光谱学与光谱分析, 2019, 39 (3): 705, 网络出版: 2019-03-19  

木材种类的近红外光谱和模式识别

Identification of Wood Species Based on Near Infrared Spectroscopy and Pattern Recognition Method
作者单位
1 华东交通大学机电与车辆工程学院, 江西 南昌 330013
2 赣州出入境检验检疫局, 江西 赣州 341001
摘要
木材的种类识别是木材加工和贸易的一个重要环节, 传统的木材种类识别方法主要有显微检测法和木材纹理识别法, 其操作繁琐, 耗时长, 成本高, 不能满足当前需求。 本研究利用木材的近红外光谱(NIRS)结合模式识别方法, 以期实现木材种类的快速准确识别。 采用近红外光谱结合主成分分析法(PCA)、 偏最小二乘判别分析法(PLSDA)和簇类独立软模式法(SIMCA)三种模式识别对58种木材进行种类鉴别研究; 5点平滑、 标准正态变量变换(SNV)、 多元散射校正(MSC)、 Savitzky-Golay一阶导数(SG 1st-Der)和小波导数(WD)五种光谱预处理方法用于木材光谱的预处理; 校正集和测试集样品的正确识别率(CRR)用于模型的评价。 采用PCA方法, 通过样品的前三个主成分空间分布图分辨木材种类的聚类情况。 在建立PLSDA模型, 原始光谱的正确识别率最高, 分别为88.2%和88.2%; 5点平滑处理的光谱校正集和测试集的CRR分别为88.1%和88.2%; SNV处理的光谱校正集和测试集的CRR分别为84.4%和84.5%; MSC处理的光谱校正集和测试集的CRR分别为83.1%和84.2%; SG 1st-Der处理的光谱校正集和测试集的CRR分别为81.8%和82.7%; WD(小波基为“Haar”, 分解尺度为80)处理的光谱校正集和测试集的CRR分别为87.3%和87.2%。 可知, 在PLSDA模型中, 木材光谱未经预处理种类识别效果最后好。 在建立SIMCA模型过程中, 原始光谱的校正集和测试集的CRR分别为99.7%和99.4%; 5点平滑处理的光谱校正集和测试集的CRR分别为100%和100%; SNV处理的光谱校正集和测试集的CRR分别为99.5%和99.1%; MSC处理的光谱校正集和测试集的CRR分别为99.0%和98.4%; SG 1st-Der的光谱校正集和测试集的CRR分别为81.8%和82.7%; WD处理的光谱校正集和测试集的CRR分别为100%和100%。 可知, 在SIMCA模型中, 木材光谱经平滑和小波导数处理后的识别效果最好, 且光谱的校正集和测试集CRR都为100%。 采用三种模式结合五种不同的预处理方法对木材近红外光谱进行定性建模识别时, 由于木材样本属性复杂, 主成分分布图相互交织, PCA无法识别出58种木材; 原始光谱的PLSDA模型可以得到较好的判别模型, 但校正集和测试集的CRR只有88.2%和88.2%; 木材光谱经过5点平滑或WD预处理后的SIMCA模型可达到最好的识别效果, 校正集和测试集的CRR均为100%, 且WD-SIMCA模型因子数比5点平滑SIMCA模型小, 模型更为简化, 故WD-SIMCA为58种木材种类识别的最优模型。 研究表明光谱预处理方法可以有效的提高木材种类识别精度, 有监督模式识别方法SIMCA可以用来建立有效的木材识别模型, 近红外光谱结合模式识别可以为木材种类的识别提供一种快速简便的分析方法。
Abstract
Identification of wood species is an important part of wood processing and commerce. The traditional methods of wood species identification mainly include microscopic detection and wood texture recognition which are complex, time-consuming and costly. They cannot meet the current needs. Near infrared spectroscopy (NIRS) of wood combined with pattern recognition methods were used to identify wood species. NIRS combined with three kinds of pattern recognition methods including principal component analysis (PCA), partial least squares discriminant analysis (PLSDA) and soft independent modeling of class analogy (SIMCA) were used to identify fifty-eight wood species. Five spectral preprocessing methods including 5 point smoothing, standard normal variable (SNV), multiplicative scatter correction (MSC), Savitzky-Golay first derivative (SG 1st-Der) and wavelet derivative (WD) were used to spectral transform. The correct recognition rate (CRR) of calibration and test sets were used for evaluation index of models. The results showed that the wood species could not be identified by using the first three principal components. In PLSDA model, the CRR values of calibration and test sets for original spectra model were the highest, which were 88.2% and 88.2%, respectively. The CRR values of calibration and test sets for 5 points smoothing model were 88.1% and 88.2%. The CRR values of calibration and test sets for SNV model were 84.4% and 84.5%. The CRR value of calibration and test sets for MSC model were 83.1% and 84.2%. The CRR values of calibration and test sets for SG 1st-Der model were 81.8% and 82.7%. The CRR values of calibration and test sets for WD (the wavelet basis is “Haar” and the decoposition scale is 80) model were 87.3% and 87.2%. In PLSDA models, the original spectra model had the best results compared to others. In SIMCA model, the CRR values of calibration and test sets for original spectra were 99.7% and 99.4%. The CRR values of calibration and test sets for 5 points smoothing were 100% and 100%. The CRR values of calibration and test sets for SNV model were 99.5% and 99.1%. The CRR values of calibration and test sets for MSC model were 99.0% and 98.4%. The CRR values of calibration and test sets for SG 1st-Der model were 98.4% and 99.0%. The CRR values of calibration and test sets for WD model were 100% and 100%. Compered to others spectra processed by 5 points smooting and WD had a best results in SIMCA models, the CRR values of calibration and test sets were 100%. Three kinds of pattern recognition methods combined with five spectral preprocessing methods were used to classify 58 kinds of wood. It could be concluded that the PCA method can’t explicitly classify 58 wood species because of complex properties of wood leading to the scatters of each wood species interwined with each other in PCA distribution diagram. The PLSDA model of original spectra could get a better result with the CRR value of 88.2% and 88.2% for calibration and test sets, respectively. The best SIMCA models were constructed by 5 point smoothing or WD preprocessing methods with the CRR of 100% for calibration and test sets. However, the factor of the WD-SIMCA model was smaller than 5 point smoothing method, and the model was more parsimonious, so WD-SIMCA model was an optimal model. The paper showed that spectral preprocessing methods can improve the accuracy of identificationof wood species, and SIMCA supervised pattern recognition method can be used to build effective identifying model and NIR combined with pattern recognition method can provide a rapid and simple method for identification of wood species.

郝勇, 商庆园, 饶敏, 胡远. 木材种类的近红外光谱和模式识别[J]. 光谱学与光谱分析, 2019, 39(3): 705. HAO Yong, SHANG Qing-yuan, RAO Min, HU Yuan. Identification of Wood Species Based on Near Infrared Spectroscopy and Pattern Recognition Method[J]. Spectroscopy and Spectral Analysis, 2019, 39(3): 705.

关于本站 Cookie 的使用提示

中国光学期刊网使用基于 cookie 的技术来更好地为您提供各项服务,点击此处了解我们的隐私策略。 如您需继续使用本网站,请您授权我们使用本地 cookie 来保存部分信息。
全站搜索
您最值得信赖的光电行业旗舰网络服务平台!