基于高光谱和卷积神经网络的西兰花农药残留检测  

Pesticide Residue Detection in Broccoli Based on Hyperspectral Technology and Convolutional Neural Network

作　　者：王丹 栾雨晴 谭佐军 魏薇 WANG Dan;LUAN Yuqing;TAN Zuojun;WEI Wei(College of Informatics,Huazhong Agricultural University,Wuhan 430070,China;College of Engineering,Huazhong Agricultural University,Wuhan 430070,China)

机构地区：[1]华中农业大学信息学院,湖北武汉430070 [2]华中农业大学工学院,湖北武汉430070

出　　处：《食品工业科技》2025年第6期1-8,共8页Science and Technology of Food Industry

基　　金：国家自然科学基金(No.42271357)。

摘　　要：农产品农药残留检测是保证农产品食用安全的重要环节,而传统检测方法步骤繁琐、成本高昂。本文利用高光谱技术结合机器学习算法和深度学习算法,以西兰花农药残留检测为样本,提供了一种简便快速、成本低、无损的西兰花农药残留检测方法。研究通过采集喷洒了不同种类农药和清水的西兰花样本400~1000 nm高光谱图像,经过多元散射校正(MSC)、Savitzky-Golay卷积平滑(SG平滑)两种数据预处理方法,和主成分分析法(PCA)、竞争性自适应重加权算法(CARS)、连续投影算法(SPA)三种数据降维后,建立支持向量机(SVM)识别模型进行农药残留判别。得到SVM-SG-SPA组合判别效果最好,其对高效氯氰菊酯、毒死蜱、吡虫啉和清水的识别精度分别达到92.86%、94.29%、91.43%和92.86%。用原始光谱数据建立一维卷积神经网络(1D-CNN)模型,其对高效氯氰菊酯、毒死蜱、吡虫啉和清水的识别精度达到94.29%、95.71%、94.29%和97.14%,识别精度均高于SVM模型。结果表明,高光谱成像技术结合一维卷积神经网络的深度学习算法,不仅简化了对西兰花农药残留的识别过程,还提升了识别效率和识别精度。The detection of pesticide residues in agricultural products is an important step in ensuring the food safety of agricultural products,while traditional detection methods are cumbersome and costly.Using broccoli as a sample,this article used hyperspectral technology combined with machine learning algorithms and deep learning algorithms to provide a simple,fast,low-cost,and non-destructive method for detecting pesticide residues in broccoli.The study collected hyperspectral images in 400~1000 nm of broccoli samples sprayed with different types of pesticides and clean water.Two data preprocessing methods,namely multivariate scattering correction(MSC)and Savitzky-Golay smoothing(SG smoothing),as well as principal component analysis(PCA),competitive adaptive reweighted sampling(CARS),and successive projection algorithm(SPA)were used to reduce the dimensionality.A support vector machine(SVM)recognition model was established for pesticide residue discrimination.The SVM-SG-SPA combination has the best discrimination effect,with recognition accuracy of 92.86%,94.29%,91.43%,and 92.86%for high-efficiency cypermethrin,chlorpyrifos,imidacloprid,and water,respectively.A one-dimensional convolutional neural network(1D-CNN)model was established using raw spectral data,which achieved recognition accuracy of 94.29%,95.71%,94.29%,and 97.14%for high-efficiency cypermethrin,chlorpyrifos,imidacloprid,and water,all of which were higher than the SVM model.The results indicated that the combination of hyperspectral imaging technology and deep learning algorithms as 1D-CNN not only simplified the recognition process of pesticide residues in broccoli,but also improved recognition efficiency and accuracy.

关 键 词：高光谱技术 西兰花 农药残留识别 卷积神经网络 

分 类 号：S24[农业科学—农业电气化与自动化]