Geo Clustering: Find Hotspots in a List of Coordinates

What does geo clustering show?

It shows where your points bunch together. Give the tool a list of coordinates, such as geolocated incidents, photo locations or sensor hits. It reports which points sit in dense groups (clusters) and which stand alone (noise). The method is DBSCAN, a density-based algorithm published by Ester, Kriegel, Sander and Xu in 1996 scikit-learn docs (read 16 Oct 2026). Two numbers drive it: a distance in metres and a minimum number of points.

The page is inspired by Bellingcat’s open-source geoclustering tool, a command-line program. It clusters the lat/lon rows of a CSV with DBSCAN or OPTICS and writes JSON, GeoJSON, text, CSV and a kepler.gl map Bellingcat on GitHub (read 16 Oct 2026). Bellingcat’s project is MIT-licensed. No code was copied: this is a separate JavaScript implementation that runs in your browser and needs no installation.

1. Paste a CSV or drop a file. Columns named lat/latitude and lon/lng/longitude are found automatically; a label or name column and a date column are optional.
2. Set the distance and minimum points. The defaults (200 m, 4 points) suit street-level data. See the next sections before you rely on a result.
3. Read the table and the map. Clusters are numbered by size, largest first. Each row gives the point count, the centre, the spread from the centre and the first and last date.
4. Export. Download the points with a cluster_id column as CSV or GeoJSON, or a one-row-per-cluster summary.

Settings at a glance

How does DBSCAN group the points?

DBSCAN looks at every point and counts its neighbours within the distance. A point with enough neighbours (the minimum points, counting itself) is a core point. Core points that are within reach of each other join one cluster. Points within reach of a core point, but without enough neighbours of their own, join as border points. Everything else is noise scikit-learn docs (read 16 Oct 2026).

Distances are measured along the Earth’s surface with the haversine formula, which gives the “as-the-crow-flies” great-circle distance Movable Type Scripts (read 16 Oct 2026). That means a 200 m setting means 200 m in Oslo and in Quito alike, which a setting in raw degrees would not. Points on either side of the 180° meridian are treated as neighbours when they are close. A grid index keeps the search fast, so the 20,000-point limit is practical (see the tests below).

Bellingcat’s tool also offers OPTICS, which it describes as giving tighter clusters but running slower Bellingcat on GitHub (read 16 Oct 2026). This page offers DBSCAN only.

How do I choose the distance and minimum points?

• Match the distance to your data’s precision. Phone GPS and geolocated photos are often good to tens of metres; geocoded addresses or place names can be off by hundreds. A distance below the error of your coordinates creates noise; one far above it merges separate places.
• Raise minimum points for noisy data. With 2 points, any pair becomes a cluster. With 5 or more, a cluster needs a real concentration.
• Test the stability. Run the same data at half and double the distance. A hotspot that persists is a finding; one that appears at a single setting is a parameter effect.
• Watch the spread column. DBSCAN chains along lines of points, so roads and rivers can produce long clusters. scikit-learn’s documentation calls the distance “not a maximum bound on the distances of points within a cluster” scikit-learn docs (read 16 Oct 2026).
• Remember duplicates count. Ten reports at exactly the same coordinate are ten points. Remove repeats first if they are the same event.

Why is there no street map?

Choice: no map tiles. The map is a plain SVG scatter plot drawn in your browser, with a coordinate grid and a scale bar. Map tiles would tell a tile server which area you are investigating, and heavy use of free tile servers is restricted. The OpenStreetMap Foundation’s policy allows normal interactive viewing of its standard tiles. It forbids bulk downloading, prefetching and offline use, offers “no SLA or guarantee” and can block heavy use without notice OSMF tile policy (read 16 Oct 2026).

For context, each cluster row has two links. OpenStreetMap opens the cluster centre on openstreetmap.org; Geo tools hands it to our geolocation tools. Both send only that one coordinate, and only when you click. For a full basemap, export the GeoJSON and open it in a desktop GIS such as QGIS, or in a mapping tool you already trust.

What can clustering not tell you?

• Why points are close. A cluster can reflect where people report, post or travel, such as a busy square or a road. It does not show that events are related.
• A single right answer. Different settings give different clusters. Report the distance and minimum points you used with every finding.
• Precision you do not have. Clusters inherit the errors of the coordinates. Rounded or geocoded locations produce artificial clusters on rounding grid points or town centres.
• Anything about time, unless you split the data yourself. The date column is summarised per cluster, but the clustering itself uses location only.

Use it responsibly. Clustering one person’s location history can reveal their home, workplace and routine. Use the tool on incident, photo or sensor data you are entitled to analyse, not to track a specific individual. Check the law on location data where you work.

How was the tool tested?

We tested the engine (geo-clustering-engine.js) in Node.js and the page in headless Chromium at 375 and 1280 pixels (16 Oct 2026):

• Haversine distance for 1° of longitude on the equator equals 6,371 km × π / 180 (111,194.9 m).
• The grid-indexed DBSCAN matched a brute-force DBSCAN on 15 random datasets. They covered a city, the high Arctic (78° N), the 180° meridian near Fiji, the North Pole and a global spread.
• A chain of points 100 m apart is one cluster at 110 m and all noise at 90 m. Minimum points 1 makes every point a cluster.
• Parsing covered header synonyms (latitude, lng), semicolon files with decimal commas, tab and space separators and GeoJSON. It also covered swapped, out-of-range and empty coordinates and CSV formula injection.
• 20,000 points at a 1 km distance clustered in about 0.1 seconds in Node.js.
• In the browser we tested the sample data, pasted text, file upload, column overrides, the tooltip, zoom and all three downloads. Hostile labels such as <img onerror> stayed text, and clustering made no network requests.

Sources

• Bellingcat, geoclustering README on GitHub, MIT licence (read 16 Oct 2026)
• scikit-learn 1.9.1 documentation: sklearn.cluster.DBSCAN (read 16 Oct 2026)
• Chris Veness, Movable Type Scripts: “Calculate distance, bearing and more between latitude/longitude points” (read 16 Oct 2026)
• OpenStreetMap Foundation: Tile Usage Policy (read 16 Oct 2026)