Editor's note
AI in forestry has left the demo phase.
Six months ago, half of this newsletter was about "look what AI can do in a forest." Now what crosses my desk is operational: AI quietly running on factory lines, models retrained for local species, classifications running on sensor streams already deployed. The hype cycle has rotated.
For builders, the implication is concrete. The market has shifted from buying convincing to buying delivery. Customers want AI that does the dull job tomorrow morning at a unit cost low enough nobody argues about it. The technology bar is lower than it was; the integration bar is higher.
The question now is what hasn't been automated yet. Plenty of forestry work that should be running on inference already isn't, often because nobody has bothered to wire it up. That backlog is where the next year of stories will come from.
Axel
WHAT GOT ME THINKING
NIBIO Adds Defect Detection and a Three-Tier Maintenance Rating to RoadSens
Two new papers from NIBIO's SFI SmartForest centre this month expand RoadSens into a combined geometry-and-condition toolkit for forest roads. The newer paper detects six surface defect types (potholes, ruts, gullies, stones, washboards, vegetation) and rates each road segment by maintenance priority on a three-tier scale; the earlier one classifies road geometry automatically against the Norwegian forest road standard from the same sensor pass.
Axel's notes: Forest road networks are big, dispersed, and expensive to inventory at any useful resolution. Driving every kilometre with a total station has cost enough that most owners survey on a sample basis, or lean on the contractor who knows the roads best. A traffic-light view of the whole network from a single drive-by changes that cost line, and turns the experienced operator's running judgement into a layer the next person can act on.
I keep coming back to this one and it keeps proving out. The most productive tech work in forestry right now has very little to do with new sensors, and a lot to do with what AI and modern models can pull out of data that was already collected on rigs already deployed. The defect map here comes from the same drive, the same data files, no new survey campaign. Mabema and Biometria did the same trick on Södra's pulpwood last month (BTB #041): existing GPV stack cameras, no new hardware, new question answered in real time. Forestry data is heading that way.
Arboair runs its first British forest survey at the Duchy of Cornwall
Arboair is in Cornwall this week surveying woodland at the Duchy of Cornwall, with Airpelago handling drone data collection. The deliverable is a single-tree-based forest management plan built from the imagery via Arboair's AI. The duchy manages over 6,000 acres of woodland, and this is the Linköping-based company's first British engagement.
NUS lab uses AI to map biodiversity in Southeast Asian rainforests from sound
NUS's LAPIS lab is building a bioacoustic monitoring stack for Southeast Asian rainforests. Autonomous recorders capture continuous forest audio. An AI pipeline turns each stream into a spectrogram, then classifies it at two levels: human-made noise versus wildlife, and bird species. The team is retraining classifiers on locally-collected audio after Northern-hemisphere bird models missed regional species.
Holmen, SCA, Stora Enso and Sveaskog commit SEK 20M to AirForestry's drone-thinning pilots
The Swedish AirForestry Pilot Project (SAPP) deploys autonomous electric thinning drones in live operations at each of the four forest owners, with the system locating, pruning, felling and transporting individual trees from the air. The funding is contracted pilot revenue rather than grant money, marking the company's first customer-paid field deployment alongside its existing Vinnova-funded PADA consortium (BTB #043). The announcement landed the day after AirForestry felled its first tree by drone in a real forest outside the test site (BTB #046), pairing the technical milestone with a commercial one.