Utility-scale solar and wind plant surveys.

The mission

A hundred-megawatt solar park, three hours from the depot, flanked by a ridge of turbines. The survey crew flies the rows with a thermal payload: hot cells, cracked panels, the junction box running forty degrees above its neighbours — the defect list that decides warranty claims and this quarter’s maintenance plan. On the wind side, the same crew flies blade and nacelle passes tower by tower. The contract is priced per site per day, the analysis pipeline is back at the office, and the plant’s operations room wants to know about anything dangerous before the crew leaves the gate — not when a memory card is couriered in on Thursday.

What breaks

The site has one bar of signal, on one carrier, from a cell built for the village down the road. A morning’s thermography is tens of gigabytes; the uplink is single-digit megabits on a good day. Uploading from site means the crew — the most expensive line in the survey’s budget — sits in a parked vehicle for hours watching a progress bar, so in practice nobody uploads: the data drives home as luggage, the pipeline starts a day or two late, and the finding that needed action today is discovered on Thursday. Meanwhile the genuinely urgent frame — the junction box that is a fire risk — is trapped in the same queue as ten thousand routine images, with nothing to tell the network which one matters.

The architecture on this mission

The survey drone and the crew vehicle each run a node, and the vehicle is the hub. Imagery lands on it over one short local hop at full link speed the moment it is captured — the aircraft never waits on the wide-area network at all. Shoreward of the vehicle, the bond gathers whatever actually reaches the site: the one thin carrier, a second one if a mast exists on the horizon, and the plant’s own service network where the operator offers it. All of it becomes one encrypted tunnel to the operations centre, measured several times a second.

Traffic is classed, and the classes do the triage the memory card never could. A finding the operator flags in flight — the hot junction box, the cracked blade root — travels immediately as high-priority traffic that fits comfortably through even the thin uplink, so the plant can isolate the string within the hour. The full-resolution dataset rides as bulk: it queues on the vehicle, uses every good interval the site link offers, and then does most of its travelling on the drive home. As the vehicle rolls back toward the motorway, coverage strengthens cell by cell; endpoint roaming carries the session across every handover, and the queue drains automatically, priority first, resuming cleanly across each gap. The drive — dead time in every other version of this operation — becomes the fattest part of the pipe.

How the survey day unfolds

1. Arrival. The vehicle parks at the O&M building; its node brings up the bond — one thin carrier plus the plant’s service network. The operations centre sees the crew on site before the first battery is clipped in.
2. The rows. The drone flies grid after grid. Every frame lands on the vehicle over the short hop at local-link speed; the queue builds; nothing waits on the wide-area network.
3. The finding. Row 40: a junction box glowing in thermal. The operator flags it, and the frames travel high-class through the thin uplink at once. The plant isolates the string the same hour — the survey has already paid for its day.
4. Wrap and roll. Last grid flown, aircraft cased, and the crew departs on schedule. The dataset’s remaining gigabytes are a passenger, not a prisoner — the queue rides the vehicle.
5. The drive. Coverage strengthens toward the motorway and the drain accelerates, resuming across every handover and dead patch without anyone touching anything. The site’s dead zone is now behind the data, not in front of it.
6. The debrief artefact. The dataset is in the pipeline before the crew is home, every image timestamped through the journal — which also shows exactly what the site’s uplink delivered, the evidence for whether a permanent node at the plant would pay for itself.

What each mechanism contributes

• Vehicle-as-hub topology — the aircraft offloads at local-link speed and the vehicle owns the wide-area problem; the crew never waits on a progress bar. Shipped.
• Class-based triage — flagged findings travel now through the thin pipe; bulk imagery fills whatever remains. Shipped, opt-in.
• Store-carry-forward — the resumable queue that turns the drive home into the delivery leg; interrupted transfers continue, they do not restart. Shipped, opt-in.
• Endpoint roaming — the session survives every cell handover and re-addressing on the move; the drain never resets. Shipped.
• Continuous measurement — a per-carrier record of what the site really offers, accumulated survey by survey, informing the case for a fixed node or a different modem mix. Shipped.

The honest boundary: Atlas cannot turn a megabit site into a gigabit one — on site, the guarantee is triage and an honest queue, not a fast upload. The aircraft, the thermal analytics, the defect judgements and the plant’s electrical decisions are the operator’s. And a site with genuinely zero signal moves the whole delivery to the drive — which the same mechanisms handle — or justifies a fixed node or satellite link joining the bond.

What a pilot should prove

• Finding-to-notification latency: a flagged defect on the operations centre’s screen, timed from the flag, over the site’s real uplink.
• Full-dataset delivery time against the current courier-or-wait baseline, for an identical survey.
• Queue resumption across the coverage gaps of the actual drive route, shown in the journal — continuations, not restarts.
• A per-carrier record of the site’s usable capacity across the day, good enough to decide on a permanent node.

One plant, one survey day, one drive home. The evaluation format covers the structure.