NASA HiRISE Anomaly Investigation

Investigation methodology

This project applies a disciplined, falsifiable process to searching public NASA

imagery for anything that stands out. The bar for claiming an anomaly is high on

purpose: nearly all "anomalies" in planetary imagery are known sensor, compression

or viewing artifacts. A rigorous process is what makes a genuine finding credible.

Pipeline

1. Acquire — download only from official sources (see SOURCES.md), record URLs.
2. Ingest & catalog — build_catalog.py records path, mission, size, sha256, plus solar/geometry metadata parsed from the PDS label (pipeline/pds.py, pipeline/metadata.py).
3. Verify — verify_downloads.py confirms hashes, sizes, duplicates; build_catalog.py --check-immutable enforces that data/raw is byte-identical to the snapshot taken at ingest.
4. Preprocess — read native PDS/.IMG EDRs directly (pipeline/pds.py handles PDS3 labels, PDS4 XML, endianness, bit masks, BSQ/BIL/BIP), convert to 16-bit where possible, destripe/denoise, correct contrast.
5. Enhance — enhance.py: percentile stretch, denoise, unsharp, upscale, --native16.
6. Stack — stack.py: sigma-clipped median-stack repeated frames (--clip), column destriping (--destripe); residual maps highlight changes/transients.
7. Detect — detect.py: multi-scale local-contrast outlier flagging. --method box (classic) or --method annulus (each pixel vs. the annular ring around it — the "military-grade" background estimator that only flags things standing out against their immediate surroundings).
8. Photometry & geometry — pipeline/photometry.py normalizes for solar incidence (Lambert / Lommel-Seeliger), computes the sun-vector and shadow direction, and scores each candidate's shadow alignment with the sun. pipeline/metadata.py derives pixel scale, spacecraft altitude, solar elevation/azimuth from the product label so a shadow that disagrees with the sun marks a candidate down, and a physical object size can be estimated.
9. Triage — triage.py: HTML page of candidates with zoomed crops.
10. Adjudicate — for every candidate, work the artifact checklist (ARTIFACTS.md) before believing it; cross-band agreement, denoising persistence, compactness, terrain texture, solar-geometry shadow check, FDR correction (pipeline/adjudicate.py).
11. 3D check — scripts/check_stereo.py runs block-matching stereo disparity on a candidate within a stereo pair: real topography shows parallax against the surrounding ground, a 2D albedo/shadow patch does not.
12. Change detection — pipeline/changedet.py phase-correlates two frames of one site, registers them, and maps only the residuals that remain after alignment (same-lighting, same-season passes).
13. Confirm — find independent confirmation before recording a finding.

Controls (do these, they make the search meaningful)

• Negative controls: run detect.py on known-clean calibration frames and a random sample of "boring" terrain. Anything it flags there is a false-positive baseline; tune --z and --min-size until the baseline is near zero.
• Synthetic positives (injection): draw a few simulated bright/shadowed blobs of known size into images, run the detector, and measure recall (what fraction you catch) and precision. This tells you what your pipeline can actually see at a given pixel scale.
• Blind review: have a second person triage candidates without knowing which images they came from. Only candidates both reviewers flag proceed.

Confirmation standard

A candidate becomes a finding only when:

• It survives the full artifact checklist with documented reasoning, AND
• It is seen in at least one independent image (different pass, angle, or lighting), AND
• Its size and position are physically consistent across those images (checked via map projection / Trek overlays), AND
• It was flagged independently in blind review.

Evidence grading

| Class | Meaning |

|---|---|

| 4 | Unexplained after all checks; consistent across independent images; strongest claim level |

| 3 | Candidate; survives artifact checks but only seen in one image / one angle |

| 2 | Weak candidate; plausible artifact not fully excluded |

| 1 | Explained artifact or geometry |

| 0 | Hoax / retouched / meme-source / misattributed |

Rules

• Never claim without provenance (source URL + product ID + hash).
• Never "enhance" until the original is documented; always keep originals unmodified (data/raw is read-only after download).
• Never cite processed JPEGs from social media as primary evidence; go to the EDR.
• Record lighting geometry — the #1 cause of apparent "new objects" is different sun angle between passes.
• If a candidate disappears under a different lighting/pass, it is shadow or artifact, not object.
• A candidate whose shadow direction disagrees with the solar azimuth is self-invalidating.
• A candidate that shows no stereo parallax against its surroundings is 2D (albedo/shadow/artifact), not elevated — no matter how it looks.
• Thermal inertia differences (THEMIS IR) that do not show in visible light are geophysical, not structural, evidence.

