OrbitSenseExplorerMethod and resultsSubmission

OrbitSense terrain comparison method

OrbitSense compares measured regional terrain around two planetary references with a small catalogue of Earth locations. The app exposes the measurements, their provenance, the comparison scale, and the contribution of each descriptor. It supports an initial search for experimental field settings. Each geological, environmental, or operational interpretation carries its own evidence.

Reference areas and observations

The exact reference centres, body-specific coordinates, analysis footprints, elevation products, nominal posting sizes, and retrieval records are stored in data/features.json and data/sources.json. Downloaded crops preserve the actual observations used. Candidate windows follow the same declared physical size and analysis grid as their planetary reference.

The nominal analysis uses a 25.6 km square and 800 m cells, giving a 32 by 32 grid. Each square is centred at the coordinates below in a local azimuthal equidistant projection with metre units. This metric definition is the authoritative footprint, especially near the lunar pole where a geographic bounding box crosses the longitude seam.

ReferenceCentreFootprint interpretationObservation
Shackleton region89.67 degrees south, 129.78 degrees eastMixed crater, rim, and exterior terrainAdjusted LRO LOLA south-polar DEM with 80 m postings
Jezero western floor18.4447 degrees north, 77.4508 degrees eastWestern floor and adjacent regional reliefMGS MOLA global DEM with nominal 463.09 m postings

All thirteen Earth windows use NASADEM HGT v001. They lie within its 60 degrees north to 56 degrees south coverage. NASADEM is a reprocessing of the February 2000 SRTM observations. In particular, the Kilauea terrain snapshot predates later volcanic changes.

The adjusted polar LOLA product replaces the initially considered global 118 m grid. It supplies finer nominal sampling in the selected polar setting. The USGS public S3 copy of the MOLA product replaces an inaccessible planetarymaps download route. Both substitutions and exact asset URLs appear in the source records. The app uses no CTX, Landsat, Copernicus DEM, or WorldCover observations in this release.

Each location uses a local metric grid. Earth processing uses Earth coordinates. Lunar and Martian processing uses the corresponding body radius and coordinate convention. Relative elevations are measured within each window, so different absolute vertical datums do not enter the similarity calculation. Elevation is aggregated to the common analysis cell before terrain derivatives are calculated. Interpolation does not establish finer observational support.

The source metadata distinguishes nominal pixel spacing from effective support, missing cells, and filled or interpolated areas. Missing observations remain missing. A location with an incomplete core feature vector receives no score.

Descriptors

DescriptorUnitsDefinition
SlopeDegreesMedian slope derived from elevation gradients in metres on the common metric grid
RoughnessMetresRoot mean square elevation residual after subtracting a fitted planar surface
Relative reliefMetresDifference between the 95th and 5th elevation percentiles

These descriptors describe terrain at the declared grid and window scales. Roughness depends strongly on scale. Detrending reduces the effect of a uniform regional tilt on roughness, while slope retains that tilt as a separate measurement. Correlated descriptors can still influence a ranking together. The weight controls and relief-only baseline make that dependence visible.

Similarity

For feature j, candidate observation x, reference observation t, and positive tolerance tau, descriptor agreement is exp(-abs(x - t) / tau). The app normalizes the three positive feature weights to sum to one and reports 100 * sum(weight * agreement).

Every tolerance is recorded beside the reference measurements with its derivation. A tolerance states the scale of difference used by this demonstration. It requires scientific validation for a particular experimental purpose. The score is dimensionless agreement under this model. It is not a probability, an Earth-wide percentile, or a certification of suitability.

For each descriptor, the recorded tolerance is the maximum of a declared floor and twice the robust spread over the nominal reference and seven reference perturbations. Robust spread is 1.4826 times the median absolute deviation. The floors are 1 degree for slope, 100 m for roughness, and 250 m for relief. These are design scales rather than calibrated measurement errors.

