Pretreatment of near-infrared spectroscopy of wood based on wavelet compression

Due to the multi-dimension of near infrared spectrum (NIRS) and large volume of data, huge storage space is needed for data processing, which directly affects the speed and accuracy of online data analysis. This study aims to discuss the feasibility of pretreatment of near-infrared spectroscopy of wood based on wavelet compression as well as its effect on prediction accuracy of Betula costata Trautv wood density using NIR technology. The NIRS data of B. costata wood were compressed using wavelet transform algorithm with the aid of Matlab. Results showed that the global threshold value based on balance sparsity norm and the heuristic threshold value were observed to be the best with decomposition layer of 6 for the sym2 wavelet. With the method, the 2 151 variables were compressed into 38 wavelet coefficients, and the corresponding energy reserved component, zero coefficient component and compression ratio were 99.66%, 98.34% and 56.61%, respectively. The partial least squares (PLS) models were developed based on both the original NIRS and the 38 wavelet coefficients after compression. The inner cross validation was used and the external validation was applied to both the original and the compressed dataset. The best prediction results were associated with the calibration model developed with the compressed NIR data with determination coefficient (R2) of 0.913 9. This study indicates that the wavelet compression method could effectively simplify NIRS data and improve the prediction accuracy.