From Drone Capture to a Roof Drawing: A New AI Agent for Tiled Roof Analysis

Q: Is one drone photograph sufficient for measurement?

The described process relies on threedimensional reconstruction and suitable source photographs. A single image does not by itself verify areas, slopes or heights.

Can drone capture of a tiled roof become a useful geometric drawing for solar planning?

Understanding roof geometry is still a substantial manual task when planning rooftop solar systems. Flat roofs are often simpler to interpret. Pitched tiled roofs introduce a much more complex arrangement.

A roof can contain planes at different angles, ridges, eaves, hips, valleys, changes in height, chimneys, solar water heaters, collectors, skylights, trees obscuring the building and other features that affect the area available for installation.

At Skylens, we are developing a spatial agent that explores how this work can move from manual drawing to a process grounded in three-dimensional data.

The goal goes beyond detecting that a roof is present. We want to understand how the roof is built.

The problem: an image of a roof is not its geometry

Vision and segmentation models can identify image regions that appear to belong to a roof. For solar planning, that is only part of the task.

We want to reach a structure such as:

Building → Roof → Roof Planes → Structural Lines → Measurements

This means identifying the geometric components as well as the outline:

Individual roof planes.
Ridge lines.
Hip lines.
Valley lines.
Eaves and roof boundaries.
Junction points.
Slope and orientation of each surface.
Area of each surface.
Relative and absolute height.
Obstacles and equipment on the roof.

This distinction also matters in the Skylens data model: the visible roof outline (roof_outline) is not necessarily the building's footprint on the ground.

Looking at geometry as well as pixels

The approach we are testing starts with the three-dimensional information produced from drone capture.

The workflow begins with a Reality Model of the building—for example, a point cloud, photogrammetric model, digital surface model (DSM) and the original drone photographs.

After isolating the building, the agent analyzes roof points and estimates the orientation of local surfaces. The proposed processing chain is:

Point Cloud → Surface Normals → Region Growing → Plane Fitting → Plane Merging → Plane Intersections → Roof Topology

Instead of asking only where the roof appears in an image, we ask: which points belong to the same physical roof surface?

Step 1: identifying planes

A normal can be estimated for each point in a cloud. It is a vector describing the orientation of the local surface.

Thousands of points on the same roof slope should have similar orientations. Combining normal similarity, height and spatial connectivity allows points to be grouped into candidate roof planes.

Plane fitting then attempts to fit a mathematical plane to each group. A cloud containing millions of points starts to become a simpler semantic representation:

Roof Plane A: area, slope, orientation and height.
Roof Plane B: area, slope, orientation and height.
Roof Plane C: area, slope, orientation and height.

These remain candidates whose support and boundaries need review.

Step 2: from planes to roof lines

The mathematical intersection of two adjacent, non-parallel planes defines a line. That line becomes a candidate for a physical roof feature.

Using plane orientation and topology, the system can attempt to classify structural lines and boundaries as ridge, hip, valley, eave or boundary.

This changes the extraction process: some lines can be derived from the intersection of reconstructed three-dimensional surfaces, rather than only from edges in a two-dimensional image. Eaves and outer boundaries also require evidence for where the roof actually ends.

Step 3: building a roof graph

The intended result is a topological model rather than a collection of unrelated lines. For example, a proposed Roof 01 might describe:

Plane A connected to Plane B through a ridge.
Plane B connected to Plane C through a hip.
Plane C ending at an eave.

The roof becomes a graph of surfaces and their relationships.

Skylens follows an entity-first approach: objects receive stable identities, geometry, evidence and spatial relationships instead of remaining isolated files or one-off AI masks.

What about solar panels?

This is where the connection to the solar industry becomes particularly useful. Once roof planes are understood, an additional detection layer can describe existing solar installations and other roof features.

The Skylens data model defines a solar array as a separate entity that can be linked to a building and the roof surface supporting it. It can hold area, azimuth, estimated tilt, roof coverage and an estimated module count when source quality supports the estimate.

The intended workflow would use a site capture campaign to establish:

Roof Geometry + Existing Solar Arrays + Obstacles + Available Roof Areas

That information could then support planning.

Why could this matter to solar companies?

In traditional workflows, some of the information needed for roof planning is collected and entered manually. The agent aims to reduce that initial work.

Instead of starting from a blank drawing, a planner could receive a roof draft generated from the captured geometry, containing:

Roof boundaries.
A division into planes.
The area of each plane.
Slope.
Azimuth.
Height.
Ridge, hip and valley lines.
Existing roof features.
The evidence used to derive the geometry.

The planner receives an initial model, checks it, corrects it where necessary and approves it. In the intended architecture, AI output remains an editable, reviewable draft until review establishes its approved status.

Human in the loop: supporting the planner

The system is not presented as a replacement for an engineer, surveyor or solar designer. The goal is to give the professional a better starting point.

Instead of drawing the entire roof again, the professional could review a proposal based on the reconstructed geometry: approve a surface, move a line, split or merge surfaces, remove an incorrect detection or add a missed feature.

In the future, those corrections could provide better training examples. Only reviewed and approved outputs should be treated as trusted labels.

The goal: roof intelligence beyond roof detection

Roof detection is the first step. Our goal is a model that can answer questions about each roof:

What is the area of the south-facing plane?
What is its slope?
What is its azimuth?
How much of the surface already contains panels?
Which roof areas are available?
Where are the ridges and obstacles?
What has changed since the previous capture?

When the building is captured again, maintaining its identity would make it possible to compare versions over time.

Where are we today?

The system is under development, with algorithms being tested on real roofs.

We have moved beyond the simple idea of Segment Roof → Convert Mask to Polygon toward an approach in which three-dimensional geometry is the primary measurement source, while images support classification, evidence and verification.

The current challenge is turning the collection of planes and lines into a stable, measurable roof topology that is easy to correct.

The test is whether the result saves time in professional work, not simply whether an algorithm produces an attractive image. This article reports a development direction; it does not report validated accuracy or measured time savings.

Looking for a pilot partner

We are looking for a company in the solar industry willing to evaluate this workflow on several real roofs.

For the pilot, we want projects with both capture data and completed professional designs, allowing us to compare Skylens output against the existing work.

We plan to evaluate:

Geometric accuracy: agreement between proposed lines and planes and reference measurements.
Area accuracy: differences in total roof and individual plane areas.
Slope and orientation accuracy: tilt and azimuth.
Completeness: how many real planes and lines were identified.
Correction time: how long a planner needs to turn the AI draft into an approved result.
Time savings: a comparison between working from scratch and reviewing an automatically generated roof draft.

The business question is straightforward: can drone capture produce a reliable roof draft that substantially shortens the path from an existing roof to solar planning?

If it can, a process currently combining capture, measurement and manual drafting could become a spatial workflow in which much of the initial work is automated, while the professional retains review and approval responsibility.

Skylens — Turning captured reality into structured spatial intelligence.

Frequently asked questions

Does the system already produce approved solar designs?

No. The agent is under development and testing. Its goal is an evidence-based roof draft that a professional reviews and corrects before approval.

Is one drone photograph sufficient for measurement?

The described process relies on three-dimensional reconstruction and suitable source photographs. A single image does not by itself verify areas, slopes or heights.

What will the pilot evaluate?

The pilot will compare geometry, area, slope, orientation, completeness and correction time against existing professional work. No validated accuracy or time-saving results are reported here.