为了实现库尔勒香梨依据可溶性固体含量(SSC)分级定等和按质论价, 推动采后处理向标准化、 产业化方向健康发展, 利用高光谱成像技术研究出了一种快速、 有效、 无损检测库尔勒香梨SSC的方法。 以表面无损伤的157个库尔勒香梨作为研究样本, 应用高光谱成像采集系统获取400~1 000 nm波长范围内高光谱图像并用ENVI5.3软件提取感兴趣区域(ROI), 获得高光谱数据。 采用Kennard-Stone(KS)样本集划分方法将全部样本按照2∶1的比例划分为校正集(105)和预测集(52)。 对比标准变量变换(SNV)、 多元散射校正(MSC)、 一阶导数(FD)和二阶导数(SD)等数据预处理方法对建模精度的影响, 最终选用SNV方法对光谱曲线进行平滑去噪。 该研究提出竞争性自适应重加权算法与平均影响值算法的组合算法(CARS-MIV)选择特征波长。 在竞争性自适应重加权算法(CARS)方法中, 建模样本由蒙特卡罗算法随机选择生成, 变量回归系数会随之发生变化, 因而回归系数的绝对值不能全面反映变量重要性, 从而影响模型检测精度。 为降低这种影响, 应用平均影响值(MIV)算法对选出的自变量进行二次筛选, 筛选出相关性较大的变量用以建模分析, 并与CARS、 连续投影算法(SPA)、 蒙特卡罗无信息变量消除算法(MCUVE)等经典特征波长选择算法进行比较。 最后分别以全波长(FS)光谱信息和四种特征波长选择方法得出的光谱信息作为输入矢量, 应用支持向量回归(SVR)建立库尔勒香梨可溶性固体含量定量预测数学模型, 以校正集相关系数(Rc)、 校正集均方根误差(RMSEC)、 预测集相关系数(Rp)和预测集均方根误差(RMSEP)四个参数来评估模型的预测精度。 比较分析发现, CARS-MIV-SVR模型效果最佳, 校正集相关系数(Rc)为0.985 94, 预测集相关系数(Rp)达到0.946 31, 校正集和预测集均方根误差分别为0.185 85和0.403 33。 结果证明: CARS-MIV特征波长选择方法能够有效增强库尔勒香梨光谱数据特征波长选择的稳定性和精确性, 提高模型的预测精度。 利用高光谱技术结合CARS-MIV-SVR模型能够满足库尔勒香梨可溶性固体含量测定需求, 实现库尔勒香梨的分级定等和按质论价。

In order to classify and set different prices on the basis of soluble solid content (SSC) of korla pears and promote the development of post-harvest processing healthily in standardization and industrialization, a fast, precise and nondestructive method to detect soluble solid content of korla pears was determined by applying hyperspectral reflectance imaging technology. 157 korla pears freshly and with no surface damage were collected as samples. Hyperspectral images with a spectral range of 400~1000 nm of pears were acquired by hyperspectral imaging system. Then the region of interest (ROI) function of ENVI 5.3 software was used to conduct spectral data extraction from each hyperspectral image of pear. Totally, 157 pear samples were divided into calibration set (105) and prediction set (52) based on the Kennard-Stone(KS)sample set partitioning method. The research compared the influence of accuracy of modeling in terms of the spectrum pretreatment methods of original spectrum, standard normal variate (SNV), multiplicative scatter correction (MSC), first derivative (FD) and second derivative (SD). The SNV was applied for smoothing and denoising of the original hyperspectral data. A variable selection method combining competitive adaptive reweighted sampling and mean impact value (CARS-MIV) was utilized to extract the characteristic variables from full spectrum (FS). The modeled samples of competitive adaptive reweighted sampling (CARS) are generated by random selection of Monte Carlo sampling, and the regression coefficients of variables will change accordingly. The absolute value of regression coefficients cannot fully reflect the importance of variables, and affect the accuracy of the model. To lower the impact, the mean impact value (MIV) algorithm is applied to select the independent variables for secondary screening, and the variables with bigger correlation are selected for modeling and analysis. In this paper, the variables selected by CARS, successive projection algorithm (SPA) and Monte-Carlo uninformative variable elimination (MCUVE) were used for comparison. Finally, the spectral information selected from full wavelength and the spectral information selected from four characteristic wavelength selection method were taken as input vector to build support vector regression(SVR)model to predict soluble solid content of korla pears. The performances of the models were evaluated by the root of mean square of calibration (RMSEC), the root of mean square of prediction (RMSEP), the correlation coefficient of calibration (Rc) and the correlation coefficient of prediction (Rp). By means of comparison, the CARS-MIV-SVR models achieved the optimal performance with the Rc reaching 0.985 94 and Rp up to 0.946 31. The RMSEC and RMSEP are 0.185 85 and 0.403 33 respectively. These experimental results demonstrated that CSRS-MIV method can efficiently improve the stability and accuracy of wavelength selection, and optimize the precision of prediction model. The hyperspectral technique combined with CARS-MIV-SVR model can meet the needs of determination of soluble solid content and be used to classify and set different prices on the basis of SSC of korla pears.