ResearchTechnology and Digital PracticeAugmented Reality (AR) Visualization

Augmented reality for landscape design: the evidence

What studies measured when designs were shown in AR on site, what Apple, Google and tool makers document about placing a model, and the limits.

Field Method · AI-assisted research roundupPublished Sources checked

Augmented reality (AR) draws a design model over the live camera view of a phone or tablet. On site, the proposal appears where it would stand. On a table, several people look at a small model together. The studies we read show that people accept AR and find it engaging. They show much less about whether people read a design more accurately with it, and they report again and again that placing the model in the right spot is the weak point.

This article covers handheld AR. Immersive headsets are covered in VR for landscape design review. We ran no tests of our own.

What the on-site studies measured

AR against photomontage, Bernburg, Germany. At a survey station on a pedestrian bridge, staffed for two days in June 2016, passers-by viewed a historic portico proposed for reconstruction. They looked through tablets that used printed markers of 45 by 70 cm, then compared the experience with photomontage posters. Eighty-two people filled in the questionnaire. Of these, 86.2% found AR totally or partially interesting and 89.8% found it partially or very useful. On the summed ratings AR and photomontage were close: 293 against 310 for ease of understanding and 271 against 269 for ease of use. The authors name marker tracking as the main technical weakness, because it is not robust where trees and crowds move, and they report that sunlight and reflections on the tablet screens degraded the experience. Goudarznia et al. 2017

Construction poles and AR on two neighbouring projects, Zurich, Switzerland. This field study used two projects. A residential complex of four buildings was staked out with the poles that mark a planned building’s outline on Swiss sites. A planned school building nearby was shown in a phone AR view. The same thirty participants, aged 21 to 58, judged the poles first and were then guided to the school site, where each saw one of three levels of model detail.

At the residential site, participants underestimated the height of the poles by about 34% on average, with a mean error of −9.6 m. At the school site, using AR, they overestimated the school’s height by about 31%, with an average error of 6.27 m. The average error for the school’s width was 6.57 m, and its length was underestimated by a mean of 26.9 m. As a baseline, participants judged the height of an existing residential building near the school site with an average error of 2.3 m.

Participants rated the AR view as easy to interpret, and the ratings did not differ significantly between the levels of detail. For height error, the results section reports a significant difference between the plain block model and the next level of detail (p = 0.02). The paper’s conclusion describes no significant difference in accuracy between the levels, so we read that result with caution.

The app took the viewer’s position from GPS with a stated accuracy of less than ten metres, and the experimenter calibrated the model’s position by hand for each participant. Foundation work for the school had begun, and some participants found the white barrier walls around the site hard to tell from the white model at the two simpler levels. The authors name small calibration errors, the phone’s small display and drift as it moved as possible sources of error. The authors write that using the AR view before construction started would have allowed a direct comparison between poles and AR. The two height figures come from different buildings and cannot be compared directly. Boos et al. 2023

Our reading of these two studies is that AR was liked. In Zurich, neither the poles nor the AR view produced accurate height judgements, each on a different building.

Design indicators, Honolulu, Hawaii. Passers-by used a phone app that placed an adjustable building complex on an empty corner lot, to learn indicators such as floor area ratio. Fifty valid questionnaires were collected. On a six-item test, the mean number of correct answers rose from 2.6 before the experience to 4.5 after. There was no comparison group using drawings or images. The authors report that occlusion did not work well on phones once the model was far from the device. Wang and Lin 2023

Spatial understanding, landscape architecture students. Nine graduate students viewed a landscape design of about 1,000 square metres outdoors on a tablet, in three groups of three with different freedom of movement, then drew plans and sections from memory. Drawings were more accurate in section when participants could walk through the scene. The group that could only orbit the design drew the most accurate plans and the least accurate sections. The authors report that tracking faults made the model shake and flicker on screen. They call for a larger study with a control group using plans and 2D images. Soria and Roth 2018

