How well does AI grade logs, according to the research?

It depends on what grade means. For identifying species from photographs, a 2026 study using off-the-shelf image models reached 92.6% on 5,549 log images from a real auction, with two look-alike oak species merged into one. Kept separate, accuracy fell to 79.1%. For judging value from the same photographs, the models did much worse. Market value classification scored between 40% and 55%. The study sorted each species' logs into three value bands: the bottom quarter, the middle half and the top quarter. Always guessing the middle band would be right about half the time. Two of the four species scored below that, and the best two beat it by about five points, with modest ability to rank logs by value. The authors' explanation is that value depends on things a photograph can't show.

Source: Triplat, Lukančič and Kavčič, "Evaluation of Deep Learning Models for Image-Based Classification of Timber Logs by Market Value", Forests 17(5): 518, published 23 April 2026, doi.org/10.3390/f17050518. Open access under CC BY 4.0.

What the usual answer says

Searching the exact question returns results about grading student essays and reading system logs, so there is little plain answer on the web. The research itself is hard to compare. A 2024 review of 33 studies found that "comparing performance across studies proved challenging due to varying goals and metrics"; we read its abstract. Studies using CT scans, such as a 2026 paper whose abstract reports networks that "accurately distinguish healthy and damaged regions", work on internal features a photograph can't show. This study is useful because it separates recognising a log from valuing it on the same images.

What the paper did

The researchers used photographs taken at an auction of high-value logs in 2023, before results were known. Each log's final sale price per cubic metre came from the auction's own database, merged with the images by log ID. They generated image features with two pre-trained networks, without retraining them, and classified with logistic regression, using 10-fold cross-validation. The authors call this "a deliberate design choice aimed at evaluating the applicability of accessible, off-the-shelf tools for forestry practice".

Species came first, in two versions: eight classes with the two oaks merged, and nine with them separate. Value came second, for the four species with enough images "to allow reliable subdivision into value-based quartiles". Within each species, the images "were divided into quartiles based on assortment value per m3". The classes were the lowest 25%, the middle 50% combined, and the highest 25%.

What it found

The abstract reports "high accuracy for tree species classification (up to 92.6%)" and "substantially lower accuracy for market value classification". The full text gives the detail. With the oaks separate, one of them was classified correctly 66 times out of 275 and mistaken for the other 192 times. Value accuracy was 55.3% for larch, 54.8% for sessile oak, 45.2% for sycamore and 40.4% for spruce. The ranking score (AUC) ran from 0.568 for spruce to 0.713 for oak, where 0.5 is chance. The authors attribute the gap to "the greater complexity of value determination from visual features".

The conclusions explain why. Market value appraisal, they write, "extends beyond visual attributes to encompass dimensions, morphology, and internal wood properties". Those are "challenging to derive from images alone". Value "hinges on partially non-visual characteristics indiscernible from imagery".

The spread of prices shows the difficulty. For one species, the most valuable log sold for more than 100 times the species' median price per cubic metre. A few logs like that make value extremely uneven, and the features that earn those prices are often inside the wood.

Reading it through a data lens

The paper's value accuracy sits in a range that straddles the naive baseline its own design implies. That comparison is ours, not the authors'. They call value performance "suboptimal" and present the results as "a context-specific baseline rather than a universally applicable model". A value model should always be reported against the score of the simplest possible guess and a ranking measure. Otherwise 55% can look like more progress than it is.

The authors point to the fix: "multimodal fusion, such as 3D laser scanning for precise dimensional quantification and acoustic sensors" for internal properties. In data terms, value needs more than one record per log. It needs the image, the dimensions, the internal scan and the sale outcome joined by the same log ID. The study could join image and price only because the auction kept both under one identifier.

A worked check

Say a buyer tests an image model that sorts incoming logs into three value bands, with half of all logs in the middle band. The model scores 52%. The first check is the baseline: always saying middle would score 50%. The model adds two points, which is well within the noise of a single test on a few hundred logs and nowhere near enough to price wood on. Now the buyer joins each log's scan dimensions and its sawn value from the mill to the image. Rerun on the joined record, the same test shows whether the extra data moves accuracy well clear of the baseline. The figures are illustrative.

Where it touches an operation

The paper could join image and price only because the auction kept both under one log ID. A mill that joins its own dimensions, infeed scans and sales outcomes by log or load has what the study lacked. Without that join, an image model can confirm species and flag visible defects, and little more.

What we cannot take from it

The study used high-value auction logs, not the everyday sawlogs most mills buy. It tested frozen, off-the-shelf image models, not purpose-trained networks or CT and multi-sensor systems, which read internal features directly. Its value labels are auction prices, which carry market noise from one day's bidding. And the naive baseline comparison is our reading of its class design.

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.