Log scanning, optimisation and AI in sawmills: what the open evidence shows

Sawmills have always been measurement businesses. A log arrives, someone or something decides where it is cut, and the value of the result depends on how well that decision matched what was inside the wood. What has changed over the last three decades is how much of the log the decision-maker can actually see, and how fast a computer can search the options. This article follows the chain from scanner to optimiser to machine-learning model, and asks at each link what the open literature actually reports: accuracy, value gain, cost, and limits.

Three conventions apply throughout. Every number comes from a page I opened during research, and the country or region of each study is marked. Where I offer my own reading it is labelled "Interpretation". Vendor statements are labelled as claims and are not used as evidence. The article closes with a list of sources and a list of what could not be verified.

The short version

• The best-documented gains come from using better information at a decision that is already being made: where to cut a stem, how to rotate a log, where to put the opening face, where to edge and trim a board. Reported simulation gains range from about 4% to well above 40%, depending on species, grade, price spread and baseline (tables below). They are not a universal promise.
• Machine-learning defect detection on boards and logs is reported at F-scores and mAP values from roughly 82% to 98% on research datasets. Those datasets are small by industrial standards, and no study opened here reports a controlled, in-mill, before-and-after profit result for deep learning.
• Digital twins and predictive maintenance for sawmills are, in the sources I could open, at the project-definition and laboratory stage, not at the measured-benefit stage.
• Adoption is held back by cost, uncertain returns, data volume and integration work, and the effect is strongest for lower-turnover mills (Swedish interview study, 2025).
• For a small sawmill the sensible path is incremental: measure first, fix the data, optimise the decisions you already make, and only then add learning-based components.

Scanning: what the machines actually measure

External shape: laser and optical scanners

The first generation of log scanners measures the outside of the log. In general terms, a laser line or structured light pattern is projected onto the log and cameras record its shifted position, from which a diameter profile or a 3D surface is computed. Phase-shift scanners work on a related principle. This is general technical background rather than a claim taken from one source.

One open study gives a concrete accuracy figure for a portable 3D phase-shift laser scanner. In Shenga and colleagues (BioResources, 2014; the page I opened does not state the country), cross-sectional areas from the laser scan were compared with CT areas for two logs: a birch log with bark (300 cross-sections) and a Scots pine log without bark (450 cross-sections). The mean area difference was 2.23% for pine (SD 1.54) and 3.73% for birch (SD 0.91), which the authors attributed mainly to bark and surface unevenness. The authors judged the two methods well correlated. Two logs is a very small sample, and a portable research scanner is not an in-line production scanner, so the figures should be read as an illustration of the scale of agreement, not as a performance guarantee for any commercial product.

External scanning is also moving upstream, into the forest. A Skogforsk working report published in June 2026 (Sweden) examined whether laser scanners and a camera mounted on a Komatsu harvester could measure sweep, the long curvature of a stem, in standing Scots pine near Umeå, with terrestrial laser scanning as the reference. For handheld mobile laser scanning at 1-5 m height the reported RMSE was 4.4 mm with a bias of 1.8 mm. For the harvester-mounted scanner, about 74% of trees were detected (202 of 288 reference trees within 1 m), and bow height RMSE at 1-5 m was 9.1 mm for the wide-angle scanner and 29 mm for the narrow-angle scanner (8.5 mm without one large outlier). The report gave no percentage value gain from bucking simulations; it said only that gains were largest for trees with pronounced lower-stem curvature. The same report included a first test of a YOLOv7 model for detecting the defect lyra in mobile-phone images, with an mAP@0.5 of about 0.75 from a training set of only 49 images and a validation set of 12. The authors present these as early results; I treat them the same way.

Internal structure: X-ray and CT

Outside shape only tells part of the story. Knots, resin pockets, rot and density variation sit inside. X-ray and computed tomography (CT) scanners measure attenuation through the log from many angles and reconstruct slices, so knot whorls and decay can be located before the saw. An early ancestor appears in a 1996 USDA Forest Service paper (USA) by Schmoldt, Li and Abbott, in which an artificial neural network labelled each pixel of a hardwood CT image as one of five classes (knot, split, bark, decay, clear wood) using small 3D pixel neighbourhoods; the initial tests used two oak species and one yellow poplar species. The abstract gives no accuracy number, so it is cited here only to show how old the idea is.

