.

What varies, all in the table: object size (scale_factor 0.92–1.08: hardware has fixed dimensions; impact craters 10–35 m, width_m), resolution (resolution_m_px: the tile's own, and on the Moon a coarser camera, camera_downscale 1.5 or 2), sun (incidence_deg, sun_azimuth_image_deg of the tile), brightness (albedo_factor, contrast_factor), camera (psf_sigma_px measured on the tile, its noise_sigma on the object, gain, gamma on both twins), terrain (landform), context (context, +-joined: blast-zone around landers, impact-scar around heat shields and backshells, tracks behind rovers, footpaths around crewed sites; impact craters always bring ejecta). object_px is the object's size in pixels; shadow_length_m its longest shadow; shadowed_fraction the share of the crop in shadow. A design only goes into tiles where its shadow stays under 24 px, the upper range common on real hardware (height × tan(incidence) / resolution; shadow_length_px is the reach measured on the mask along the sun, from the middle of the body, shadow_area_px its area), so tall designs appear only at high sun and Starship (50 m) only on the Moon. Faint HiRISE tiles with fewer than 20 grey levels are not used as backgrounds.

split is the background tile's: evaluate detectors trained on natural-terrain_training/ on validation only (backgrounds never seen in training; validation tiles carry extra injections for that reason).

Limits: the sun azimuth in a raw frame is computed from latitude, incidence and season up to its mirror images, and the right one picked with the tile's shading (sun_azimuth_source; shading and geometry agree within a median of 5° on the Moon, 16° on Mars). radiance_scale_dn is an estimate. Rendered objects are simple parts seen from straight above; new hardware has no real frame to check against, only the same rules. Real cut-outs carry the lighting of their source frame. Synthetic data: use it to measure sensitivity, never as ground truth.

Credit

Moon: NASA / GSFC / Arizona State University (LROC). Mars: NASA / JPL / University of Arizona (HiRISE). Legacy set: NASA PDS (Clementine, MRO CTX / HiRISE, MGS MOC), ISRO (Chandrayaan-2 TMC-2). Source data are public domain; this compilation is CC0.