LiDAR · Utility Network · Metadata, End-to-End

01The question

Can free public LiDAR replace paid utility survey flights?

Utilities pay for dedicated survey flights to map their lines. Meanwhile USGS publishes open LiDAR over most of the US for nothing, but it is far sparser: enough to see the ground, not obviously enough to find a wire.

A corridor as raw LiDAR: a cloud of uncoloured points with no labels — Raw returns: position only, nothing named.

The same corridor after segmentation, each point coloured as ground, vegetation, pole or wire — After segmentation: every point is ground, vegetation, pole or wire.

02The problem

Most open LiDAR arrives sparse, which is the data this has to work on. At that density ground and trees still hold their shape, but wires fall apart into scattered points and poles smear into the vegetation: the two classes the whole product depends on. Scores below are IoU, the overlap between prediction and truth, where 1 is perfect.

Same tile at 5 points per square metre, model trained at full density only: wires break up, poles vanish — 5 pts/m², the open-data case: wires break, poles vanish.

wire IoU 0.58pole IoU 0.42

Same tile at 5 points per square metre, model trained across densities: wires and poles intact — Trained across densities: wires and poles come back.

wire IoU 0.95pole IoU 0.65

at full density both scorewire IoU 0.96pole IoU 0.71

03What I built

LiDAR
segmentation
topology
wire separation
metadata

Segmentation that survives sparsity

The model is trained across the range of densities it will meet in the field, not just the dense end, so sparse scans are the normal case rather than the exception. A large frozen encoder is gated into a small trainable backbone, adding its knowledge at no inference cost.

Topology in one forward pass

A four-channel height map feeds a dual-head network that predicts poles and wires together in one pass. The channels that depend on point labels are dropped at random in training, so the network still works when a scan is badly classified.

4-channel DSM dual-head ResUNet-FPN

One span → individual conductors

A learned separator and a geometric one fail in opposite situations, so a gate picks the better one per span. A catenary curve is then fitted to each conductor.

A span of seven transmission wires: raw above, separated into one colour per wire below — One span: raw, then one colour per conductor.

Metadata, one record per asset

With every conductor a fitted curve, the record writes itself: wire endpoints, pole heights and vegetation clearance. Clearance is the number that tells a utility where to trim before fire season.

wire endpoints and sagpole heightvegetation clearance

04Results

Segmentation

Measured on open datasets under one protocol, with no retraining between them: real USGS 3DEP scans at survey-grade sparsity, and the DALES aerial benchmark. Against a pre-trained point transformer used as is, wires and poles go from unusable to usable on the sparse data, and improve on the benchmark too. The numbers are in a paper under review; they will be here when it is out.

Network

≥ 0.91pole detection, 0.4 to 193 pts/m², clean to badly classified
99.8%precision assigning points to individual conductors

Poles become nodes and spans become edges, assembled across tiles so a span that crosses a tile boundary is still one edge. Figures from the BMVC 2026 paper.

Two panels for one tile: the combined pole and wire channel the network reads, and the pole-and-span graph it reconstructs — One tile: the combined pole and wire channel, then the graph it reconstructs. Eight poles in, eight out.

Metadata

12–18 cmerror on conductor endpoint height and sag

Every conductor is a fitted curve through real returns, so each measurement is a simple query on the curve. Vegetation clearance is the number that decides where a wildfire can start.

A schematic span between two poles, turning slowly, with endpoint height, sag, ground clearance and vegetation clearance called out — What one span becomes: a schematic of the record.

05In production

segmentation
topology reconstruction
metadata

Live in production. Six scale-to-zero endpoints, each on the instance type it needs, so shipping a new model is a configuration change.

AWS SageMaker Terraform MLflow model aliases

06Impact

500K+line miles · US, Canada, Europe · public, customer-flown and mobile LiDAR
50K+line miles in a single job

Per-span vegetation clearance shipped as a product: where to trim before fire season.
Next: where no LiDAR was ever flown, teach a depth model the canopy heights this pipeline already produces, so imagery alone can fill the gaps.