基于高光谱成像技术的高粱农药残留种类检测研究

Detection of pesticide residue types in sorghum based on hyperspectral imaging technology research

中文关键词:  高光谱成像  高粱  农药残留  无损检测  BP-AdaBoost

英文关键词:hyperspectral imaging  sorghum  pesticide residues  non-destructive testing  back propagation neural network with adaptive boosting

中文摘要:

      目的 基于高光谱成像技术实现对高粱农药残留种类的鉴别。方法 利用近红外高光谱成像系统采集高粱农药残留样品的高光谱数据, 建立基于BP神经网络自适应增强算法(back propagation neural network with adaptive boosting, BP-AdaBoost)、轻量梯度提升机(light gradient boosting algorithm, LGBM)、极度梯度提升(eXtreme gradient boosting, XGBoost)、支持向量机(support vector machine, SVM)的高粱农药残留分类模型; 采用了3种预处理方法和4种特征波长选择方法, 并构建基于特征波长信息的农药残留分类模型, 对比分析结果。结果 标准正态变换(standard normal variate, SNV)为最佳的预处理方法, 类型提升算法(type boosting algorithm, CatBoost)相比于梯度提升树(gradient boosting decision tree, GBDT)、竞争性自适应重加权采样法(competitive adaptive reweighted sampling, CARS)和主成分分析法(principal component analysis, PCA)选择的特征波长更具有代表性; 在所有分类模型中, SNV-CatBoost-BP-AdaBoost模型农药残留鉴别效果最好, 测试集平均分类正确率为95.17%。结论 高光谱成像技术结合BP-AdaBoost算法可以识别出高粱中农药残留的种类, 为检测高粱农药残留类别提供了一种新的方法。

英文摘要:

      Objective To identify sorghum pesticide residue types using hyperspectral imaging technology. Methods A near-infrared hyperspectral imaging system was utilized to collect hyperspectral data from sorghum pesticide residue samples. Classification models for sorghum pesticide residues were established employing the back propagation neural network with adaptive boosting (BP-AdaBoost), light gradient boosting algorithm (LGBM), eXtreme gradient boosting (XGBoost), and support vector machine (SVM). Three kinds of preprocessing methods and 4 kinds of feature wavelength selection techniques were applied. Additionally, classification models based on feature wavelength information were developed, followed by a comparative analysis of the outcomes. Results The standard normal variate (SNV) transformation emerged as the optimal preprocessing method, while the type boosting algorithm (CatBoost) exhibited more representative feature wavelengths compared to gradient boosting decision tree (GBDT), competitive adaptive reweighted sampling (CARS), and principal component analysis (PCA). Among all the classification models, the SNV-CatBoost-BP-AdaBoost model demonstrated the best performance in identifying pesticide residues, achieving an average classification accuracy of 95.17% on the test dataset. Conclusion The combination of hyperspectral imaging technology with the BP-AdaBoost algorithm can be employed to identify the types of pesticide residues in sorghum. This method offers a novel approach for detecting pesticide residue categories in sorghum.