AI tools for solar site assessment: what they see, and what they can’t

What AI can reliably assess from imagery

Given sharp, recent imagery, a remote tool does a respectable job on the parts of a site that are visible from the sky:

• Roof outline and plane layout. Ridges, hips, valleys and eaves are high-contrast edges, which is exactly what vision models are good at.
• Azimuth. The direction a plane faces falls out of the outline and is usually solid.
• Pitch, as an estimate. Pitch has to be inferred from stereo imagery, oblique views or LiDAR height points. How good it is depends on resolution and point density, so treat it as an estimate unless the tool states a tolerance.
• Large obstructions and trees. Chimneys, dormers, skylights and big canopies show up. Small vents and stacks often do not.
• A first shade estimate. Tree and building heights from LiDAR or 3D imagery feed a sun-path model that is good enough to qualify or drop a lead.

NREL has reviewed these remote methods in its report on estimating rooftop suitability for PV, and the conclusion holds up: remote data is a strong screening tool, and its accuracy depends on the input data. Two limits are easy to forget. Imagery has a capture date, so a tree that grew or a vent added since is not in it. And a model can be confidently wrong on unusual roofs: tile, low-slope sections, additions at odd pitches.

What still needs someone on site

Our site survey page puts it plainly: satellite imagery doesn’t know what’s in your attic. The data that decides whether a design can be permitted and built sits under the roof deck and behind the panel cover:

• Rafter or truss size, spacing and span. The structural engineer needs members, not an outline. No image from above records them.
• Sheathing, roofing layers and condition. Delamination, soft decking and a second layer of shingles show up at a cut edge or from the attic, not from orbit.
• Main breaker and busbar ratings. For a load-side connection, the NEC 120% rule is calculated against the busbar rating, and the two labels often differ. Only an open panel answers it. The code text lives with NFPA 70.
• Meter, service type and equipment locations. The utility application and the one-line both need them, and the conduit route depends on where things actually are.
• Small obstructions and access. The vent stack exactly where row three was going, the locked side gate, the safe ladder point.

Miss any of these and the cost lands later: a second truck roll, a redesign after the homeowner has seen the layout, or a panel surprise on install day. The full minimum list, reader by reader, is in what a residential solar site survey actually needs.

AI solar site assessment approaches, compared

None of these is wrong. Each is right for a different stage of the job. The mistake is using a screening tool’s output as if it were a design input.

Why AI works best on top of measured field data

A model’s output can only be as good as its input. Put AI on a satellite tile and you get fast, useful estimates of what is visible. Put the same kind of processing on a capture made on the roof, in the attic and at the panel, and it starts removing office work instead of adding risk.

After a visit, the time goes to someone scrolling hundreds of photos for the main breaker, re-drawing planes and running shade by hand. AI is good at exactly that work. The visit makes sure there is something real to process; the models make it usable, the same way every time.

Shade shows the split. A remote shade number is a fair screen. A model built from obstructions captured on site and run through every hour of the year belongs under a production estimate. One tells you whether to send a truck; the other tells you what to build.

How to evaluate AI tools for solar site assessment

Ask any vendor these questions, including us. A good tool answers every one without hand-waving.

How Solar Survey AI approaches it

We are the third kind of company in the short answer: AI on measured field data. Our platform starts with a technician on the roof, in the attic and at the panel, and a drone over the top, capturing GPS-tagged photos to one fixed spec. The photos are sorted into structural, electrical and drone sets and pinned to where they were taken. The drone photos become a 3D model, and every roof section gets its pitch, azimuth, area and edge lengths, with each measurement’s tolerance shown next to it. The site lands in your portal once QA signs off.

Our shading analysis walks the sun through all 8,760 hours of a typical year using the trees and obstructions actually captured on site, weighted by typical-year NREL weather, and reports solar access, TOF and TSRF per roof section. The math is published, an accuracy suite runs on every engine change, and reports are marked as not yet program- or lender-approved until independent engineering review is complete. Remote shade reports from public LiDAR and aerial data, for pre-qualification, are coming soon.

The platform is in beta for select clients, and each client sees only its own sites. The site survey itself covers roof and electrical in one visit, with the deliverable 48 hours later. More practitioner notes live in Field Notes.

Questions buyers ask