OpenCellID Bulk Downloads

The largest crowdsourced record of radio networks in the world, sliced into country-sized flat files. OpenCelliD - the community cell database maintained by Unwired Labs - takes observations contributed from handsets, modems and wardriving rigs, resolves them into logical cells, and issues one file per country, so a national extract always lines up with the borders you analyse.

Scale first, because it separates a corpus from a curiosity. The project's own country roll-up verifies the United Arab Emirates at 209,465 cells - 44,621 GSM, 111,996 UMTS, 52,836 LTE and twelve NR - beside Afghanistan at 16,186 and Albania at 19,237. Three markets, three different decades of network build-out, one identical row shape. Millions of cells land in the corpus overall.

Inside Datadory's catalog - 1,744 datasets averaging 7.81 - the telecom tower REITs slice holds 13 records averaging 7.31, and this one scores 8/10, carried by field definitions verified against the publisher's database-format documentation rather than guessed from column headers. Get a sample of this dataset cut to the countries you care about before anything else.

One row, one logical cell, one composite key:

# one row = one logical cell, keyed radio + mcc + net + area + cell
radio            : LTE           # GSM | UMTS | LTE | CDMA
mcc              : 260           # Poland
net              : 06            # network code; SID on CDMA
area             : 11823         # LAC / TAC / NID
cell             : 68211         # CID / LCID / BID
unit             : 289           # PCI on LTE, PSC on UMTS, empty otherwise

lon              : 21.0122       # decimal degrees
lat              : 52.2297
range            : 1200          # estimated coverage radius, metres
samples          : 487           # measurements behind the fix
changeable       : 1             # 1 = averaged from measurements, 0 = precise
created          : 1483228800    # unix ts, first observation
updated          : 1755648000    # unix ts, most recent observation
averageSignal    : -95           # dBm or per TS 27.007 8.5

# country roll-ups published beside the rows - same corpus, different mix
# United Arab Emirates: total_cells=209465  gsm=44621  umts=111996  lte=52836  nr=12
# Afghanistan:         total_cells=16186   gsm=11945  umts=4065   lte=176
# Albania:             total_cells=19237   gsm=5669   umts=9201   lte=4367

Read the anatomy rather than the illustrative values. changeable is the honesty column: 1 means the coordinates are averages of contributed measurements and may drift; 0 means the position is precise - typically fixed infrastructure measured directly. samples sizes the evidence behind a row, because a cell backed by thousands of observations deserves more trust than one backed by three. range converts that confidence into metres of uncertainty, and averageSignal keeps received-strength context beside the position instead of forcing a join elsewhere.

The published country roll-ups travel beside the rows and make technology mix legible at a glance. The UAE leans on UMTS (111,996 of 209,465) with LTE closing fast; Afghanistan remains overwhelmingly GSM (11,945 of 16,186) with just 176 LTE cells; Albania splits almost evenly across GSM, UMTS and LTE. Same fourteen columns everywhere - which is precisely why cross-country comparison becomes a group-by instead of a research project.

Fourteen, all verified against the publisher's database-format documentation. The identity stack answers which radio: radio picks the standard from GSM, UMTS, LTE or CDMA; mcc fixes the country; net carries the network code (doubling as the System ID on CDMA); area triples as Location Area Code, Tracking Area Code or Network ID depending on standard; cell finishes the key as Cell ID, UTRAN Cell ID or Base station ID. unit joins only where it exists - Primary Scrambling Code on UMTS, Physical Cell ID on LTE - and comes up empty on GSM and CDMA, a useful tell when validating parses.

The evidence stack answers how much to trust it: samples, changeable, averageSignal, and range as the metre-scale error bar. Provenance answers when: created and updated Unix timestamps per row. Deliveries extend the dictionary with derived cuts - country naming attached, wide pivots per radio standard, structure-cluster keys - folded under additional fields on request and pinned against live rows when your sample is cut.

Three chips, drawn straight from the record:

• Geography - global, partitioned one file per country keyed by mobile country code, with per-country cell totals published by radio standard on the project's stats page.
• Temporal - a rolling 18-month observation window applies to any single cut, and every row carries first-seen and last-seen timestamps so recency stays queryable per cell rather than asserted in aggregate.
• Granularity - one row per logical cell, keyed by radio, MCC, network code, area code, cell ID and unit, with decimal-degree coordinates and a metre-scale radius.

