Partial least square regression (PLS-R) calibrations based on near infrared (NIR) spectroscopic data were developed in order to predict mechanical and physical properties of agro-based particleboards. The panels were manufactured using Eucalyptus and Pinus wood particles and sugar cane bagasse. The following panel properties were evaluated according to standard methods: modulus of elasticity (MOE), modulus of rupture (MOR), internal bonding (IB) strength, water absorption (WA24H), and thickness swelling (TS24H) after 24 hours of immersion. NIR spectra information was measured on samples cut from each particleboard and correlated with their physical and mechanical properties by PLS-R to build predictive NIR models. The NIR models for IB, WA24H and TS24H presented satisfactory coefficient of determination (0.73; 0.72 and 0.75, respectively.) The key role of resins (adhesives), cellulose, and lignin for NIRS calibrations of mechanical and physical properties of the particleboards is shown. These models can be useful to quickly verify such properties in unknown agro-based particleboards.

Particleboards can be manufactured from particles of any lignin-cellulosic material that can be combined with an adhesive and consolidated under the action of temperature and pressure. Because the raw materials in the industrial process are continually changing, the particleboard industry requires methods for monitoring the quality of their products. Hence, the aim of this paper was to evaluate the composition of the agro-based particleboards by near infrared spectroscopy. In this study, agro-based particleboards produced with different compositions of Eucalyptus and Pinus wood particles and sugar cane bagasse were evaluated by NIR spectroscopy and partial least square (PLS) regression. The PLS models to estimate the Eucalyptus and Pinus particles and sugar cane bagasse contents presented a strong coefficient of determination (0.90, 0.88 and 0.84, respectively), but also high magnitudes of standard errors of cross-validation were observed (ranging from 8.84 to 11.27%). Development work would be required in order to reduce the standard errors and improve predictive model performance to build robust models that could be applied as quality control tool.

No abstract