For a current, vendor-originated description, MiCROTEC published a news story in April 2026 about a Swedish customer with three mills and close to 800,000 cubic metres of sawn timber per year. It states that CT slices have a resolution of around ten millimetres and that up to 35 sawing alternatives are simulated per log. These figures are claims from a vendor-published page. The same page says CT combined with AI predicts output "with significantly higher precision than before" but gives no figure, no measured recovery gain, and cites no independent audit; the linked Swedish article's URL suggests sponsored content. I use it only to describe what the technology is said to do, not as proof that it does so.

Rotation, positioning and the "horns up" convention

A scanner is only useful if the decision it feeds can be executed. One executable decision is log rotation: the log is turned to a chosen angle before sawing. The conventional positions are "horns down" (Finnish bucking study) or "horns up" (Douglas-fir study below). Both are fixed rules; a CT-based approach replaces the rule with a computed best angle. The reported value of doing so is covered in the next section.

From measurement to decision: optimisation

Value-based bucking

Bucking is cutting a stem into logs. Nordic harvesters usually buck using a two-dimensional diameter profile. A Finnish study by Rummukainen, Makkonen and Uusitalo (Luke, Finland; Wood Material Science and Engineering, 2021) used a sawing simulator on 1,582 Scots pine stems and compared diameter-based bucking against value-maximising bucking using optical scanning alone, or optical plus X-ray CT scanning. They also tested two log positions: the traditional horns-down position and a CT-based value-maximising rotation. The abstract reports that optical scanning alone could raise sawn timber value by up to 5%, while CT scanning before bucking added little value for a sawmill that already uses a CT scanner to optimise rotation. Interpretation: the information has diminishing returns. If a later step already exploits the internal data, collecting it earlier may add cost without adding value.

A Swedish case study by Barth, Möller, Wilhelmsson and colleagues (Annals of Forest Science, 2015) is a reminder that bucking inputs also come from the forest inventory. Using airborne laser scanning (ALS) versus a traditional field inventory as input to bucking simulations validated against harvester data from 16 stands (139.3 ha), the average stand-level error index was 0.15 for ALS and 0.18 for the traditional inventory. At the total wood-flow level the traditional data gave a lower error index because the ALS tree lists had an overall bias. The authors concluded that predictions for log product groups were slightly more accurate with ALS. This is about planning, not about the sawline, but it shows that small improvements in input data translate into small, not dramatic, improvements in prediction.

Rotation and sawing pattern optimisation

The clearest rotation result I could open is a case study on visually strength-graded Douglas-fir by Rais, Ursella, Vicario and Giudiceandrea (Annals of Forest Science, 2017). The authors are affiliated with Microtec in Italy, a CT vendor, so the study should be read with that in mind; the logs came from experimental plots in southern Germany. Thirty-six 4.1 m logs from 20 trees were scanned with an industrial CT scanner at 120 m/min and sawn into centre boards. Across 510 knot-ratio (DAB) pairs from 172 boards, the RMSE between CT and manual measurement was 0.09. A Monte Carlo sawing simulation, including random rotational error, compared the best rotation angle against horns-up. Mean value increases were 4%, 8% and 20% under three price scenarios, with medians of 4%, 8% and 17%. The gain was largest when the price differences between strength grades were largest. Interpretation: the headline "up to 20%" is conditional on a wide price spread between grades. A mill whose products sell at similar prices would see something closer to the lower scenario.

A very large simulated gain comes from US hardwoods. Chang and Gazo (Forest Products Journal, 2009; the page does not state the country but both institutions are in the USA) scanned 29 logs of five species (black cherry, hard maple, yellow poplar, red oak, white oak) with medical-grade X-ray CT, sawed them at a mill to measure actual recovered value, and sawed virtual logs built from the scans in TOPSAW software. Simulated value gains averaged 46% across species and grades. By species the gains were 42% (black cherry), 33% (hard maple), 83% (yellow poplar), 24% (red oak) and 60% (white oak). By grade they were 27% (Grade 1), 47% (Grade 2) and 97% (Grade 3). The abstract does not spell out precisely what the baseline sawing strategy was, and the sample is 29 logs, so I would treat 46% as a statement about low-grade hardwood logs with strong internal defects under simulation, not as an expectation for a typical softwood mill. Note also that the gain is largest where logs are worst, because that is where information about defects changes the best cut most.