Change detection caveats

changedet.py aligns two frames by phase correlation and reports residuals.

These are always expected and almost always mundane:

• Different sun angle (even 1°) shifts shadow length and moves shadow/shine boundaries.
• Different season → different frost/thermal state and dust redistribution.
• Different resolution/phase → resampling ringing, not ground change.

Only a residual that survives (a) tight registration, (b) same-lighting

acquisition (compare solar elevation and azimuth in the labels first), and

(c) stereo/flatness checks, is worth investigating.

Known artifacts — the checklist

Before recording anything, rule out these. Each is a well-documented cause of

false "anomalies" in planetary imagery. Note in the finding report which were

checked and why they don't apply.

Sensor / camera

• Cosmic ray hits: bright single pixels / short streaks, random orientation; common in orbital CCDs. Verify by checking neighbors and the same spot in a second frame.
• Hot pixels / bad columns: reproducible, fixed position across frames — identical across multiple images of the same scene = sensor, not object. (analyze.py flags narrow columns that stay bright/dark beyond the flagged box as column_smear.)
• Column/row defects and dead bands: fixed stripes.
• Calibration / zero-area strips: detector edges, trailing-edge dark areas often wrongly read as "craters" or "objects".
• Blooming / smear: charge spill along a column from a bright source. (Same automated column_smear check applies.)
• Dust or debris on optics: out-of-focus donuts or smears that repeat in the same position of every frame regardless of target.

Compression & processing

• JPEG/JP2 lossy artifacts: blocking, ringing, mosquito noise around edges. Use EDR (lossless) when available; never zoom into JPEG compression as "structure".
• Decompression errors: JP2 software differences produce grid or wavy artifacts.
• Stitching seams: panoramas/mosaics mismatch brightness across frame boundaries.
• Map projection artifacts: warping near poles/limb, oversampling, resampling ringing.
• Bit-flip / telemetry corruption: random noise bars, vertical striping, dropped rows.

Optics & lighting

• Lens flare / internal reflections: arcs, glows, secondary ghosts; follow the Sun direction.
• Vignetting: dark corners mistaken for clouds/shadow. (analyze.py flags candidates whose box sits near an image corner as corner.)
• Shadow geometry: long shadows from terrain; an object's own shadow. Check solar azimuth and incidence angles.
• Foreshortening at limb/oblique angles: features look elongated/compressed.

Hardware in frame

• Spacecraft booms, arms, wheels, mast, antenna, parachute, heat shield, backshell, sky crane, footpads, MOXIE, RTG etc. Rover/lander frames frequently contain their own hardware.
• Landing hardware in remote images: parachute + backshell + descent stage + heat shield are large and look anomalous to the untrained eye. NASA publishes their exact locations.

Human & narrative artifacts

• Copy/meme sources: a "UFO in an Apollo photo" is often just a rescanned, recompressed copy of an original — compare pixel-for-pixel to the archive scan.
• Highlights & blown-out regions: film/sensor saturation flattens to pure white; white blobs inside blown highlights are not objects.
• Film grain / scratches / dust on negatives in old Apollo scans.
• Misattribution: images of Earth, launch vehicles, ISS, or Earth-circling debris mislabeled as "from the Moon."

Geometry & photometry checks (automated, then manual)

• Shadow direction vs. solar azimuth: every real object's shadow points away from the sun at the label's solar azimuth. A "structure" whose shadow direction disagrees is shadow geometry, relief illusion, or processing, not an object. photometry.shadow_alignment scores this; adjudicate.py folds it into the verdict.
• Sun elevation vs. shadow length: an object's height can be estimated from its shadow length and solar elevation (metadata.height_from_shadow_len). An implied height wildly inconsistent with the feature's own size (or > terrain relief) argues against a discrete object.
• Stereo parallax: a real elevated feature shifts between the two frames of a stereo pair; a 2D albedo/shadow patch does not. No parallax → not elevated. (This is the single strongest debunker for "towers/pyramids".)
• Photometric consistency: normalized reflectance (Lambert/Lommel-Seeliger) should be roughly angle-independent across passes. A feature that only "appears" at grazing sun angles is relief illusion.
• Thermal inertia: THEMIS-IR-only brightness differences at a spot that looks ordinary in visible light indicate material/density differences (geophysical), not engineered structure.

The default conclusion

For every candidate, the null hypothesis is: known artifact, reproducible, explained by physics or instrument. Extraordinary claims require you to defeat that null hypothesis in writing, in the finding report, with references to the specific images and metadata.