ABRUBAB/shadowastra-atlas

ShadowAstra Atlas: every land area on Earth, measured as a Moon or Mars practice site

Where on Earth can a team rehearse a Moon base or a Mars mission? We measured all 80,206 land areas on Earth (0.5° cells, about 55 km wide) with open NASA data, the same way we measured 18 Moon and Mars targets, and scored every place for every target. This dataset is the measured planet and every test behind the ShadowAstra Atlas website (the Caves and lava tubes layer, caves.json in the code repository, is not included), built by Team ShadowAstra (Dhaka, Bangladesh) for the NASA Space Apps Challenge 2026: Identify Earth Locations that Analog the Permanent Moon Base Locations and Mars.

• Live site: https://shadowastra-atlas.vercel.app
• Code and pipeline: https://github.com/ABRUBAB/ShadowAstra_Atlas (pipeline/s21_open_dataset.py rebuilds these files)
• Every test, explained: https://shadowastra-atlas.vercel.app/methods

Files

File	Rows	What it holds
places.parquet	80,206	One row per land area: names, measurements in real units, NASA EMIT minerals, slope histogram, 0 to 1 indicators, the learned Moon and Mars scores and a confidence level
target_scores.parquet	1,443,708	One row per land area and target: the match score as the website shows it for the target's main job, and each factor behind it
targets.csv, targets.json	18	The Moon and Mars targets: terrain at 50 km, main rocks, published water-made minerals, LRO Diviner, Lunar Prospector and Mars Odyssey measurements
ai_twins.parquet	648,000	The 5 closest Earth places for every 1° cell of the Moon and Mars, from our terrain AI
known_sites.csv	38	Analog sites used to train the learned model, with the jobs they were used for
validation_sites.csv	28	Published analog sites never used for training, kept as an untouched test
validation.json		Every test result shown on the How it works page

Join target_scores and ai_twins (earth_cell) to places on cell.

Key columns

places

• cell, lat, lon: the centre of the 0.5° cell. country, region, nearest_town, nearest_town_km from Natural Earth.
• Terrain over a 50 km square (SRTM via AWS Terrain Tiles): elevation_m, height_range_50km_m (2nd to 98th percentile), mean_slope_deg, slope_share_<a>_<b>_deg (share of ground with slopes between a and b degrees).
• Climate (NASA POWER, MERRA-2 climatology): air_temp_*, rain_mm_per_year, humidity_pct, sunshine_kwh_m2_day.
• Surface (NASA GIBS, Terra MODIS 2024): ground_temp_hottest_day_c, ground_temp_coldest_night_c, plants_ndvi_max, bare_ground_share, snow_ice_share.
• Context: volcano_km and nearest_volcano (Smithsonian GVP Holocene volcanoes), impact_crater (a selection of 62 confirmed craters from the Earth Impact Database), rock_type_glim (GLiM dominant class; an input to the learned model and shown in reports, not part of the rocks_craters factor).
• emit_<clay|sulfate|carbonate|iron>_share: mineral share of the ground from NASA EMIT L3 (0.5°, v002). EMIT sees dry, bare land only: it is empty where it has no clear view (plants, water, snow, cloud, towns, or beyond ±52°), so only 16,868 of the 80,206 places have values.
• indicator_*: the 0 to 1 transforms the scores use (dry, bare, cold, swing, airdry, hot, volcanic, impact, ice, remote, rugged).
• learned_moon, learned_mars: percentile rank (0 to 1) from the learned model, a rank-average of tuned XGBoost, tuned LightGBM and logistic regression trained positive-unlabelled on the 38 known sites. *_spread: how much its 24 single runs disagree. measured_share: share of measurements that are real, not gap-filled. confidence_*: high, medium or low from those two.
• known_site_within_150km, nearest_known_site_km: whether a known test site is nearby.

target_scores

• target, body, job (the target's main job, as the website picks it first).
• match_score = 40% factors_only + 60% learned, as the website shows by default.
• factors_only: the job-weighted mean of ground_shape (slope histograms and height range compared with the target at the same 50 km scale, LRO LOLA 118 m or MGS MOLA 463 m), rocks_craters, cold_heat, sunlight_angle (Sun-height match in the best month) and bare_dry_ground (little rain and few plants: loose, bare, dry ground, or regolith; the website calls this factor "Ice & loose ground", and ice counts under cold_heat).
• ground_shape_ai: the terrain AI's similarity rank for the target (an option on the website, not in match_score).

targets: temp_avg_c, temp_max_c, ice_surface_share, ice_1m_share within 10 km of the seven south-pole Moon targets (LRO Diviner polar resource product, Paige et al.); temp_avg_c, temp_max_c and temp_min_c (coldest, before dawn) over about 45 km for the three Moon targets away from the pole, from the day-and-night curve in Diviner's global temperatures (Williams et al. 2017, PDS LRO-L-DLRE-5-GCP); feo_wt_pct, tio2_wt_pct (oxides by weight) and th_ppm for every Moon target from the Lunar Prospector gamma-ray spectrometer (Prettyman et al. 2006, about 150 km footprint); h2o_wt_pct, water-equivalent hydrogen in the topsoil at each Mars target (Mars Odyssey GRS, 5° map, Boynton et al. 2007); water_made_minerals from published CRISM, OMEGA, TES and rover results.

ai_twins: for each Moon or Mars cell (body, lat, lon), rank 1 to 5 and the Earth place (earth_cell, earth_lat, earth_lon). Made by a self-supervised vision transformer (DINO ViT-S/8) trained on Kaggle (2 × T4) with 50 km height maps from all three worlds and a domain-adversarial head.

