Geolocating a photo or a point on the map

Geolocation means pinning down on a map where a photo or video was taken. It is central to verification work: confirming that an image of a flood shows this flood, that a viral video was filmed where people say it was, or that the photos in a listing show the property being advertised.

The method moves from the general to the specific and rests on a simple idea: every clue in the image is checked against an independent source until only one place fits.

Legitimate uses, and one that is not

• News verification and countering disinformation. Confirming or ruling out the location attributed to an image.
• Documenting public events: disasters, conflicts, demonstrations.
• Fraud. Listings with photos of somewhere else, or fake profiles that illustrate their supposed home town with pictures taken elsewhere.
• Organisational security. Photos posted by staff that reveal sensitive facilities or entrances.

What is not legitimate: working out where a private individual lives, works or spends their time. The techniques in this article are for verifying content and places, not for locating people.

Step 0: metadata and context

Before you study the image, check what came with it. If the file still holds GPS coordinates in its EXIF metadata, the job changes: you are no longer finding the place but checking that the coordinates are consistent with what you see. Images downloaded from social networks almost never have them.

Read the text that came with the post too, and look for the oldest copy with a reverse image search: the original context often says where the picture was taken.

Step 1: read the scene

List the clues and rank them from most to least distinctive:

• Text: signs, shopfronts, language and script, phone number formats, web domains.
• Infrastructure: which side of the road traffic drives on, road markings, sign styles, number plate format and colour, bollards, street lights, power lines, public transport.
• Architecture: materials, building heights, roofs, street furniture.
• Natural setting: vegetation, terrain, mountains on the skyline, coastline.
• Light and weather: shadow direction, cloud, snow, rain.
• Landmarks: towers, bridges, churches, stadiums, masts.

Use them to narrow down the country first, then the region, the city and finally the street. Write down each inference and the clue it rests on.

Watch out for misleading clues. A shop chain can operate in several countries, a language may be spoken on both sides of a border, and a sign may be temporary. If text appears back to front, the image has probably been mirrored, which flips directions and shadows too; correct for that before going any further.

Step 2: find candidates with OpenStreetMap and Overpass

OpenStreetMap is a collaborative map with a great deal of detail: every pharmacy, bus stop, church or bench can carry its own tags. Overpass is the API for querying that data by tags and spatial relationships, and Overpass Turbo is a website for writing those queries and seeing the results on a map.

Its strength lies in combining clues. If the photo shows a pharmacy next to a bus stop, you can ask for every pharmacy within 50 metres of a bus stop inside a given area:

[out:json][timeout:25];
node["highway"="bus_stop"](40.40,-3.72,40.43,-3.69)->.stops;
nwr["amenity"="pharmacy"](around.stops:50);
out center;

The four numbers define the search area (its south, west, north and east edges); replace them with your own. The result is a list of candidates, not an answer: each one has to be checked. And OpenStreetMap's level of detail varies from place to place, so a missing feature does not rule a location out.

Pandora runs a similar query around any point. Starting from coordinates typed into the search box (for example, 40.41689, -3.70358) or from a geocoded place, it lists the named places within 200 metres, with their type and distance: useful for checking whether what appears in the photo matches what surrounds a candidate. It is launched by hand from the location's menu, because the public Overpass server is shared and can take half a minute to respond.

Step 3: compare with street-level imagery

With one or more candidates in hand, it is time to compare the photo with the real place. Street-level imagery lets you do that without travelling:

• Mapillary collects street photos uploaded by users, with the date, the camera direction and an open licence.
• KartaView is a similar collaborative project.
• Commercial map services have their own street-level coverage.

Compare features that change slowly: building outlines, windows, kerbs, street lights, the skyline. Match the viewpoint using the camera direction, and keep each image's date in mind, because signs and shops change.

Pandora queries Mapillary around a point and lists the photos taken within about 120 metres, nearest first, with their date, the direction the camera was facing and who uploaded them.

Satellite imagery completes the comparison when the distinctive feature is best seen from above: the shape of a junction, a roof or a plot of land.

Step 4: time and shadows

Shadows point away from the sun, and the sun's position depends on place, date and time. Tools such as SunCalc calculate it for any point and moment.

A simple procedure:

1. Measure the direction of a shadow relative to something whose orientation you know from the map, such as the line of a street.
2. Work out at what times, on the claimed date, the sun sits in the opposite direction.
3. If you know an object's height, the length of its shadow gives the sun's elevation: the tangent of the elevation equals the height divided by the shadow length.
4. Check the recorded weather for that day: if the photo shows sunshine and it rained, something does not fit.

Two caveats. Shadows do not give a unique date: the sun follows almost the same path on two days of the year, symmetrical around the solstice. And the time in the metadata comes from the device clock, which may be wrong and often has no time zone. Use this check to confirm or rule out, not as proof on its own.

Step 5: confirm and document

Treat a location as confirmed when several independent clues match: not one, several. Record:

• The coordinates and the direction the camera faces.
• The matches, with screenshots from the photo and from the source.
• The sources and their dates.
• The alternatives you ruled out, and why.
• Your confidence level.

If the result is going to be published or to support a decision, ask someone else to repeat the geolocation without knowing your answer. If they reach the same spot on their own, your confidence can go up.

A note on IP addresses

IP geolocation is no use for this. At best it gives an approximate city and the network provider, and with a VPN not even that. It is useful in cybersecurity, not for placing a person.

Ethical limits

Geolocation is a powerful technique, which is exactly why it has clear limits:

• Do not geolocate private homes, or places that specific people frequent. If a legitimate investigation leads you to a private home, stop and consider whether your purpose justifies going any further.
• Do not publish locations that put anyone at risk: victims, witnesses, children.
• Keep the law in mind. Stalking and harassment are criminal offences in Spain, the UK and many other countries, and a location is personal data when it relates to an identifiable person.

Pandora looks up places, not people: Overpass returns shops, stops and named buildings, and Mapillary returns street photos. The purpose of every case is recorded in the audit log.

This article is not legal advice.