ATRAC InstituteRadioGuard estimates radio-frequency exposure across Switzerland, in the browser, from public records: the federal register of licensed mobile installations, a sector-geometry snapshot, and official settlement geography. Every number the map shows is an upper-bound estimate produced by a documented propagation model — and every assumption in that model is stated here, including where it overstates reality and by roughly how much.
Public debate about mobile-network exposure runs ahead of the evidence, in both directions: alarmed claims built on numbers no one can trace, and reassurances built on measurements no one can see. RadioGuard's answer is neither. It is a transparent estimation instrument: the same records regulators publish, run through a model whose every parameter is written down, so that anyone can reproduce a figure and challenge any assumption in it.
1. The inputs, and where they come from
Three public datasets feed the model. Each is loaded as a local snapshot, so the map renders identically on every visit and never depends on a third-party endpoint being up.
The antenna register. The Federal Office of Communications (BAKOM) publishes the official inventory of mobile installations: 22,482 register rows, one per installation and technology group, covering 20,853 distinct coordinates. Each row carries the operator, the technology generations deployed, the permitted power class, the installation limit value that applies, and coordinates in the national LV95 projection. 21,261 rows carry a generation this map models and are drawn; the remaining 1,221 are 2G-only — including all 1,201 SBB railway installations — and are excluded, because the map does not model 2G.
The sector snapshot. The register records where an installation is, not where it points. RadioGuard bundles a derived direction snapshot: per site, the azimuths it transmits on, the generations each bearing carries, and the frequency bands paired with them. This is an estimate, not a measurement — every popup that shows an angle says so.
Settlement geography. Official swisstopo municipality boundaries (simplified to ~280 m) plus every OpenStreetMap place node in Switzerland with a settlement class: 15,174 places — 10 cities, 166 towns, 3,502 villages, 6,839 hamlets and 4,657 isolated dwellings. The modelled field is sampled at each place's centre, which is how the map can rank municipalities and draw the quiet patches around them honestly rather than anecdotally.
2. The propagation model
You can point the instrument at the geometry above: this is the live model, embedded, reading the same data the full map serves. Drag it, zoom it, change the threshold — it is the same code, not a screenshot.
The shaded field is computed in your browser, per screen pixel, from the permitted power class of each nearby installation — never from a pre-rendered image. For every site the model takes:
- the permitted power class from the register (what the installation is licensed to emit, not what a particular phone moment requires);
- the sector azimuths from the direction snapshot;
- an assumed 65° half-power beamwidth for the horizontal panel pattern;
- a free-space path-loss model with a vertical pattern and beam tilt applied on top;
- and adds the contributions of all nearby sectors at each point on the screen.
Two renderings come out of that model, and the settings panel chooses which is drawn. Per-mast fans show one threshold contour per antenna — an edge you can trace back to a single installation. Combined field (the default) sums every mast and paints the ground as the model sees it — the same quantity the popups, the pointer readout and the settlement figures report. The combined view is the honest one: what reaches a point is the sum of everything transmitting toward it, not the loudest neighbour.
3. What the model deliberately overstates
A model that only ever errs on the high side is more useful than a model that sometimes flatters. RadioGuard's assumptions push every estimate upward, and the map says so:
- Full licensed power, always. Installations are modelled at their permitted maximum around the clock. Real networks scale output down when demand is low and when a phone needs only a whisper; the register does not record the current power, so the model assumes the ceiling.
- Free space, no obstacles. Free-space path loss ignores walls, terrain and foliage. Indoors — where people actually spend most of their time — real levels are typically a factor of 10 or more below an outdoor free-space estimate at the same distance.
- Worst-case pointing. Sector angles come from a derived snapshot. Where a bearing is wrong or a panel has been re-aimed since, the model can place its lobe where the real one is not.
- No height discrimination. The vertical pattern is an assumption; a penthouse above the beam axis and a garden flat below it are treated more similarly than physics would treat them.
The practical consequence: where RadioGuard shows a low patch, real exposure can only be lower still — the estimate is a ceiling. Where it shows a high patch, the true value is likely somewhat lower, and only a measurement can say by how much. That asymmetry is intentional: the instrument is built to surface places worth checking, not to close the question.
Reading the figure above
Three things in the side-view diagram carry most of the model, and none of them is a measurement. The height sets where the beam can reach: a lower antenna drags the strongest ring toward the mast, a higher one pushes it out. The down-tilt decides where the beam's axis meets the ground — at these assumptions about 140 m out — and is the single assumption a real installation can violate furthest: remote electrical tilt hardware routinely spans 2–12°, and a site tilted at the bottom of that range puts its strongest ground reading several hundred metres out, while one tilted at the top puts it close to its foot. The vertical beam is what allows the model to say the ground under a mast is quiet: a person standing there sits well above the beam's axis, where a real panel radiates tens of decibels below its maximum. Change any of the three and the shape moves; the shape itself — a ring of strongest field at mid distance, quiet under the mast, falling away beyond — is what measurement confirms.
4. Modelled numbers versus measured ones
The map's numbers are estimates of a specific quantity: power density outdoors at ground level under the stated assumptions. They are comparable to each other (the model is applied uniformly), and comparable to published thresholds in power-density terms — the settings panel offers the precautionary targets and the Swiss installation limits as reference contours. They are not a substitute for an on-site measurement, which records what your specific balcony actually receives from all sources, indoors and out, at that moment.
The conversion rule used throughout: for a single plane wave in the far field, power density and field strength relate as S = E²/377, so µW/m² = E² × 2,652. Where the map reports power density and a device reports field strength, that is the bridge between them.
5. Provenance and reproducibility
Everything the map fetches is committed alongside the code that renders it: the register snapshot, the direction snapshot, the boundaries, the settlement list. Antenna data carries the federal register's licensing terms; the settlement layer and basemap are © OpenStreetMap contributors under ODbL 1.0, with the required attribution shown on the map itself. The build tools that assemble the snapshots from the upstream sources are part of the same repository, so the chain from public record to rendered pixel is inspectable end to end.
That is the standard the institute holds its instruments to: no number without a method, no method without its assumptions in the open, and no claim the instrument cannot support. Where RadioGuard is uncertain, it says so in the same breath as the estimate — which is exactly the point of publishing it.