NIR-based Modeling Wood Density Of Major Hardwood Species In Northeastern Region

Wood density is an important property for wood quality evaluation. Rapid determination of wood density is essential for marketed-oriented stand breeding and wood utilization. In this paper, the probe light was used for near infrared spectroscopy (NIRS) collection from wood discs’surface. Partial least squares (PLS)_ modeling technique was used for model development and density prediction of unknown samples. Near infrared spectrum of the wood was pretreated with wavelet denoising and wavelet compression for model optimization. The impacts of parameter selection on the model performance were further discussed.(1) Elm wood density NIRS calibration models were developed with PLS modeling technique. The determination coefficient (R2) was 0.8347 and 0.7461 for the calibration and validation models, respectively. It was 0.8270 and 0.7570 for red oak and 0.8475 and 0.7634 for ribbed birch. For hybrid model, R2 was 0.8820 and 0.7614 for the calibration and validation models, respectively. Applied the model to forecast the wood density of unknown samples, he best predicting results were associated with the ribbed birch calibration model with R’of 0.8390. This study indicated that NIRS can be used for wood density determination in the field.(2) Wavelet transform were applied to the original wood NIRS for the noise elimination. With sym5 and wavelet decomposition layers of three, heuristic & hard threshold was associated with the best denoising effects for elm wood. With db5 and wavelet decomposition layers of two, sqtwolog & hard threshold was associated with the best denoising effects for oak wood. With db5 and wavelet decomposition layers of two, heuristic & hard threshold was associated with the best denoising effects of ribbed birch wood. PLS models were developed based on the denoised NIRS. For elm wood, R2 was 0.8655 and 0.7925 for the calibration and validation model, respectively. It was 0.8918 and 0.8039 for oak wood and 0.8671 and 0.7824 for ribbed birch. For the hybrid model, R2 was 0.8820 and 0.8130 for the calibration and validation model, respectively. The best predicting results were associated with the ribbed birch wood NIR calibration model developed with the denoised NIR data with R2 of 0.8923, RMSEP of 0.0276, and SEP of 0.03, respectively. The study indicates that wavelet transform can eliminate the noise of NIRS and improve the model accuracy.(3) The NIRS data of wood were compressed with wavelet transform algorithm in this study. Balanced sparse standard hard threshold was shown to be the best wavelet compression method for the NIRS in this study. The best model performance was db2 with wavelet decomposition layers of six for red oak wood, sym4 with wavelet decomposition layers of seven for elm wood, and db2 with wavelet decomposition layers of six for ribbed birch. Wood density prediction models were developed based on PLS and compressed NIR. The R2 of calibration model was 0.8355,0.8545, and 0.8479, for elm wood, oak wood and ribbed birch wood, respectively. For hybrid model, R2 of the calibration was 0.8469. The best predicting results were associated with the oak wood density calibration model developed with the compressed NIR data with R2 of 0.8717. This study indicated that the wavelet compression method could effectively simplify NIRS data and improve the prediction accuracy.