College of Mechanical and Electrical Engineering, Shihezi University, Shihezi 832003, China.
College of Mechanical and Electrical Engineering, Shihezi University, Shihezi 832003, China; Key Laboratory of Northwest Agricultural Equipment, Ministry of Agriculture and Rural Affairs, Shihezi 832003, China.
Food Res Int. 2024 Nov;196:115010. doi: 10.1016/j.foodres.2024.115010. Epub 2024 Sep 4.
Pesticide residues are identified as a significant food safety issue in Hami melons, and their rapid and accurate detection is deemed critical for ensuring public health. In response to the cumbersome procedures with existing chemical detection methods, this study explored the potential of identifying different pesticide residues in Hami melon by microfluorescence hyperspectral imaging (MF-HSI) technology combined with machine learning. By simulating the actual agricultural production, three pesticides, Beta-Cypermethrin, Difenoconazole, and Acetamiprid, were sprayed on Hami melons. The Hami melons with pesticide residues were collected as samples, from which spectral and image information were extracted. Competitive adaptive reweighted sampling (CARS), genetic algorithm (GA), and sequential projection algorithm (SPA) methods were used to extract characteristic wavelengths. Partial least squares discriminant analysis (PLS-DA), extreme learning machine (ELM), and k-nearest neighbor (KNN) classification models of the original spectra and characteristic wavelengths were established. 9 feature variables-Red (R), Green (G), Blue (B), Hue (H), Saturation (S), Value (V), Lightness (L), Red-Green axis (a) and Yellow-Blue axis (b) information was extracted by the color statistical method. 4 important color information of the image (G, B, V, L) were identified through Pearson correlation analysis, and the optimal feature wavelength was fused to enhance the identification accuracy of models. The results indicated that the SPA-PLS-DA model demonstrated the highest accuracy for the characteristic wavelength dataset, achieving accuracies of 89.35 % for the training set and 86.99 % for the testing set, which was better than the model established by the full wavelength dataset. Using the dataset that fuses feature wavelengths with 4 important image features, the SPA-PLS-DA model demonstrated superior performance, with recorded accuracy, precision, specificity, and sensitivity metrics on the testing set at 93.48 %, 93.81 %, 96.63 %, and 93.36 %, respectively. Consequently, MF-HSI technology combined with machine learning offers an approach to analyze pesticide residues in Hami melons accurately, and it provides a technical basis for detecting pesticide residues in other fruits and vegetables.
农药残留被认为是哈密瓜食品安全的一个重大问题，快速准确地检测农药残留对于保障公众健康至关重要。针对现有化学检测方法繁琐的问题，本研究探索了利用微荧光高光谱成像（MF-HSI）技术结合机器学习识别哈密瓜中不同农药残留的可能性。通过模拟实际农业生产，在哈密瓜上喷洒了三种农药，分别为β-氯氰菊酯、三唑酮和啶虫脒。采集有农药残留的哈密瓜作为样本，提取其光谱和图像信息。采用竞争自适应重加权抽样（CARS）、遗传算法（GA）和序贯投影算法（SPA）方法提取特征波长。建立原始光谱和特征波长的偏最小二乘判别分析（PLS-DA）、极限学习机（ELM）和 K 最近邻（KNN）分类模型。通过颜色统计方法提取 9 个特征变量-红（R）、绿（G）、蓝（B）、色调（H）、饱和度（S）、值（V）、亮度（L）、红-绿轴（a）和黄-蓝轴（b）信息。通过皮尔逊相关分析确定图像的 4 个重要颜色信息（G、B、V、L），融合最优特征波长以提高模型的识别精度。结果表明，SPA-PLS-DA 模型在特征波长数据集上表现出最高的准确性，在训练集和测试集上的准确率分别为 89.35%和 86.99%，优于全波长数据集建立的模型。使用融合特征波长和 4 个重要图像特征的数据集，SPA-PLS-DA 模型表现出优异的性能，在测试集上的准确率、精度、特异性和敏感性分别为 93.48%、93.81%、96.63%和 93.36%。因此，MF-HSI 技术结合机器学习为准确分析哈密瓜中的农药残留提供了一种方法，为检测其他水果和蔬菜中的农药残留提供了技术基础。