Tree planting with young people, Oslo, Norway. In 2020 and 2021, five groups of 9 to 18 young people from eight city districts used tablets to place virtual trees on real sites as part of a municipal tree plan. The study is qualitative. The author reports that participants placed objects at the intended spot and at correct scale within seconds, and that their schemes influenced the final plan. The same paper reports that location tracking and positioning were imprecise and often faulty, which irritated users. Tree positions could not be taken from the app into the city’s planning system and were redrawn by hand in CAD or GIS. Experts were still needed to judge which proposed trees were viable. Reaver 2023

Tabletop AR

Tabletop AR puts a terrain or site model on a desk for a workshop. Tomkins and Lange describe such a tool for landscape design and flood visualisation. It supports printed markers and markerless environmental tracking. The paper notes that a marker gives a group one shared, fixed location for the model, while environmental tracking lets each user place it anywhere. The paper presents the tool and a proposed series of three workshops. We found no user study in it, so it gives no measured effect. Tomkins and Lange 2019

How accurately a model can be placed

One open-access study measured placement error directly. It placed a 3D model of one building by several methods, indoors and outdoors. It measured reference points with a robotic total station, repeated each measurement four times, and reported root mean square error (RMSE). El Barhoumi et al. 2022

We give the results by type of method, as the paper groups them. This is one test at one site. It does not rank products.

  • Printed markers. With ARCore, RMSE was 9 cm indoors and 16 cm outdoors. With the Vuforia toolkit, it was 13 cm indoors and 20 cm outdoors.
  • Manual placement in two commercial apps. Trimble SiteVision gave 7 cm indoors and 7 cm outdoors. GAMMA AR gave 8 cm indoors and 8 cm outdoors.
  • Markerless, matched to an existing 3D model of the building. 9 cm indoors and 11 cm outdoors.
  • Markerless, with no prior model. 19 cm indoors and 40 cm outdoors.
  • Automatic placement by satellite positioning (GNSS) in SiteVision. 3.47 m, tested outdoors only.
  • The phone’s inertial sensors alone. 7.67 m, given in the paper’s text.

Two conditions belong with these figures. First, the authors attribute the 3.47 m result to low coverage of the correction service at their site and to the receiver used. Second, with the Vuforia marker method the error grew with distance from the marker, reaching 17 cm at 6 m and 20 cm at more than 8 m.

The study concludes that placement error ranges from a few centimetres to metres depending on the method.

What the platform makers document

Apple ARKit geotracking. ARKit’s geotracking configuration places location anchors at a latitude, longitude and optional altitude. Apple’s page states that it requires iOS 14 or later, a device with an A12 chip or later and cellular (GPS) capability, and an internet connection, and that it works outdoors only. The mechanism is a comparison of the live camera image with imagery Apple collected in advance from public streets. Apple states that the imagery covers specific areas in over 20 countries. Apple, ARGeoTrackingConfiguration Apple also states that geotracking does not support areas that are gated or open only to pedestrians. An app checks availability at the user’s location before it starts. Apple, Tracking geographic locations in AR An anchor should be within about 5 kilometres of the user. Apple, ARGeoAnchor The pages we read give no accuracy figure in metres.

In our reading, the street-imagery condition matters most for landscape work. The interior of a park or a trail is usually open only to pedestrians, which is the case Apple says is not supported.

Google ARCore Geospatial API. Google’s API matches the camera view to a model built from Street View images, which it calls the Visual Positioning System (VPS), and combines that with GPS. Google notes that a small number of ARCore devices do not support the API. Google, Geospatial API

For phones, Google’s Android guide says that under typical conditions VPS can be expected to give positional accuracy of approximately 5 metres and rotational accuracy of 5 degrees. It adds that outdoors, with few or no overhead obstructions, GPS may be enough. Google, Enable the Geospatial API

Google gives a different figure for a different class of device. Its guide to the same API in the Jetpack XR software kit lists wired XR glasses and audio and display glasses as the applicable devices. That guide says that when VPS can determine the user’s position it can often offer sub-metre positional accuracy and sub-degree orientation accuracy. Android Developers, Geospatial API for Jetpack XR Glasses are outside this article. We report the statement only to show that the maker’s figures differ by device class. Neither page states test conditions, and neither figure is a specification for a site.