The window is the constraint worth designing around. Any one cut reaches back about eighteen months of observations, so trend work longer than that needs accumulation rather than archaeology - which is exactly why successive captures under a Datadory delivery stack into a lengthening panel instead of overwriting one another. Per-country extracts themselves scale from kilobytes to hundreds of megabytes uncompressed, so scope by market rather than pulling blind.

API, files, or your warehouse. Daily, weekly, or hourly.

The upstream corpus moves on its own rhythm; your delivery cadence gets set separately, to match the decision you are feeding. Pick files sized for overnight loads, an API for live lookups, or a direct pipe into Snowflake, BigQuery or Redshift - the rows arrive identical either way: one logical cell per row, fourteen documented columns.

Every delivery ships the field dictionary unchanged and sample rows for validation, so your first join happens against evidence rather than optimism. Derived cuts ride along on request - country naming attached, wide pivots per radio standard, structure-cluster keys for tower-level questions. Name the countries and cadence you need and the sample comes back shaped to them before any commitment.

Ranked by how directly one row settles the day job:

1. Infrastructure analysts running registered-versus-active ratios. Join observed cells against the FCC's antenna structure universe and the disagreement between the two - structures with no measurable radio, radios on no listed structure - is the finding.
2. Data scientists engineering radio-environment features. Coordinates, radii and sample-backed confidence feed models of signal environment, with per-row timestamps keeping staleness under control.
3. Competitive teams mapping radio technology by market. The radio/mcc/net fields turn technology mix per country into a countable split rather than a vendor's narrative.
4. Developers shipping geolocation products. Resolve a device-reported cell key to coordinates and radius offline - the classic embedding job on a schema that holds still.
5. Investors reading carrier capex proxies. Network-density growth by country shadows operator investment cycles without waiting for earnings calls.
6. Siting and policy teams screening gaps. Low sample counts and wide radii flag thin cells - candidate territory for new builds or municipal small-cell programmes.

The workflow companion lives at OpenCellID bulk download by country, and the wider strategy frame at the telecom tower REITs data guide.

Data scientists and ML engineers hold the top relevance score of 3 in this slice's persona pack - millions of observed cells with per-row timestamps are exactly what radio-environment models starve for. Developers and data-product builders also score 3, automating recurring country extracts against a stable documented schema.

At relevance 2: market researchers and consultants quantifying network buildout across countries, competitive intelligence teams mapping rival radio technologies market by market, and journalists, academics and students citing a community corpus with a stated observation window. Investors and quants round out the pack at relevance 1 with the capex-proxy read.

This slice splits cleanly by question. The FCC Data Hub Licensing and Tower Data owns the US regulatory side - antenna structures with owners, heights and determination histories - but lists steel, not signals, and stops at the border. The TowerMaps US Cell Tower Database consolidates 600,000-plus commercial site locations with owner-level attributes, issued quarterly and US-centric. The Data.gov Cell Tower Datasets Catalog harvests 46-plus municipal and state GIS layers for city-scale work. This record is the only global layer of observed radios in the slice - and the head-to-head against the municipal-GIS route is scored line by line in the comparison with the antenna catalog.

They complement rather than compete: registries say what was authorised, consolidations say who owns the structure, and this corpus says where radios are measurably operating, worldwide. Serious tower work runs at least two of the three. Browse the rest on the telecom tower reits data hub or the best telecom tower reits datasets ranking.

Four things, stated plainly.

First, cells are not towers. A logical cell is a radio sector identified by its key; several sectors typically share one structure. Comparing cell counts against a REIT's tower portfolio needs a stated mapping assumption - clustering by proximity and shared keys gets you a defensible estimate, and structure-cluster keys built for exactly that ship on request.

Second, any single cut observes a rolling 18-month window. History deeper than that comes from accumulation: successive captures under a Datadory delivery stack into a panel rather than replacing one another.

Third, the illustrative values above are shapes, not records. Definitions are verified; example coordinates, timestamps and signal readings get pinned against live rows when your sample is cut, which is also where formatting locks for your pipeline.

Fourth, file weight varies enormously by market - kilobytes to hundreds of megabytes uncompressed per country - so scope the sample to the countries that matter and nothing else ships. Get a sample of this dataset scoped to your markets before anything is scheduled.