ReferenceSlope tolerance in degreesRoughness tolerance in mRelief tolerance in m
Shackleton2.77153081100250
Jezero1100250

All three core descriptors must be finite, and all tolerances must be positive. The all-zero weight setting cannot define a comparison and prompts the user to restore a positive weight. Changing weights changes the contribution of existing observations. It does not generate new measurements.

Evaluation

processing/evaluation.py reproduces the browser score in Python. It records the nominal ranking, a relative-relief-only baseline, and sensitivity to feature weights, window sizes, and analysis resolutions represented in the saved observations.

The weight experiment uses 200 deterministic Dirichlet(1,1,1) draws. It reports top-three acceptance, rank quantiles, top-three overlap, and Spearman correlation. These shares describe chosen perturbations. They are not statistical confidence intervals. Resolution and window diagnostics retain the nominal reference tolerances to isolate changes in observed features.

The seven measured perturbations use 19.2 km and 32.0 km windows at 800 m, a 25.6 km window at 1600 m, and nominal-size windows shifted 3.2 km east, west, north, and south. Reference and candidate windows receive the same perturbation. Every recorded perturbation passed the usable-elevation threshold. There is no product-swap or finer-grid sensitivity experiment in this release.

Comparison sites are labelled explicitly. Their measured results are retained even when they score highly. The diagnostic fraction of candidate-comparison pairs ordered as expected describes this selected catalogue. The catalogue does not supply independently verified binary truth labels for these exact planetary references, so the diagnostic does not establish general classifier accuracy.

Scientific interpretation

Published analog relationships are property-specific. NASA documents Death Valley terrain for Mars exploration activities, volcanic landscapes in Hawaiʻi for lunar field research, Atacama soils for biosignature investigations, and Lake Salda sediments in relation to Jezero. These documented relationships do not guarantee that an arbitrary local window reproduces the regional topography of the selected Shackleton or Jezero reference.

The source for these contextual examples is NASA's Planetary Analog Explorer. NASA also describes the Haughton crater setting and lunar geology training in Iceland. Only sources corresponding to locations actually included should appear as candidate-specific evidence.

Terrain measurements cannot establish mineral composition, subsurface ice, habitat suitability, radiation shielding, geotechnical conditions, or operational safety. Gravity, atmosphere, illumination, weathering, and biological activity differ across bodies. Evidence cards identify the experimental activity that the measured geometry could help study and the observations still required for a stronger recommendation.

Scope relative to the revised design

This working release implements the current request for a transparent, reproducible terrain comparison with measured sensitivity variants. The larger plan_2 design additionally proposes a systematic Earth catalogue, named-unit child-window retrieval, reference-derived background percentiles, cross-scale admission, and an independently labelled benchmark. Those research components require separate implementation and validation before their claims can enter the results.

OrbitSense measured evaluation

Small, handpicked demonstration catalogue. Measurements are real. Rankings describe declared terrain descriptors at the analysed scale. No discovery, safety, habitability, or broad predictive-accuracy claim is made.

Results below are generated by processing/evaluation.py from saved observations. The nominal model uses equal weights for slope, detrended roughness, and relative relief.

Shackleton region

RankSiteRoleScore
1Himalaya Relief Controlcomparison15.87
2Mauna Keacandidate1.73
3Death Valleycandidate0.26
4Wadi Rumcandidate0.18
5Atacama Yungaycandidate0.15
6Tibesti Volcanic Terraincandidate0.14
7Kilaueacandidate0.14
8Cima Volcanic Fieldcandidate0.13
9El Pinacatecandidate0.11
10Amazon Surface Controlcomparison0.06
11Aden Crater Regioncandidate0.06
12Kansas Plains Controlcomparison0.05
13Qaidam Basincandidate0.05

The relief-only baseline and the full ranking share 3 of the top 3 sites. Their Spearman rank correlation is 0.819.

Across 200 deterministic weight draws, the mean top-three overlap is 1.000 and median Spearman correlation is 1.000. These are sensitivity diagnostics.