Two further notes bear on design use. The yaw (compass direction) accuracy value is the radius, in degrees, of the 68th percentile confidence level, and Google says yaw can vary with VPS data availability and conditions at the location. Google, Geospatial pose By our arithmetic, a yaw error of 5 degrees moves an object 100 metres away sideways by about 8.7 metres. Google also warns that coordinates taken from Google Earth carry an error margin of up to several metres. Google, Geospatial anchors

High-accuracy GNSS: Trimble SiteVision. Trimble’s help lists five ways to place a model. Automatic and measured placement both require high-accuracy GNSS. Measured placement uses two known points on site, such as control points, and Trimble recommends a survey range pole for it. Trimble calls marker placement, with printed QR markers at known locations, the recommended method. Trimble, Placing a model · Trimble, measured placement The help pages we read give no accuracy figure. The test above is the only measured result we cite for this product.

Design software: SketchUp Viewer. SketchUp’s help describes AR in its Viewer for Mobile app. The user scans a flat surface such as a table or floor, and the model is placed on it and anchored with a tap. It appears at 1:1 unless it is larger than the detected surface, in which case it is rescaled to fit by default. Move, rotate and scale tools adjust it by hand. The page requires a device that supports ARKit or ARCore and a paid licence for the user’s own models. SketchUp Help We did not find placement by coordinates on that page.

Practical limits

Position. In the one placement test we cite, at one site, errors ran from 7 cm with manual placement in one app to several metres with satellite positioning or phone sensors alone. Printed markers gave 9 to 20 cm. The Zurich, Oslo and student studies also report drift, faulty positioning or jitter.

Light. Apple’s guidance is that world tracking depends on a clear camera image and degrades when the scene is too dark or lacks visible detail, and that fast movement blurs the image. Apple, Understanding world tracking Bright sun is a separate problem. The Bernburg authors reported reflections on the screen, and the placement study’s authors attribute less accurate marker results outdoors to exterior conditions and lighting.

Occlusion. By default the model is drawn over everything in the camera view, so a proposed tree appears in front of an existing building that should hide it. Apple documents occlusion by people on devices with an A12 chip or later. Apple, Occluding virtual content with people Google’s Depth API supports occlusion by real objects. Google states that it can estimate depth from 0 to 65 metres away, and that results are most accurate when the device is half a metre to about five metres from the scene. Google, Depth In our reading, the most accurate band covers a table or a small courtyard and is short for most site views. A flood visualisation app instead hid virtual content behind simple blocks that the user drew on site to stand for existing buildings. Its authors note that this trades model accuracy for versatility. Haynes and Lange 2016

Devices. Google keeps a list of certified ARCore devices and states that, as of November 2026, over 85% of active devices support the Depth API. Google, ARCore supported devices A public event cannot assume every visitor’s phone will work. The Honolulu and Oslo studies gave participants prepared devices.

A short protocol, as our own suggestion

These steps are ours. They follow from the sources above and have not been tested.

  1. State the purpose. The studies report that participants accepted AR and found it interesting. None shows that it engages people more than photomontage or a physical model. We found no study showing that people judge sizes accurately in AR. In the one study that tested it, on one building, the average height error was about a third.
  2. Decide how close the overlay has to be, and say so. The one placement test we cite found 7 cm to several metres at a single site. We give no tolerance. Where the position of the overlay will be relied on, the control points, the placement and any statement of positional accuracy are matters for a licensed land surveyor.
  3. Test at the site before promising geolocated AR. Ask the app’s maker which positioning method it uses, then run the app at the site on the devices the meeting will use. Apple’s and Google’s documentation ties geolocation to street imagery, so expect park interiors and trails to be the weak case.
  4. Put a known object in the model. An existing wall, pole or building corner in the model lets viewers see whether the overlay is aligned at that object. It does not show alignment farther away.
  5. Bring a second medium. In Bernburg, photomontage scored about the same as AR. Printed views cover glare, unsupported phones and tracking failure.
  6. Record what was shown. Keep the device, app and version, the placement method, the check against the known object, and screenshots with time and position.
  7. Treat positions marked in AR as sketches. In Oslo they had to be redrawn in CAD or GIS and reviewed by experts.

