geolocate-from-pixels

Geolocate from pixels

Every photograph taken outdoors contains enough information to place it. The constraint is never the image; it is your patience and your reference knowledge.

The beginner mistake is searching before inventorying. People see a mountain, type "mountain with two peaks" into a search box, and get nothing. The method is the opposite: extract every clue first, rank them by how much of the planet each one eliminates, and only then start searching — because the clue that pins the country is usually not the one your eye went to.

Rank your clues before you search

Work down this list. Each row eliminates far more of the world than the one below it, so a single row-one clue is worth twenty row-six clues.

Every variation and what it implies: reference/regional-indicators.md.

Method

1. Inventory. Write a numbered list of every clue in the frame before you search anything. Include the negatives — no snow, no palms, no overhead wires — because negatives eliminate regions just as well. Zoom in on every sign, every vehicle, every pole. Run the preprocessing recipes in find-the-original-image to read underexposed or small detail.
2. Fix the country. Combine your tier-1 and tier-2 clues until they agree. If two contradict — Cyrillic signage with right-hand-drive cars — that contradiction is a finding: an imported-vehicle market, a border region, or a composited image.
3. Read the text properly. Transcribe, then translate, then search the transcription verbatim in the local language. Searching a translation loses you the match. If the script is unfamiliar, get the script identified before you attempt letters — Georgian, Armenian, Amharic, Khmer, Thai, Lao and Sinhala are all frequently misidentified as each other's neighbours by people guessing.
4. Narrow to a locality. Named businesses go into a mapping search restricted to the country. Chains are useful in reverse: a chain that only operates in three provinces eliminates the rest of the country.
5. Query the map for the geometry, not the place. When you have no names but you do have structure — a water tower next to a rail crossing next to a football pitch — query OpenStreetMap features directly with Overpass rather than panning around. This is the step most people skip and the one that most often works.

[out:json][timeout:90];
area["ISO3166-1"="RO"]->.a;
nwr["man_made"="water_tower"](area.a)->.t;
foreach.t -> .w (
  nwr(around.w:400)["leisure"="pitch"]["sport"="soccer"];
  out center;
);

6. Confirm in imagery. Match the candidate against satellite/aerial and street-level sources. Compare invariants: building footprint shape, roof colour, the count and spacing of windows, kerb line, tree positions, the exact arrangement of a fence. Do not match on things that change — parked cars, awnings, signage, foliage density.
7. Chronolocate. Sun position for time of day and date band, season from vegetation, weather archives for corroboration. Procedure in reference/chronolocation.md.
8. Score it. Three independent features aligning, or stop.

Imagery sources and where each one wins

Where this goes wrong

• Confirmation bias is the failure mode of this discipline. You will find a building that looks right and then start explaining away the differences. Set your falsification criteria before you look: "if the pole on the left is on the wrong side of the road, this candidate is dead." Then honour them.
• Imagery is dated, and you are comparing across time. A missing building may have been demolished; a present one may be newer than the photo. Check the capture date of the imagery, and check the historical timeline before you reject a candidate.
• Ornamental and introduced vegetation lies constantly. Eucalyptus grows on five continents. Palms are planted far outside their native range. Vegetation is a tier-five clue for a reason — it corroborates, it does not decide.
• Global brands and franchised signage tell you almost nothing except where a company operates. A ubiquitous fast-food logo is not a clue; the local-language sub-brand and phone number on the same sign are.
• Compression invents detail. Text you "read" at the JPEG artifact level is frequently not there. If a plate or a sign only becomes legible after upscaling, it is a hypothesis, not a reading. Go back to the original pixels.
• Reflections and mirrors flip everything. Text in a shop window, or a scene shot into a mirror, reverses. So does a mirrored repost. If the driving side and the text direction disagree, suspect a flip before you suspect a country.
• Photos are not necessarily of one place. Composites exist, and a video can be cut from footage of several locations. Geolocating one frame does not geolocate the video. Verify frames independently, and hand suspicion to is-this-photo-real.
• Border regions and enclaves break single-clue logic. Signage, plates, currency and infrastructure all mix within a few kilometres of a border, and in territories with disputed or transitional administration.
• The claim shapes what you see. If you are told the photo is from a particular city, you will find that city. Try to do the inventory before you read the caption, and when you can't, run the exercise as though the caption said somewhere else.
• Long lenses compress and wide lenses stretch. Apparent distance between a foreground subject and a background mountain is a function of focal length. Do not judge "how close the hills are" without accounting for it.

Confidence grading

• Confirmed location — a specific coordinate where at least three mutually independent, non-transient features match reference imagery: for example building footprint geometry, the position and count of utility poles, and terrain profile. Independence is the requirement — three photos of the same sign is one feature, not three. You should be able to reproduce the camera position and bearing and state a radius in metres.
• Probable location — the correct locality with a plausible specific site; two independent features match, or three match but one reference source is undated or low-resolution. Express as a named place plus a radius, not a coordinate.
• Region only — country or province established from tier-one and tier-two clues with no site match. This is a perfectly respectable result and is often all a case needs. Say "somewhere in this province", not a point.
• Unconfirmed — a candidate that looks right but rests on transient features, a single matching element, or your own sense of resemblance.
• Excluded — you can affirmatively rule the claimed location out. Often easier and more valuable than finding the true one; a disproof needs only one hard contradiction, such as driving side.

Always report a radius with a coordinate. A bare six-decimal coordinate implies sub-metre certainty you do not have.

Worked example

An image is circulated as an attack on a fuel depot "in country A". No metadata.

Inventory: white centre line, right-hand traffic, concrete utility poles with a single horizontal crossarm and stubby brown insulators, a warning sign in a Latin script with a diacritic that does not exist in country A's language, one shop sign partly legible, low scrubby vegetation, bare deciduous trees, snow patches in shadowed ground only, a mountain ridge on the left horizon.

The diacritic already contradicts the claim — that is the finding that matters. Script and orthography narrow to two neighbouring countries. The pole and insulator style matches one of them.

The shop sign OCRs to a fragment. Searching the fragment as a business name gives a chain with outlets in one province. Overpass query for fuel depots within that province returns eleven candidates.

First candidate looks right in satellite view — same tank count. Killed by the falsification test: the access road approaches from the wrong side and the ridge would be behind the camera. Dead end, and a good one, because it was cheap.

Fourth candidate matches on tank arrangement, the perimeter fence corner, and the ridgeline profile generated from elevation data for that viewpoint. Street-level crowdsourced imagery from a nearby road shows the same pole line.

Chronolocation: shadow azimuth and a shadow-length ratio off the fence post give mid-morning and a solar elevation consistent with two date bands. Snow in shade only, plus bare deciduous trees, selects the late-winter band over the early-autumn one. A weather archive for the nearest station shows precipitation days earlier and clear skies that morning, consistent.

Result: location confirmed, 100 m radius, in country B not country A. Date probable to a two-week window. Time of day probable, mid-morning local.

Pivots

Legal and ethical notes

Reading public imagery and public map data is passive and lawful. Two limits are real. First, geolocating a private individual's home, school or routine from their own posted photographs is the core mechanic of stalking, and the fact that the technique is impressive does not make the output legitimate; do it for missing-persons work, authorized investigation, threat assessment, or to show someone their own exposure, and not otherwise. Publish a rounded location or a region rather than a doorstep coordinate. Second, in conflict work, publishing a precise location can endanger the people in the frame or make them a target. Both of these are judgement calls you must make explicitly and record. See ../../ETHICS.md.