Assessment of machine learning algorithm-based grading of Populus x euramericana I-214 structural sawn timber

The efficiency of visual grading standards applied to structural timber is often inappropriate, and timber properties are either under or over-graded. Although not included in the current UNE 56544 visual grading standard, machine learning algorithms represent a promising alternative to grade structural timber. The general aim of this research was to compare the performance of machine learning algorithms based on visual defects, non-destructive techniques and sawing systems ( cut type ) with UNE 56544:1997 visual grading in order to predict the qualifying efficiency of Populus x euramericana I-214 structural timber. Visual evaluation, ultrasound and vibrational non-destructive testing, and sawing systems register (radial, tangential and mixed) were applied to characterize 945 beams. In addition, in order to retrieve actual physical-mechanical values, density and static bending destructive testing (EN-408:2011 + A1:2012) was also carried out. Several machine learning algorithms were then used to grade the beams, and their predictive accuracy was compared with that of visual grading. To do so, three scenarios were considered: a first scenario in which only visual variables were used; a second scenario in which cut type variables were also included; and a third scenario in which additional nondestructive variables were considered. Results showed a poor level of performance of UNE 56544:1997, with an apparent mismatch between the strength values assigned for each visual grade (established by the EN 338 standard) and the actual values. On the opposite, all algorithms performed better than visual grading and may thus be deemed as promising timber strength grading tools.

Automated summary

The efficiency of visual grading standards for structural timber is often inadequate, and machine learning algorithms could be a promising alternative solution. This study compared the performance of machine learning algorithms with visual grading in predicting the qualifying efficiency of structural timber. The results showed that machine learning algorithms outperformed visual grading and could be considered as promising tools for timber strength grading.