What we could not settle

  • Does AR improve design decisions or project outcomes? No study we found followed a landscape project to construction.
  • Do people judge size better in AR than with poles, drawings or photomontage? The Zurich study tested poles and AR on different buildings, and we found no like-for-like test.
  • How does tabletop AR compare with a physical model or a shared screen? We found a tool description and no peer-reviewed test for a landscape design task.
  • What accuracy do ARKit geotracking and the ARCore Geospatial API reach in parks and on trails? We found no independent open-access measurement.
  • Does the sub-metre statement Google makes for XR glasses apply to phones? The phone guide we read does not say so.
  • What do other design-software makers document for AR? We read SketchUp’s page and did not reach Esri’s or Vectorworks’ current documentation.
  • Does occlusion by terrain and vegetation work at site distances on current devices?

Field Method tools

Field Method has no AR tool. Photo tools places site photographs on a map and against a site’s 3D model, with each location reviewed by the user. It works with photographs of existing conditions and does not composite a design into a photograph. Its key view map, which numbers each photograph where it was taken, can go beside the printed views in step 5. Positions come from the phone’s GPS record or the user’s own entry. They are not survey positions.

Sources

  1. Goudarznia, Pietsch and Krug (2017) — Testing the Effectiveness of Augmented Reality in the Public Participation Process: A Case Study in the City of Bernburg, Journal of Digital Landscape Architecture 2-2017
  2. Boos, Reichenbacher, Kiefer and Sailer (2023) — An augmented reality study for public participation in urban planning, Journal of Location Based Services 17(1)
  3. Wang and Lin (2023) — Public participation in urban design with augmented reality technology based on indicator evaluation, Frontiers in Virtual Reality
  4. Soria and Roth (2018) — Unreal Reality: An Empirical Investigation of Augmented Reality Effects on Spatial Cognition in Landscape Architecture, Journal of Digital Landscape Architecture 3-2018
  5. Reaver (2023) — Augmented reality as a participation tool for youth in urban planning processes: Case study in Oslo, Norway, Frontiers in Virtual Reality
  6. Tomkins and Lange (2019) — Interactive Landscape Design and Flood Visualisation in Augmented Reality, Multimodal Technologies and Interaction 3(2), 43
  7. El Barhoumi, Hajji, Bouali, Ben Brahim and Kharroubi (2022) — Assessment of 3D Models Placement Methods in Augmented Reality, Applied Sciences 12(20), 10620
  8. Apple Developer Documentation — ARGeoTrackingConfiguration
  9. Apple Developer Documentation — Tracking geographic locations in AR
  10. Apple Developer Documentation — ARGeoAnchor
  11. Google for Developers — Build global-scale, immersive, location-based AR experiences with the ARCore Geospatial API
  12. Google for Developers — Enable the Geospatial API for your Android app (Kotlin/Java)
  13. Android Developers — Work with the Geospatial API using ARCore for Jetpack XR
  14. Google for Developers — Obtain the device camera's Geospatial pose (Android, Kotlin/Java)
  15. Google for Developers — Geospatial anchors (Android, Kotlin/Java)
  16. Trimble — SiteVision Help: Placing a model
  17. Trimble — SiteVision Help: Placing a model: measured placement
  18. SketchUp Help — Viewing Your Models in Augmented Reality
  19. Apple Developer Documentation — Understanding World Tracking
  20. Apple Developer Documentation — Occluding virtual content with people
  21. Google for Developers — Depth adds realism (ARCore Depth API)
  22. Haynes and Lange (2016) — Mobile Augmented Reality for Flood Visualisation in Urban Riverside Landscapes, Journal of Digital Landscape Architecture 1-2016
  23. Google for Developers — ARCore supported devices

These tools support site review and design studies. Read each tool’s stated capabilities and limits before use.

Explore Photo tools
Back to Research