How does AI lumber grading work?

As we read the published systems, an automated lumber grader turns images of each board into a grade in four steps. Sensors image every face of the board: colour cameras, lasers for shape, and sometimes x-ray for density. A model finds each defect and classifies it as a knot, split, wane, stain or pocket. The grading rules are then applied to those defects, and the system decides where to trim and which grade the board makes. The AI is mostly in the second step, and the grade comes from the rules. That is why an automated grader's accuracy has to be read two ways: whether it gets the grade right, and whether it gets the value right.

What the usual answer says

Vendor pages describe the result. One vendor's page says its system uses "deep learning models trained on millions of board images, enabling real-time detection of knots, splits, wane, checks, stain, and dimensional deviations at line speeds exceeding 400 feet per minute". Pages like it stress consistency against human inspectors, with accuracy figures that cite no study. What they rarely give is a validated accuracy with its method, split in a way a mill can use.

What a validation measured

Rado Gazo and colleagues published a validation in 2018 in Computers and Electronics in Agriculture, read here from its abstract. They ran kiln-dried hardwood lumber of nine commercial species through a multi-sensor scanner "equipped with color cameras, dot-grid and profile lasers and an x-ray sensor to locate and classify defects at a speed of 980 linear feet per minute". Over 1,000 boards of each species were graded by the scanner and checked by a trained human inspector.

The result: "Across the entire volume of boards scanned, the automated grading system was 99.50% on-value and 92.22% on-grade accurate". About one board in thirteen got a different grade from the inspector's. The value of the run came out almost the same. Our reading is that most disagreements fell between adjacent grades, some up and some down, and largely cancelled when the boards were priced. The abstract doesn't say so, and a newer 2025 paper on AI visual grading could not be opened to compare.

Detection is not the grade

An older test shows why the steps matter. A 2011 prototype study by Forest Service and university researchers identified defects "on a pixel-by-pixel basis with an accuracy of 96.7%". Yet "For 40% of the boards, both our software and the NHLA trained grader assigned the same grades". Much of the gap came from one misclassification: the system "occasionally classified natural stains incorrectly as defects". A small error in detection became a large error in grade.

The same paper notes that "certified graders are not always in agreement". No study we opened gives a rate for inspector against inspector. Without one, a 7.78% disagreement between a scanner and one inspector may be close to the disagreement between two inspectors.

Why the two figures matter differently

On-value accuracy is what the accounts see across a run, though an average can still hide errors on particular grades or customers. On-grade accuracy is what the customer sees. A buyer who ordered one grade and receives a board of the grade below has a complaint, whatever the average value of the load. A mill selling by grade to demanding customers needs the on-grade figure for the grades it ships most, species by species.

How to evaluate one

Ask for on-grade and on-value accuracy separately, by species and by grade, measured against a trained inspector on your own boards. Ask how errors split between detection and rules, since the fixes differ: training data or sensors for one, rule settings for the other. After installation, have an inspector regrade a sample each month and track both figures, so model drift shows up as a trend rather than a complaint.

When it doesn't apply

Strength grading for structural timber rests partly on measured stiffness as well as visual rules, and its accuracy is judged against mechanical tests rather than an inspector. Mills that sell mostly on value rather than grade may care little about on-grade accuracy. And the 2018 and 2011 figures describe particular scanners and species. They show what to measure, not what any system today will achieve.

Quarri for sawmills is built around how a sawmill runs, from log intake to shipped order.

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.