Grading logs: letter classes, standards and veneer-log scaling

Measuring tells you how much wood there is; grading tells you what it is worth. The grade depends on a standard, on who applies it, and on how consistent the application is.

European standards and national frameworks

Visual quality classification of round timber sits in the EN 1927 series for softwoods and EN 1316 for hardwoods; feature measurement is in EN 1309-3:2018. These standards are paywalled and their class limits are not quoted here. A contract should name the part and edition.

The German RVR classes are:

with a 3 m minimum section length for A and B; mixed grades such as B/C must be agreed. Poland uses a similar A to D ladder, and Austria applies the Holzhandelsusancen unless the sales letter says otherwise. The letter schemes look alike but are not equivalent, so never compare "B" prices across countries without checking the rule.

A case study: grading birch veneer logs by computer vision

A Latvian study (Magaznieks, Millers and Gabrans, 2026, Forests; the three authors are employed by the developer of the system) tested a deep-learning vision system on an operating line, using the Latvian standard LVS 80-7:2024 with Class I, II, III and reject. The class depends on defect size and on whether it reaches the peeling cylinder, the part that becomes veneer.

The system found defects, then a rule module applied the standard. Test set: 153 logs, enriched with rejects (18 percent against roughly 5 percent in normal production), so class-level numbers do not transfer to daily flow.

Two results: core discolouration was found with recall 97 percent and precision 91 percent; dead and rotten knots (52 logs marked by the scaler) had recall 44 percent and precision 77 percent.

Recall is the share of the scaler's defects that the system found. Precision is the share of its flags that the scaler confirmed.

Worked example 1: recall and precision

Dead and rotten knots: the scaler marked 52 logs; recall 44 percent means about 52 x 0.44 = 22.9, so roughly 23 found and 29 missed. Core rot: 21 system indications, 7 confirmed, so precision 7 / 21 = 33 percent; the other 14 were false alarms. The main weakness for knots is missed knots, not false alarms, because bark makes birch knots hard to see.

Worked example 2: class agreement from the confusion matrix

Rows are the remote scaler's class, columns the algorithm's:

• Exact matches (diagonal): 26 + 35 + 5 + 21 = 87, and 87 / 153 = 56.9 percent, the reported 57 percent (Cohen's kappa 0.40, fair agreement).
• Stricter than the scaler (above the diagonal): 6 + 3 + 0 + 20 + 10 + 1 = 40, which is 40 / 153 = 26 percent.
• More lenient (below the diagonal): 13 + 0 + 7 + 0 + 3 + 3 = 26, which is 26 / 153 = 17 percent.

The system rejected 0 + 10 + 1 + 21 = 32 logs, of which 11 (10 Class II and 1 Class III) the scaler considered usable.

The human baseline

Seven line scalers agreed with a control scaler on 79 percent of shortlogs, individual scalers from 69 to 88 percent. This is not directly comparable with the 57 percent (shortlogs versus whole logs), but manual grading is not perfectly repeatable either. The system is birch-specific; the paper offers no evidence for sawlogs.

Key points

• Contracts should name the grading standard, its part and edition; letter classes A to D from different countries are not equivalent.
• The EN 1927 and EN 1316 class limits are not quoted here because they are paywalled.
• Recall is the share of real defects found; precision is the share of flags that were right.
• In the Latvian birch study class agreement was 87 of 153 logs (57 percent); the system was stricter in 40 logs (26 percent) and more lenient in 26 (17 percent).
• Knot recall was 44 percent; human line scalers agreed with a control scaler 79 percent of the time.