采用荧光高光谱成像技术对脐橙表面不同浓度毒死蜱和多菌灵进行判别。 实验通过由氙灯光源激发的高光谱成像系统（392～998.2 nm）分别采集浓度为0, 0.5, 1, 2 mg·kg-1的毒死蜱和0, 1, 3, 5 mg·kg-1多菌灵的高光谱图像。 使用ENVI软件获取样本的感兴趣区域（ROI）; 对原始光谱数据采用卷积平滑（SG）、 标准正态标量变换（SNV）及一阶导数（FD）方法进行预处理; 采用区间变量迭代空间收缩法（iVISSA）、 无信息变量消除算法（UVE）和竞争性自适应加权算法（CARS）进行一次提取特征波长, 二维相关光谱（2D-COS）方法进行二次提取特征波长。 最后采用主成分分析与线性判别分析相结合算法（PCA-LDA）和偏最小二乘算法（PLS-DA）建立基于两次提取特征波长脐橙表面不同浓度毒死蜱和多菌灵残留的判别模型。 将原始光谱数据与经过预处理的3种光谱数据进行建模分析, 结果发现毒死蜱和多菌灵的光谱数据经过SG处理后模型效果最优。 对经SG预处理后的毒死蜱光谱数据和多菌灵光谱数据进行特征波长一次提取, 最佳特征波长分别为iVISSA法和CARS法, 分别提取出26个和30个特征波长; 再采用二维相关光谱（2D-COS）算法对这26个和30个特征波长进行二次提取, 分别得到10个和12个特征波长。 对一次提取特征波长和二次提取特征波长后的光谱数据分别建模。 结果表明, 对于不同浓度的毒死蜱, 基于iVISSA-2D-COS建立的PCA-LDA模型判别效果最佳, 其校正集与预测集判别正确率分别为98.61%和95.83%; 对于不同浓度的多菌灵, 基于CARS-2D-COS建立的PCA-LDA模型判别效果最佳, 其校正集与预测集判别正确率分别为97.22%和95.83%, 均高于全波段光谱数据模型和一次提取特征波长模型判别正确率, 说明2D-COS可以捕捉可用的荧光光谱信息。 该研究采用2D-COS对一次提取最优特征波长进行二次提取后建模, 研究结果为脐橙表面不同浓度农药残留的快速无损判别提供了一定的参考。

In this study, fluorescence hyperspectral imaging technology identified different concentrations of chlorpyrifos and carbendazim on the surface of navel oranges. Hyperspectral images of the concentrations of chlorpyrifos at 0, 0.5, 1 and 2 mg·kg-1 and carbendazim at 0, 1, 3 and 5 mg·kg-1 were acquired by a hyperspectral imaging system (392~998.2 nm) excited by a xenon light source. The sample’s region of interest (ROI) was captured by ENVI software. Raw spectral data were pre-processed by a spectral pre-processing methods, including SG, SNV and FD. The interval variable iterative spatial shrinkage (iVISSA), uninformative variable elimination algorithm (UVE) and competitive adaptive reweighted sampling (CARS) were used for the primary extraction of feature wavelengths and the two-dimensional correlation spectroscopy (2D-COS) method for secondary extraction of feature wavelengths. PLS-DA and PCA-LDA model developed primaryand secondary feature wavelength extraction at different concentrations of chlorpyrifos and carbendazim residues on the surface of navel oranges. 3 methods studied the spectral pretreatment. The results showed that the model effect of SG methods was best. A total of 26 feature wavelengths were extracted by the iVISSA method for the spectral data using the SG chlorpyrifos; A total of 30 feature wavelengths were extracted by the CARS method for the spectral data using the SG method of carbendazim. The 2D-COS algorithm was used for the secondary extraction of 26 and 30 feature wavelengths, resulting in 10 and 12 feature wavelengths, respectively. Discriminant models based on spectral data of primary and secondary extraction of feature wavelengths were established to identify the samples. The results showed that the PCA-LDA model based on iVISSA-2D-COS was the best with the calibration set and prediction set discrimination rates of 98.61% and 95.83% for different concentrations of chlorpyrifos. The PCA-LDA model based on CARS-2D-COS was the best with the calibration set and prediction set discrimination rates of 97.22% and 95.83% for different concentrations of carbendazim, respectively, which were higher than the discrimination rates of full-band spectral data and once-extraction feature spectral data. In this study, secondary extraction of the optimal feature wavelengths by 2D-COS has developed discrimination models, and the results can provide some reference for rapid and non-destructive discrimination for different concentrations of pesticide residues on the surface of navel oranges.