Road/Extraction

Blog/ 2026-11-09

Roundabout vs traffic circle: telling them apart in satellite imagery

A curving two-lane road winding through green countryside under a clear sky.

If you digitize roads for a living, you've hit this one: a closed ring of pavement with legs coming off it, no obvious signal heads in the imagery, and a tagging decision to make before the geometry goes into the graph. Is it a roundabout that gives way on entry, or a traffic circle that might run signals or stop control at the approaches? Get it wrong and the routing engine either stacks up unnecessary stops at a free-flowing yield junction, or sends traffic through what it thinks is a yield when drivers are actually facing a red light.

What separates the two on the ground

A roundabout is a specific geometry with a specific rule set: traffic entering yields to traffic already circulating, there's almost always a raised or painted splitter island at each approach, and the circulating lane doesn't stop for anything inside the ring. Most roundabouts built in the last twenty years are deliberately small in diameter to keep speeds down, with a central island that's landscaped or just curbed, not paved for turning movements.

A traffic circle (older engineering texts and a lot of US cities still call them rotaries) is the earlier design. Larger diameter, sometimes signalized at one or more entries, sometimes stop-controlled rather than yield-controlled, and historically built to move higher volumes at higher speed rather than to calm an intersection. Plenty of circles built before the 1990s still operate this way today, and plenty of newer construction has converted them to modern roundabout rules without changing the pavement footprint much at all.

Reading the geometry in imagery

At 0.5-1 m GSD you can usually pick out the splitter islands directly: small triangular or teardrop shapes sitting in each approach leg, separate from the central island. Their presence is the single strongest visual signal that you're looking at yield-controlled roundabout geometry rather than a circle. Crosswalk striping set back from the ring, rather than hugging it, is another tell, since modern roundabout design pulls pedestrian crossings away from the circulating lane.

Signal poles are harder. At 1-2 m you're often working from shadow and a faint dot at the corner of an approach rather than anything you'd call a pole in the image. If you can make out a mast arm shadow crossing the roadway at one or more legs, treat that as a circle with signal control until something else in the imagery or the legacy network contradicts it. Diameter alone is a weaker cue than it used to be. Plenty of agencies have shrunk old traffic circles down to roundabout footprints during resurfacing, so a large ring with no visible splitter islands could be either an unconverted circle or a converted roundabout where the paint hasn't been refreshed in the capture window.

Where this actually bites in the graph

Classifying junction geometry from imagery matters less for the shape of the line and more for the attributes hanging off it. A roundabout node typically carries give-way logic on every approach and no stop or signal delay cost assigned to the ring itself. A circle with signal control needs those delay costs preserved, or your routing engine will start recommending the circle as faster than an equivalent signalized four-way, which it usually isn't. Get the classification wrong across a metro area with a few dozen legacy rotaries and the error shows up as a consistent routing bias, not a one-off glitch, which makes it harder to catch in QA until a customer complaint points you at it.

In production work, splitter islands and crosswalk setback resolve the yield-vs-signal question on most junctions straight from the imagery. The ones they don't resolve get flagged for manual review instead of tagged on a guess. Pole shadows and stop bar paint fill in some of what's left. For anything still ambiguous, cross-referencing against whatever signal inventory or legacy OSM tagging already exists in that metro beats inferring control type from pixels that are two meters on a side.

New roundabout construction and circle conversions are exactly the kind of change that sits in a blind spot between survey cycles, which is the gap Road Extraction is built to close. It runs on a quarterly vector diff of the road network, extracted from high-res imagery, so a new or reshaped junction shows up as a mergeable segment in your layer instead of waiting on the next field pass.

Next capture window, that junction geometry change lands as a vector diff you can merge straight into your graph.

Start a project