Optimisation for small mills

A study aimed squarely at small mills is Lin, Wang and Thomas (Wood and Fiber Science, 2011; USA, central Appalachia). They built a 3D log sawing optimisation system that creates 3D log models, selects the opening face, simulates sawing, grades lumber by National Hardwood Lumber Association rules and uses heuristic and dynamic programming algorithms; internal defect position and shape came from a USDA Forest Service model. Using log data from five local sawmills, average lumber value recovery rose by 10.01% with the heuristic algorithm and by 14.21% with dynamic programming, and recovery of select-or-better lumber rose by 16-30%. The authors concluded that the system could help small sawmill operators improve performance. Interpretation: the interesting detail is the gap between the two algorithms. The better search found about four additional percentage points, which suggests that the optimisation algorithm matters as well as the sensor, and that algorithmic improvement can be a lower-capital route than new hardware. Again this is a simulation, validated against sawmill values, not a measured production trial.

Real-time versus simulation

Two words are often blurred, so it helps to separate them. Simulation-based studies, which dominate the open literature, take scanned logs and test sawing strategies on a computer after the fact. They answer "how much could this be worth?" Real-time optimisation runs inside the mill's cycle time: the scan arrives, the optimiser searches candidate solutions, and the result goes to the positioning equipment before the log reaches the saw. The Microtec story above, for example, describes testing up to 35 alternatives per log in that setting, though as a vendor claim. The open studies I opened do not report cycle-time budgets or the compute needed to run production-speed optimisation, so I cannot give figures for that. What the simulation studies can tell a mill is the ceiling; the achievable fraction depends on positioning accuracy (the Douglas-fir study explicitly modelled a random rotational error for this reason), on how often the scanner data are wrong, and on whether the downstream process can execute the plan.

Boards, edgers and trimmers

After the headrig, the same logic applies to boards. A rough, green board has wane, knots, splits and decay, and the edger and trimmer decide where to cut it. A USDA Forest Service prototype from Lee, Abbott, Schmoldt and Araman (USA, 2003) is a documented early example. It scanned rough, green hardwood boards using commercial lasers and a video camera, recording thickness and reflectance at 1/16-inch (1.6 mm) resolution. Wane boundaries were detected with an average error of 3/16 inch (5 mm). A modular artificial neural network classified clear wood, knots and decay with a pixel-level accuracy of 96.7%. A software step then chose cuts maximising commercial value at current prices and reported lumber grades. The authors noted that scanning right after the headrig can cut losses more than scanning later in the process, but few researchers had studied it because fibres and debris on green surfaces make imaging harder. It is a 2003 prototype, not a recent result.

A 2000 survey by Bowe (hardwood sawmills; the page says "nationwide" but does not name the country, and it is from a US university repository) gives a rare view of adoption in this area. Of mills responding to a mail survey sent to over 2,000 mills, with average annual production of 7.6 million board feet, headrig optimisation was used by 27%, edger-optimisers by 10% and trimmer-optimisers by 5%. Improved raw material recovery and higher lumber revenue were the most highly rated adoption factors, and the author concluded that adoption in the hardwood sawmill industry was low. That is a 1999 snapshot and certainly out of date as a number; I include it to show the order of magnitude of uptake once, and because the factors it named echo later findings.

Machine vision and deep learning for defects

Deep learning enters the picture mainly through image-based detection. The studies below were all found in open-access venues; none is a long-term in-mill trial.

How should this table be read? First, the metrics are not comparable across rows: F-score, mAP at a given threshold, pixel accuracy and species accuracy measure different things, and datasets differ. Second, the best number (97.7%) comes from a dataset in which, according to the paper, all knot types were grouped as a single category, which is an easier problem than distinguishing sound knots, dead knots, and decay. Third, the FDD-YOLO result of 82.3% mAP, on a dataset partly collected at a sawmill with seven defect types, is a more sobering indication of what harder, more realistic data can give. This is my interpretation, but it follows from the dataset descriptions in the papers. Fourth, the Skogforsk test shows what happens with a very small training set: usable but modest.

