What works in AI for forest management, and what is hype?

AI works in forest management where it measures what a sensor sees. It also works where it has been calibrated and checked against the operation's own records. The hype is in generic claims of volume, value or product mix that were never tested locally. A national-scale LiDAR model was released as a preprint in June 2026, trained and evaluated on 32,052 inventory plots. It read dominant height with about 12.3% relative error and total volume with 39%. A 2025 study that calibrated LiDAR on local plots and checked it against logs scaled at the mill found stand-level volume errors of −8 to 6%. The difference between those results is the test for any AI claim in forestry.

What the usual answer says

The top results for this question say AI is revolutionising forest management, and then list challenges: data quality, generalisation from one forest to another, transparency. The FAO's forest monitoring team puts the condition most precisely. In a February 2026 blog post it writes that for applications needing policy-grade credibility, "remote sensing outputs need systematic calibration and independent validation". That is correct, and it leaves a manager asking where the line falls.

What a national-scale model shows

Emilie Vautier, Clément Mallet and Cédric Vega released FLORA in June 2026, a deep learning model that predicts six forest attributes from airborne LiDAR collected under varied conditions. It is a preprint and has not been peer reviewed. The headline result, in their words: "FLORA achieves an rRMSE of about 12.3% (R2 = 0.88) for dominant height and 39% (R2 = 0.74) for total volume".

For the same model on the same plots, the relative error for volume is about three times the error for height. In our reading, that is because the laser measures height almost directly, while volume depends on density and species as well. The authors report that auxiliary ecological and spatiotemporal variables "provide modest overall gains but contribute more strongly to species-specific volume prediction". The model was built to work across a whole country, sensors and seasons. The authors present it as a baseline. It is a fair measure of what an uncalibrated, general model does on volume at plot scale.

What local calibration shows

A 2025 study in Forests by Sparks and colleagues, read from its abstract, tested the other case. It trained LiDAR models on local stem-mapped plots in a mixed conifer forest. It then validated volume against the gross volume of harvested logs, "tracked by load and location and scaled at the processing mill". At stand level, the LiDAR methods "produced stand-level volume estimates with similar errors (−8 to 6%) to the cruise estimated volume (−16 to 6%) when compared with scaled volume".

That is what working looks like: a local model, a decision-scale estimate, and a check against a record the business already trusts.

Where AI works best

It works for detection and change, where the product is a location to visit first. It works for attributes the sensor measures directly, such as height and canopy cover. It works for volume when calibrated locally and checked against scale. It also works for forecasting from an operation's own history, where new records to check against keep arriving.

From Quarri's own work with a forestry operation: a 15-day harvest forecast, trained on four years of daily data, was cross-validated to within about 260 tonnes a day.

Where the hype sits

The hype is in claims that skip the calibration or the check. Examples are volume by product for a stand from a general model, merchantable value for a timber sale from imagery, or carbon to the tonne without field plots. The more a claim depends on species, grade or product mix, the further it sits from what any sensor records. It then needs local validation, which such claims don't always mention.

The question to ask

Ask any vendor or study for the error on the attribute you sell, at the scale you make decisions at, measured against scale records or a cruise you trust. A good answer names the attribute, the scale, the error and the test data. An answer that gives accuracy for height or canopy when you asked about volume is answering a different question.

When it doesn't apply

For monitoring tasks, such as detecting fire, storm damage or insect outbreaks, the product is detection, and a height-against-volume comparison says nothing about them. Research programmes estimating national totals can accept plot-level error that a stand-level sale cannot.

Quarri for forest management is built around how a forest operation runs, from the cruise to the settled account.

Quarri is an AI-native data platform for the timber supply chain. It connects buying, production, sales and inventory for forest management, sawmill, wood products and pulp, paper and packaging operators.