Logo image
Back
Preliminary development of an interpretable machine learning tool to verify treated timber penetration compliance
Conference paper   Peer reviewed

Preliminary development of an interpretable machine learning tool to verify treated timber penetration compliance

Zidi Yan, Zhuoyang Xin, Yuyang Zhu, Jack Norton and Tripti Singh
Proceedings IRG Annual Meeting, pp.1-18
Annual Meeting of International Research Group on Wood Protection (IRG), 56th (Beijing, China, 24-May-2026–28-May-2026)
IRG Secretariate
2026
url
https://www.irg-wp.com/irgdocs/details.php?1d8aa586-0547-4155-b8dc-aad39c15ef43View
Webpage

Abstract

Timber engineering Spot test, treated pine preservative penetration computer vision machine learning
Preservative penetration checks for copper chrome arsenate treated timber in Australia are commonly performed using the manual spot tests; PAN (1-(2-pyridylazo)-2-naphthol (CAS 85-85-8)) for copper, and Variamine Blue RT salt (VBRT) ((N-phenyl-p-phenylenediamine, diaozonium salt) (CAS 4477-28-5) for sapwood in conifers) testing, followed by visual interpretation against AS/NZS 1604.1 requirements. In practice, these checks are subjective, especially when colour evidence is discontinuous, low contrast, or affected by lighting, surface texture, and timber anatomy. This can lead to inconsistent " Pass " or " Fail " outcomes between assessors, slower release decisions, and harder to defend quality assurance records. This paper presents a machine learning assisted workflow that converts paired PAN and VBRT images into repeatable measurements and traceable decision outputs aligned with H3 penetration logic for 35 mm × 70 mm timber boards. The workflow standardises image size through cropping, uses manual annotations to generate reference masks, and then derives quantitative metrics including treated sapwood coverage, detection of major surface connected untreated gaps, penetration depth from each exposed face with a 5 mm target, and the fraction of unpenetrated heartwood under an allowance check consistent with AS 1604.1 logic. These metrics serve two purposes. First, they support a rule-based compliance check using predefined thresholds. Second, they are used as input features for supervised classification. A dataset of approximately 2,000 paired images was planned to be established, with 497 specimens used for model evaluation and training at the current preliminary stage. The trained model achieved an overall accuracy of 73.2% across these 497 specimens. A separate verification test on 125 additional unseen specimens achieved 80% accuracy, demonstrating the feasibility of an AI assisted decision tool for improving consistency and transparency in treated timber penetration assessment. A full 2,000 data set will be analysed, and the algorithms will be improved. The current workflow provides a practical baseline that addresses the core consistency issue, and the next stage is focused on systematic improvement through continual learning and deployment feedback.

Details

Metrics

1 Record Views
Logo image