How good is it?

Ranks are "top X% of all land": 1% means only 1 in 100 land areas scores higher.

Test	Result
Hidden known sites (each site and its neighbours within 300 km hidden, model retrained)	the typical (median) site ranks in the top 2.4% as match_score scores it; the learned model alone gets top 1.2% (Moon 0.5%, Mars 2.4%)
28 published sites never used for training	the typical site ranks in the top 17.8% as match_score scores it (Moon 7.2% from only 7 sites, Mars 25.8%); learned model alone 15% (14.7%); a simple hand-made rule about top 40% on these, and 31 to 35% on the known sites
Mission factors alone (factors_only)	never-seen sites top 41% versus 40% for the simple rule, so only about as good as the rule there; known sites top 19% versus 35% for the rule
Scoring a site for the job it was really used for	ranks it higher for 78% of 78 site-and-world pairs (66 distinct sites; sign test p < 0.001)
Random weight changes of about 25% (each weight multiplied by exp(N(0, 0.25)))	9 of the top 10 places stay
Terrain AI, ground shape alone	known sites top 44% vs formula 62%; new sites 58% vs 71%; expert pairs 35% vs 34%
Best 10 km spot inside a place (website only; the test is in validation.json)	the 10 km piece holding the real site has a median rank of 42% where chance is 50% (66 field sites, p = 0.13): a hint, not a passed test

New data only changes the score if it ranks the known sites better. EMIT minerals did not: as a learned-model input they helped slightly overall (1.19% to 1.14%) but hurt where EMIT can see the ground (1.63% to 1.83%), and as a Mars mineral term in Rocks & craters they made the known Mars sites slightly worse (22.4% to 22.8%) and the independent ones better (41.3% to 38.2%). So they are measurements here, not part of match_score, and they cover only the 16,868 places where EMIT can see the ground. Details: validation.json and the How it works page.

Use it

import pandas as pd

base = "hf://datasets/ABRUBAB/shadowastra-atlas/"
places = pd.read_parquet(base + "places.parquet")
scores = pd.read_parquet(base + "target_scores.parquet", filters=[("target", "==", "jezero")])

best = scores.merge(places, on="cell").query("known_site_within_150km.isna()").nlargest(10, "match_score")
print(best[["region", "country", "lat", "lon", "match_score", "emit_clay_share", "emit_carbonate_share"]])

With 🤗 Datasets: load_dataset("ABRUBAB/shadowastra-atlas", "places").

Limits

• 55 km cells: many famous analog sites were chosen for something much smaller (an ice cave, a lake, a hot spring), so they can rank low here.
• Earth can never copy gravity, vacuum, radiation or lunar nights. The website shows these next to every score; this dataset scores only what Earth can copy.
• The Moon and Mars targets are 18 named places; ai_twins covers every 1° cell of both.
• Scores rank places for study. Visiting any site needs permission, local partners and care for the land; Antarctic sites need Treaty permits.

Sources and licence

This dataset is released under CC BY 4.0: please credit "Team ShadowAstra, ShadowAstra Atlas (NASA Space Apps Challenge 2026)" and the original sources: NASA POWER (LaRC); NASA GIBS and Terra MODIS; NASA EMIT L3 ASA v002 (JPL, LP DAAC, doi:10.5067/EMIT/EMITL3ASA.002); SRTM via AWS Terrain Tiles; LRO LOLA and MGS MOLA (USGS Astrogeology); LRO Diviner (polar product near the south pole, global day and night product for the three non-polar Moon targets), Lunar Prospector GRS and Mars Odyssey GRS (PDS Geosciences); NASA MODIS water mask (MOD44W, via GIBS); Smithsonian Global Volcanism Program; Earth Impact Database (PASSC); GLiM (Hartmann and Moosdorf 2012, CC BY); Natural Earth (public domain). NASA does not endorse this project.

Team ShadowAstra, Dhaka

Abdullah Rubab (lead) · Tanvir Hossin (planetary terrain) · Md. Tanvir Hasan (scientific validation) · Mohammad Saniul Alam (experience design) · Tasnuva Arif (development and testing) · Mohammed Mahmud Parvez (film and motion).

@misc{shadowastra2026,
  title  = {ShadowAstra Atlas: Earth analog sites for Moon and Mars missions},
  author = {Team ShadowAstra},
  year   = {2026},
  note   = {NASA Space Apps Challenge 2026, Dhaka},
  url    = {https://huggingface.co/datasets/ABRUBAB/shadowastra-atlas}
}