A 2026 systematic review in Forests by Kucuk (Türkiye) and Buehlmann (USA) gives the wider picture. The authors searched Web of Science on 5 November 2025, found 7,349 records and included 50 peer-reviewed English-language publications from 2014 to November 2025, with an annual growth rate of 23.28% and a considerable increase in AI-focused research after 2023. The literature reviewed was led by China, Canada and Poland, with China having the most corresponding-author publications (9). Explainable AI was used in only three of the studies. Interpretation: a field with 50 qualifying papers in eleven years, mostly laboratory or dataset studies, is still young, and the review's conclusion about the limited use of explainable AI matters for mills, because an operator who cannot see why a model rejected a board will find it hard to trust or correct it.

Species and grade classification

The one species-identification study I opened, Liu and colleagues (Forests, 2025, China), uses microscopic wood sections: 3,926 images of 32 species, mapped to 25 IAWA anatomical feature codes, reaching 94.1% accuracy. Microscopy is not a sawmill-line technique. It is relevant for laboratories and for checking species claims, not for sorting logs on a conveyor. I did not open a study that demonstrates automated species or grade classification at line speed in an operating mill, so I make no claim about it. Interpretation: for most European sawmills processing a small set of species, species recognition is not the bottleneck; grade and defect prediction are.

Predictive maintenance and digital twins

Predictive maintenance

Saw condition affects both yield and downtime. One open article, Muhammad, Khan, Kamal, Ahn and Oh (IEEE Access, 2024; the page does not state the country, though the listing I saw came from the Seoul National University of Science and Technology repository, so South Korea is likely), collected six-axis vibration data from a vertical bandsaw in normal and faulty states and trained classifiers. The fault types included loose or tight wheel pulleys, blade and motor belts, blade guide looseness and cutting-bed tilt. The best classifier reached 93.22% accuracy on clean data and 87.34% on corrupted data. It is a machine-condition result on a vertical bandsaw, not a result from a production sawmill line, and it says nothing about avoided downtime or money. I did not find a peer-reviewed open study that reports maintenance savings at a sawmill, so I cannot quote one.

Digital twins

A digital twin is a data-linked model of a machine or a process that is kept in step with the real thing. In sawmills the documented open examples are early-stage. A Vinnova-funded initiation project (Sweden, June to November 2019, SEK 500,000, coordinated by RISE) ran workshops and calls to define opportunities, requirements and use cases for a digital twin of a sawmill in Karlskoga and to map existing information. The project page reports process outcomes only, such as a larger consortium, and no measured improvement in production or maintenance. A Rosewood 4.0 factsheet describes a related Swedish "Digital Sawmill" project with Moelven Industrier, RemaSawco, Schneider Electric, RISE and DigiWood, whose goals were to raise process efficiency by 15%, raise product value by 10% and cut energy use by 10%, and states a calculated potential of about SEK 5 million per year. Both numbers are targets or projections, not outcomes, and the page gives no method or data, so I mark them as unverified. The same factsheet describes the product family as focused on log-to-board traceability.

Data infrastructure: the part that decides whether the rest works

A pattern runs through the sources. The Nowik master's thesis (SLU, Sweden, 2025, based on semi-structured interviews with sawmills that do and do not use CT) found that implementation is difficult because of the large volumes of data involved. It also reported that early adopters use CT for log-to-product traceability and for determining possible products before sawing, and that new regulations make traceability measures necessary. That is an interview study, so it reports perceptions rather than measured performance.

The practical content of "data infrastructure" is mundane. Each log needs an identifier that survives from yard to scanner to sawline to board. Scanner outputs, saw settings and board results need to be stored with timestamps in a form that can be joined. Without that join, nobody can say whether the optimiser's predicted product matched the actual one, and without that feedback no model can be improved or even audited. Interpretation: this is the most underrated investment, and it is also the one that makes later steps cheaper. The Douglas-fir and TOPSAW studies both depended on matching scanned logs to the boards sawn from them, which is the same join done by hand for research.

