Comparison of genetic algorithm optimized and PLS regression density models for Acacia auriculiformis

Abstract

Partial least squares (PLS) regression models based on genetic algorithm (GA) representative near infrared (NIR) wavelengths for estimating wood properties provide improved calibration and prediction statistics compared to PLS models based on all available NIR wavelengths. However, the utilization of predicted data, obtained from full NIR wavelength and GA selected NIR wavelength models, in a practical application, has not been explored. Our application was to examine radial density variation in Acacia auriculiformis Cunn. Ex Benth. clones at a resolution of 10 mm. One hundred and forty A. auriculiformis samples representing seven clones and two radial positions (adjacent to pith and bark respectively) had NIR hyperspectral images (wavelength range 931-1718 nm) collected from their transverse surface. Two PLS density models (all NIR wavelengths and GA representative NIR wavelengths) were developed using 134 NIR spectra extracted from the images. The models were then used to predict density in 10 mm increments of 144 radial samples from the same clones. A PLS density model using only 15 representative NIR wavelengths provided a mean (0.506 g/cm³) which was not statistically significantly different to measured density (0.503 g/cm³), whereas the mean for PLS model using all wavelengths was 0.522 g/cm³. However, the PLS model with 15 representative NIR wavelengths had greater variation (standard deviation of 0.060) compared with measured data (0.052) and the full NIR wavelengths PLS model (0.047). Radial density variation was less than 0.09 g/cm³ for six of the seven clones examined.