A roof that looks straightforward from the street can be difficult to interpret from above. Dormers interrupt a slope, shadows obscure edges and small rooftop objects can be mistaken for usable surface. RoofeNet-Multi preserves how people revise AI proposals for those features, creating a dataset for research on rooftop photovoltaic potential and annotation uncertainty.[1]
The Scientific Data paper describes 5,359 Czech aerial-image tiles with 12,761 annotation sets: 5,359 initial model proposals and 7,402 human reviews. Of the images, 1,502 received at least two human reviews. The remaining 3,857 received one. Counting annotation sets as separate images would substantially overstate the dataset’s size.[1]
A record of review, rather than independent agreement
The model supplied starting labels for roof areas and obstacles, and reviewers accepted or changed them. Every human annotation therefore begins with an AI suggestion. Multiple reviews reveal variation in that process, but they do not show what independent annotators would have drawn without seeing a proposal. A plausible initial box can influence where a reviewer looks and what they leave untouched.[1]
That makes the provenance of each annotation valuable. Researchers can compare the model’s starting point with human changes, identify features that repeatedly need correction and study disagreement where several reviews exist. Collapsing those records into a single supposedly definitive label would discard part of what distinguishes the release.[1]
The sample and the geometry define its uses
The imagery comes from areas around the 15 most populous Czech cities after excluding Prague. Tiles are centred on buildings, rather than drawn as a random sample of every kind of landscape. That selection matters for a model later asked to search large areas containing empty land, unusual buildings or architecture unlike the sampled cities.[1]
The labels are axis-aligned rectangles for roof planes and obstacles. They do not supply georeferenced polygons or the full geometry needed to calculate usable roof area, pitch, shading and an installation design. A box can help a model locate a candidate feature without becoming a trustworthy physical measurement.[1]
RoofeNet’s immediate contribution is therefore to the step between aerial imagery and a dependable roof interpretation. It lets researchers inspect how AI suggestions survive human review, where they change and where people still disagree. Those distinctions are useful before any prediction is turned into a solar estimate.[1]
Sources & further reading
- Human-reviewed Rooftop Photovoltaic Annotations Initialized by a Vision-language Model — public manuscript Scientific Data / author manuscript on ResearchGate · 24 Oct 2026 · Accessed 26 Oct 2026