What the economics look like

Collecting the reported gains in one place, with their caveats:

Two observations. First, all four values are simulated or modelled gains. None is a before-and-after mill result with the same logs and market. Second, the spread is enormous, from 4% to 97% for Grade 3 logs, which says the gain is highly situational: species, log quality, price spread and the existing process all matter. Interpretation: a mill should treat these figures as an argument for measuring its own baseline, not as a forecast. The only cost-side evidence I opened is qualitative: the Swedish interview study names high upfront cost, uncertain returns and a complicated implementation process, and says CT is spreading slowly, especially in mills with lower turnover. No study I opened gives purchase prices, and vendor prices are not public in the sources I used, so I give none.

Limits and barriers

Evidence quality. The numbers above are mostly simulations, small samples and benchmark datasets. The sample sizes are 29 logs, 36 logs, 1,582 stems, a few hundred images up to a few thousand.

Transferability. Hardwood grading rules (NHLA in the USA) differ from European softwood strength and appearance grading, and Nordic pine differs from Douglas-fir or Appalachian oak. A result in one setting says little about the next.

Domain shift. A model trained on sanded, clean samples or on lab images may perform differently on rough, wet, dirty, moving boards. The Lee et al. paper notes that fibres and debris near the surface make green lumber imaging harder.

Explainability. The 2026 review found explainable AI in only three of 50 studies. Mills need to know why a system decided what it did.

Integration and data volume. The Swedish interview study identifies large data volumes as a hard part of CT implementation.

Evidence of maintenance and twin benefits. I found none that is measured and open. Promises in this area should be treated as unverified until a mill reports figures.

Vendor claims. The vendor page I opened contained no measured gains and cited no independent verification. Vendor demonstrations can be valid, but a buyer should ask for the data on a comparable mill, with the baseline stated.

Small and medium sawmills

The Appalachian work by Lin and colleagues is the only open study I opened aimed explicitly at small mills, and its message is that optimisation software on top of modest scanning data can recover several to over ten percent of value in simulation. The Swedish thesis is the main evidence on barriers: cost, uncertain return, complexity, and slower spread among lower-turnover mills, with the author expecting that future CT scanners may become smaller and easier to integrate. The 2000 US survey indicates that, historically, adoption was concentrated in headrig optimisation, and edger and trimmer optimisation was rarer. Interpretation, which is not drawn from a single study: the small mill's problem is less "which AI?" than "what is my current recovery, which decisions are made by eye, and which of them is most worth automating?" Outsourced or shared solutions (service-based software, shared scanning of purchased logs, regional co-operation) may fit smaller budgets, but I did not open evidence on how well these work, so this is a hypothesis.

A roadmap for a small sawmill

This roadmap is my own synthesis from the evidence above, not a validated method.

1. Measure the baseline (months 0-3). Record, per log batch, input volume, diameter class, species, sawn output by product and grade, and prices. Without a baseline, no later gain can be demonstrated. The simulation studies above all rely on a comparison against a baseline, and so must you.
2. Fix identity and data (months 1-6). Give each log or batch an ID that follows it through the process, store the data in a joinable form, and record the saw settings. This is the cheapest step and the one that every later step depends on.
3. Use a decision where you already have information (months 3-9). If you already have a diameter or profile scanner, check how its data are used in bucking, sorting and sawing pattern selection. The Finnish bucking study suggests the gain from adding better external data can be a few percent; the algorithm comparison by Lin and colleagues suggests software search quality matters.
4. Add scanning where the value spread is largest (months 6-18). The Douglas-fir study and the hardwood CT study agree that gains grow with price spread between grades and with lower log grades. If your product range has a narrow price spread, the case for internal scanning is weaker. Request vendor trials on your own logs, with the baseline stated.
5. Pilot learning-based inspection on one narrow problem (months 9-24). For example, one defect class at one station, with human review in parallel. Keep a labelled image archive, and measure precision and recall on your own boards, not on a benchmark.
6. Postpone digital twins and predictive maintenance until you have sensor data and a maintenance log. Start with simple condition monitoring on the most failure-prone machine, such as a bandsaw, and compare alerts with actual events before relying on them. The open evidence for savings is not yet there.
7. Review annually. Compare realised recovery and value against the baseline and against the figures in the literature.

