Machine learning based density estimation of light red meranti (Shorea spp.): a segmented approach to multiple regression of self-organising maps colour clusters using custom made ‘KayuSort’ colour sorting software

Abstract

Wood density is an important characteristic of wood which correlates to its strength. This study proposes an algorithm using multiple regression on pre-segmented colour images of the wood to estimate the density of light red meranti (Shorea spp.) (LRM). Two batches of LRM timber were randomly selected from a factory (Batch 1: 119 samples, Batch 2: 79 samples). Timber samples were kiln-dried, free of sapwood and major visual defects, and freshly surfaced 2 sides. The apparent density and moisture content (MC) of each timber sample were measured. The samples were then imaged and colour-sorted using KayuSort, an in-house industrial timber colour sorting prototype that uses the self-organising map (SOM) algorithm. Otsu thresholding was applied to several different colour space components to obtain features. Multiple regression was applied to obtain an equation to estimate the density of the wood. Coefficients of determination (\(\hbox {R}^{2}\)) and 95\(\%\) Limits of Agreement (LoA) were used to assess performance. Performing colour segmentation to the dataset using KayuSort for average \(\text {YC}_b{\text{C}}_r\) colour space scored an \(\hbox {R}^{2}\) of 0.7109 and an LoA of ±146.8 \(\hbox {kgm}^{-3}\). Therefore, it is possible to estimate the density of LRM using colour features of the wood using KayuSort, with the caveat that timber is kiln-dried to under 15% MC, freshly surfaced, without major defects and sapwood, and within the thickness range of 26.9 to 30.6 mm.