The NIRS technology, together with multivariate methods, has already proved to be capable of discriminating wood from different forest species that are visually similar and determining the provenance within the same species. However, models developed with dry wood under controlled conditions of humidity and temperature present prediction errors or high outlier exclusion rates when applied to the analysis of samples under different moisture conditions. This work proposes applying external parameter orthogonalization (EPO) to minimize the effect of moisture in the spectra of wood from native Brazilian species under different moisture conditions and enable its analysis by PLS-DA discrimination models. After correction, the models showed high-efficiency rates and a significant reduction in the number of outliers. For the model trained with oven-dried samples and validated with spectra measured on samples under environment conditions, it was observed that the rate of samples with inconclusive results (RIR) was reduced from 87.8% to 18.4% by using EPO. Under these conditions, efficiency rates of 93.5% were obtained for the identification of Carapa guianensis Aubl. (Andiroba), 98.6% for Swietenia macrophylla King (Mogno) and 100.0% for Cedrela odorata L. (Cedro), Erisma uncinatum Warm. (Cedrinho) and Micropholis melinoniana Pierre (Curupixá). For the second validation set, with samples in more extreme moisture conditions, the same model without EPO correction had a RIR of 100.0%. In comparison the model with EPO had a RIR of 43.8%, demonstrating the feasibility of EPO in the correction of moisture interference in samples under moisture conditions different from those used in the model development. Applying more flexible criteria in the identification of samples by PLS-DA also favored the reduction of errors. Thus, EPO and soft-PLS-DA proved to be an effective strategy to enable the application of NIR technology in field conditions, such as in the inspection of illegal timber trade.
Read full abstract