Sources used

• Rummukainen, Makkonen, Uusitalo - Economic value of optical and X-ray CT scanning in bucking of Scots pine (Luke/VTT record, 2021)
• Chang, Gazo - Measuring the effect of internal log defect scanning on the value of lumber produced (Forest Products Journal, 2009)
• Rais et al. - First industrial X-ray CT scanner increases lumber recovery value: Douglas-fir (Annals of Forest Science, 2017)
• Barth et al. - Swedish case study on product recovery from stem profiles based on airborne laser scanning (Annals of Forest Science, 2015)
• Lee et al. - A system for optimal edging and trimming of rough hardwood lumber (USDA Treesearch, 2003)
• Lee et al. - A prototype scanning system for optimal edging and trimming of rough hardwood lumber (USDA Treesearch, 2003)
• Schmoldt, Li, Abbott - CT imaging of hardwood logs for lumber production (USDA Treesearch, 1996)
• Lin, Wang, Thomas - 3D log sawing optimization system for small sawmills in central Appalachia (Wood and Fiber Science, 2011)
• Lin, Wang, Thomas - same study, USDA Treesearch record
• Lindgren et al. - Measurement of sweep and timber quality with mobile sensors on harvesters (Skogforsk Arbetsrapport 1275, 2026)
• Shenga et al. - 3D phase-shift laser scanning of log shape (BioResources, 2014)
• Fang et al. - Surface knots on sawn timbers using YOLO-v5 (BioResources, 2021)
• Wang et al. - FDD-YOLO: surface defects in wood (Forests, 2025)
• Yang et al. - Improvement of YOLO detection strategy for defects in bamboo strips (Forests, 2025)
• Liu et al. - Interpretable wood species identification from microscopic images (Forests, 2025)
• Reich et al. - Multimodal surface defect detection from wooden logs for sawing optimization (arXiv preprint, 2025)
• Kucuk, Buehlmann - Systematic review of AI applications for wooden materials (Forests, 2026)
• Muhammad et al. - Intelligent troubleshooting of vertical bandsaws (IEEE Access, 2024; DOAJ record)
• Vinnova - Digital twin of sawmill for efficient production and maintenance (2019)
• Rosewood 4.0 factsheet - DigiWood digital sawmill traceability tools (targets, not measured results)
• Nowik - The adoption process of CT scanning in sawmills (SLU master's thesis, 2025)
• Bowe - Modeling the adoption decision process of scanning and optimizing technology in hardwood sawmills (Virginia Tech, 2000)
• MiCROTEC - With CT, X-ray and AI, sawmills can see inside the wood (vendor page, 2026; claims only)

Limits of this article

• Research method. I read pages one at a time and used abstracts and, where available, the page text. Full texts were not read in every case, so details in methods sections may differ from what I report. Some pages were paraphrased by the retrieval tool and numbers were taken as given.
• Not found or not opened. I did not open peer-reviewed open studies reporting measured, in-mill, before-and-after profit from deep learning, digital twins or predictive maintenance. The statement that such evidence is thin reflects my search, not a proof that none exists.
• Unverifiable points. The SEK 5 million and the 15%, 10% and 10% targets from the Digital Sawmill factsheet (targets and a calculation, no method). All MiCROTEC statements (vendor claims). The country of three studies is not stated on the pages I opened (Chang and Gazo 2009, Shenga 2014, Muhammad 2024; the first is inferred from affiliations, the third from a repository listing). The baseline definition in the Chang and Gazo abstract.
• Dates and currency. Several sources are old (1996, 2000, 2003). They are used for context and labelled as such.
• General background. Descriptions of how laser scanners and CT work, and the definition of a digital twin, are general technical background and not taken from a single opened source.
• Interpretation. Statements labelled "Interpretation", the roadmap and the small-mill reasoning are my own and are not findings of the cited studies.
• Not advice. Gains in simulation studies do not guarantee gains in your mill. Seek independent technical and financial advice before investing.