Identification Of Dalbergia Spp. Wood Based On Hyperspectral Imaging Technology

It is contribute to standardize the Dalbergia spp. market order and protect the interests of consumers if common Dalbergia spp. of wood on the market could be identified rapidly and nondestructively. The current Dalbergia spp. recognition methods have lots of shortcomings, such as time-consuming, labor-intensive, high-cost and could not identify wood nondestructively. This research as to identity four kinds of Dalbergia spp.(D. cochinchinensis, D. bariensis, D. oliveri,and D. retusa) by combining hyperspectral imaging technology and different discriminant models.At first, the region of interests were extracted from the Dalbergia spp. hyperspectral images that wavelengths from 400 to 1000 nm and wavelengths from 900 to 1700 nm. Then, the wavelengths from the region of interests were preprocessed by Savitsky-Golay smoothing(SG),standard normal variate(SNV), and Multiplicative Scatter Correction(MSC), principal component analysis(PCA), regression coefficient(RC), speccial wavelengths and successive projections algorithm(SPA) were used to selecte sensitive wavelength. Moreover, the textural features in monochrome images and PC images were extracted from the Dalbergia spp. hyperspectral images that wavelengths from 400 to 1000 nm. At last, a partial least square-discriminant analysis(PLS-DA) and an extreme learning machine(ELM) were used to build discriminant models based on full wavelengths, selected sensitive wavelengths and textural features. By comparing the effects of different models between two kinds of spectral ranges to find the best model.The results showed that, among the 144 samples included the calibration set and the prediction set, the classification accuracy of models build based on textural features not very well,the classification accuracy for both the calibration set and the prediction set were below 85%. The models build based on the textural features in monochrome images is better than the textural features in PC images. The models build based on spectral data that wavelengths from 400 to 1000 nm is better than wavelengths from 900 to 1700 nm. Selected sensitive wavelengths using SPA from preprocessed spectral, ELM models obtained the best classification accuracy for both the calibration set and the prediction set were more than 97.92%. Thus, this study provided a new method to identify Dalbergia spp. wood rapidly and nondestructively.