VariantWindow kmCell mTop-three overlapSpearman rho
window-small19.28002/30.890
window-large32.08003/30.978
resolution-coarse25.616003/30.879
shift-east25.68003/30.978
shift-west25.68003/30.995
shift-north25.68003/30.978
shift-south25.68003/30.995

Descriptive separation of a handpicked candidate group from labelled comparison sites. The groups are not independently validated binary terrain-analog labels, so this fraction is not a general accuracy estimate.

Candidate-comparison pairs evaluated: 30. Fraction with the candidate above the comparison, counting ties as one half: 0.567.

Documented analog context checks

Published analog relationships refer to their documented dimension and planetary body. These exact regional reference windows have a separate measured terrain result.

SiteDocumented dimensionTerrain rankTerrain score
Kilaueageology70.14

Kilauea scores 0.14, compared with Kansas Plains Control at 0.05. The measured score difference is 0.08 points.

Kilauea scores 0.14, compared with Himalaya Relief Control at 15.87. The measured score difference is -15.73 points.

Kilauea scores 0.14, compared with Amazon Surface Control at 0.06. The measured score difference is 0.08 points.

Jezero western floor

RankSiteRoleScore
1Cima Volcanic Fieldcandidate63.78
2El Pinacatecandidate52.59
3Atacama Yungaycandidate46.59
4Wadi Rumcandidate42.17
5Tibesti Volcanic Terraincandidate37.70
6Kilaueacandidate37.25
7Aden Crater Regioncandidate26.33
8Amazon Surface Controlcomparison25.86
9Qaidam Basincandidate22.10
10Kansas Plains Controlcomparison21.60
11Death Valleycandidate7.56
12Himalaya Relief Controlcomparison0.54
13Mauna Keacandidate0.38

The relief-only baseline and the full ranking share 3 of the top 3 sites. Their Spearman rank correlation is 0.929.

Across 200 deterministic weight draws, the mean top-three overlap is 0.900 and median Spearman correlation is 0.984. These are sensitivity diagnostics.

VariantWindow kmCell mTop-three overlapSpearman rho
window-small19.28002/30.813
window-large32.08003/30.973
resolution-coarse25.616002/30.973
shift-east25.68003/30.989
shift-west25.68003/30.995
shift-north25.68003/31.000
shift-south25.68002/30.962

Descriptive separation of a handpicked candidate group from labelled comparison sites. The groups are not independently validated binary terrain-analog labels, so this fraction is not a general accuracy estimate.

Candidate-comparison pairs evaluated: 30. Fraction with the candidate above the comparison, counting ties as one half: 0.800.

Documented analog context checks

Published analog relationships refer to their documented dimension and planetary body. These exact regional reference windows have a separate measured terrain result.

SiteDocumented dimensionTerrain rankTerrain score
Death Valleyterrain117.56
Atacama Yungayprocess346.59

Death Valley scores 7.56, compared with Kansas Plains Control at 21.60. The measured score difference is -14.03 points.

Death Valley scores 7.56, compared with Himalaya Relief Control at 0.54. The measured score difference is 7.02 points.

Death Valley scores 7.56, compared with Amazon Surface Control at 25.86. The measured score difference is -18.29 points.

Atacama Yungay scores 46.59, compared with Kansas Plains Control at 21.60. The measured score difference is 25.00 points.

Atacama Yungay scores 46.59, compared with Himalaya Relief Control at 0.54. The measured score difference is 46.05 points.

Atacama Yungay scores 46.59, compared with Amazon Surface Control at 25.86. The measured score difference is 20.73 points.

Limits of this check

The catalogue was selected for a small demonstration. Existing analog documentation often concerns a different property or scale from these exact planetary reference windows. A comparison site can rank highly on terrain despite being unsuitable for another purpose. Such a result is retained. No formal classifier performance, statistical significance, or proof of novelty follows from this catalogue.

Nominal tolerances are kept fixed in the resolution and window-size diagnostic. This isolates the change in measured features. It does not implement the complete multi-scale site-to-landscape admission design proposed in plan_2.