高光谱结合二维相关光谱检测灵武长枣中半纤维素的含量

利用高光谱成像技术与二维相关光谱（2D-COS）结合化学计量学检测灵武长枣半纤维素含量。采用定量瘀伤装置获得0，Ⅰ，Ⅱ，Ⅲ，Ⅳ级瘀伤长枣模型，通过高光谱和分光光度计分别获得样品高光谱图像和半纤维素含量。蒙特卡洛异常值检测法剔除异常样本后，分别用随机划分法（RS），Kennard-Stone法（KS）、光谱-理化值共生距离法（SPXY）和3∶1比例法对样本集划分校正预测。采用基线校准（Baseline）、去趋势（De-trending）和标准化（Normalize）对长枣原始光谱预处理后建立偏最小二乘回归模型（PLSR），优选最佳样本集划分及预处理方法。利用2D-COS将光谱信号扩展到第2维，在全光谱范围内寻找与半纤维素含量相关的敏感波段区间。采用竞争性自适应加权算法（CARS）、引导软收缩（BOSS）、区间变量迭代空间收缩方法（iVISSA）、变量组合集群分析法（VCPA）以及iVISSA+BOSS，iVISSA+CARS和iVISSA+VCPA方法在2D-COS敏感波段区间进行特征波长提取，并建立基于特征波长的PLSR模型。结果表明，样本集经3∶1划分和Baseline预处理后建立的基于全波段的PLSR模型最优，故最佳样本集划分方法为3∶1，预处理方法为Baseline，用于后续特征波长提取。通过2D-COS分析发现3个与半纤维素相关的自相关峰（401，641和752 nm）；在2D-COS敏感区域（401～752 nm范围内），采用BOSS，CARS，iVISSA，VCPA，iVISS+BOSS，iVISS+CARS，iVISS+VCPA分别提取了14，26，39，12，15，22和11个对应的特征波长，占总波长的18.9%，35.1%，52.7%，16.2%，20.2%，29.7%和14.8%。对比2D-COS和特征波建立的PLSR模型，2D-COS+iVISSA-PLSR模型效果较好，其R2_C=0.747 9，R2_P=0.604 7，RMSEC=0.043 8，RMSEP=0.060 3。研究表明，利用高光谱成像技术结合2D-COS可实现灵武长枣半纤维素含量的快速检测。

A Combination of Hyperspectral Imaging With Two-Dimensional Correlation Spectroscopy for Monitoring the Hemicellulose Content in Lingwu Long Jujube

In this paper, hemicellulose content in Lingwu long jujube was determined by hyperspectral imaging and two-dimensional correlation spectroscopy (2D-COS) combined with stoichiometry. A quantitative bruising device was used to obtain the level 0，Ⅰ，Ⅱ，Ⅲ and Ⅳbruising model of jujube. Hyperspectral images and hemicellulose content of samples were obtained by hyperspectral and spectrophotometer, respectively. After the outliers were eliminated by the Monte Carlo cross-validation method, sample sets were divided into corrected and prediction sets by random sampling (RS)，kennard-stone method (KS)，sample set partitioning based on joint X-Y distances (SPXY) and 3∶1 partitioning method, respectively. The original spectrum of long jujube was preprocessed by baseline calibration, de-trending and normalising, and then a partial least square regression model was established to determine the optimal sample set division method and spectral pretreatment method.The spectral signal was extended to the second dimension by 2D-COS, and sensitive wavelength areas related to hemicellulose content were searched in the full spectral range. Competitive adaptive reweighted sampling (CARS), bootstrapping soft shrinkage (BOSS), interval variable iterative space shrinkage approach (iVISSA), variables combination population analysis (VCPA), iVISSA+BOSS, iVISSA+CARS and iVISSA+VCPA combination methods were used to extract characteristic wavelengths in the 2D-COS sensitive wavelength areas, and establish PLSR model based on characteristic wavelengths.The results showed that the PLSR model of full band established after the sample set was divided by 3∶1 and Baseline preprocessed was optimal. Therefore, the optimal sample set division method is 3∶1, and the spectral pretreatment method is Baseline, which isused for the subsequent characteristic wavelength modeling. Three autocorrelation peaks containing 401, 641 and 752 nm were found by 2D-COS analysis, respectively. The BOSS, CARS, iVISSA, VCPA, iVISSA+BOSS, iVISSA+CARS, iVISSA+VCPA methods were applied to selected 14, 26, 39, 12, 15, 22 and 11 corresponding characteristic wavelengths from 2D-COS spectra, accounting for 18.9%, 35.1%, 52.7%, 16.2%, 20.2%, 29.7%, 14.8% of the total wavelength, respectively. Comparedwith the PLSR model established by 2D-COS and characteristic waves, the 2D-COS+iVISSA-PLSR model had the best performance, with R2_C=0.747 9, R2_P=0.604 7, RMSEC=0.043 8, RMSEP=0.060 3. The results showed that hyperspectral imaging technology combined with 2D-COS could be used to detect hemicellulose content in Lingwu long jujube quickly.