Data-driven assessment of log characteristics and processing factors associated with veneer volume recovery and grades in industrial veneer production : a South African case study of Pinus patula and Pinus elliottii

Industrial veneer production involves highly variable roundwood, processing conditions, and resulting veneer defects, but the combined effects of these variables are difficult to quantify under mill conditions. This study developed a data-driven assessment of industrial veneer production using Pinus patula and Pinus elliottii roundwood from a South African mill as on-site industrial research. A total of 988 sampled veneer logs were processed. Three analytical levels were examined. At the group-level, 12 combinations of wood species, freshness, and log size were compared on volume recovery and the proportion of high-grade veneer yielded. At the log-level, multiple linear regression, generalized least squares (GLS), and random forest (RF) models were developed to explain and predict relative green veneer recovery (RGVR). At the veneer-level, logistic regression (LR) and RF models were implemented to examine and predict operational veneer-grade classes. The group-level analysis showed that no group maximized both volume recovery and high-grade veneer proportion, hence processing priority should be dependent on the specific production target requirements. Fresher or larger logs generally yielded better volume recovery, whereas older or smaller logs tended to produce a higher total proportion of A-, B-, and C-grade veneer. At the log-level, freshness, log size, log geometry, and post-conditioning log temperature measured at the middle portion of the log, were among the most useful variables associated with RGVR. GLS models provided the preferred explanatory framework by accounting for heteroscedasticity, while the three modelling approaches showed similar internal cross-validated predictive performance, with R² values of approximately 0.37~0.40. At the veneer-level, sheet length, drying conditions, freshness, and visual defects contributed to grade classification. The three-level RF model predicted “Layup” and “Not accept” veneer more effectively than “Face” veneer, while the LR model provided more interpretable associations. Overall, this research shows the value of integrating multilevel industrial data to support veneer production decisions. Routinely collected industrial data can support log processing prioritization, veneer volume recovery analysis, and veneer grade prediction, but incomplete measurement and traceability constrained model performance. Improving log-to-veneer linkage, moisture content measurement, conditioning records, and defect characterization is likely to provide greater gains than increasing model complexity alone.