LGeo and ReCover Initiative reveal an automatic rooftop HVAC unit detection tool
- Jamie Blackley
- 4 days ago
- 3 min read
Updated: 3 days ago
ReCover Initiative identified a practical data gap that started this work:
There is no simple, city-scale inventory of rooftop HVAC units on large buildings in Canada.
This matters because rooftop HVAC units (also referred to as RTUs) are part of the building energy and emissions picture, but they are not easy to count. A full inventory of RTUs traditionally means manual review, site visits, or relying on incomplete records. Without a city-wide rooftop HVAC inventory, it is harder to estimate retrofit opportunities, plan outreach, and decide where more detailed energy and emissions analysis is needed.
Our objective was to support ReCover in filling this gap. We tested whether high-resolution aerial imagery and machine learning can produce a usable rooftop HVAC inventory. Moncton was our test city and we found that indeed you can! This blogpost shows you how you can view what that model looks like and provides some examples of how you could use this information to support your work. Some additional details on the model and approach are also included.
The dashboard
The dashboard demonstrates some features that allow you to explore the RTU context in Moncton. It shows four layers:
RTU detections: the blue outlines describe what is inferred as an RTU and what isn't. This layer can be filtered by model confidence thresholds.
The layer answers the main question: where are rooftop HVAC units in Moncton?
Model validation: Three colours describe how well the model performs.
🟢 green describes a correctly detected RTU
🔴 red describes a rooftop feature that isn't an RTU but the model incorrectly thinks it is
🟡 amber describes an RTU that the model missed.
Building footprints, from NB open data, which show popups when clicked giving a building-level RTU summary
Masked ortho, also from NB open data, which shows the high-resolution imagery that the model used to detect RTUs.
What you can do with this information
This inventory is a starting point for planning.
High-confidence detections can support early screening and city-scale estimates of buildings that may require HVAC retrofit.
To dataset helps to understand how HVAC equipment is distributed in your city’s buildings
Counting the amount of equipment that needs to be replaced helps with planning! For example, bulk procurement programs can be designed, and contractors.installers/suppliers can gather valuable market information.
The next step is to connect the rooftop unit inventory with other building information so it can support retrofit planning, emissions analysis, and targeted follow-up.
Conclusion
To achieve our climate goals, building retrofits will be fundamentally important. We believe that more readily available data on the building stock can help to develop optimal retrofit plans. We hope to continue to partner with ReCover, retrofit accelorators, and those working in this space in the future to modernize the retrofit landscape!
Appendix for nerds
New Brunswick benefits from high quality open data. For this project, we used GeoNB’s 2024 orthophotos at 7.5 cm resolution. That is to say, very high resolution imagery in which features on individual rooftops are visible from above. The same workflow can also be repeated in other communities with comparable imagery.
Our workflow has three steps:
preprocess hundreds of gigabytes of high resolution imagery;
run a Faster R-CNN object-detection model on that imagery; and
validate the detections against hand-labelled data.
How the model performs
For the final output, we kept detections with a confidence score of 85% or higher. The model detected 2,955 rooftop units across the buildings analyzed. To check the results, we hand-labelled a validation set of 508 buildings containing 2,415 rooftop units. A detection counted as correct only if its box overlapped a labelled unit by at least half, using PASCAL VOC matching at IoU ≥ 0.5 (close guesses did not count).
The validation results were:
2,160 correct detections
248 false positives
255 missed units
89.4% recall
89.7% precision
10.3% false-positive rate
In plain terms, the model found about nine out of ten real rooftop units, and about nine out of ten detections were correct. These results are not perfect. They are also not a black box. The errors were kept visible so the output can be checked and used with the right level of confidence. As we iterate on other cities, we expect both the confidence and adaptability of the model to improve over time.
Comments