基于高光谱成像技术和自适应增强网络的水蜜桃产地溯源

Origin tracing of Prunus persica based on hyperspectral imaging technology and adaptive boosting

中文关键词:  高光谱成像技术  AdaBoost算法  水蜜桃  产地溯源  机器学习

英文关键词:hyperspectral imaging technology  adaptive boosting algorithm  Prunus persica  origin tracing  machine learning

中文摘要:

      目的 基于高光谱成像技术和自适应增强网络(adaptive boosting, AdaBoost)算法建立最佳的不同产地的水蜜桃分类模型。方法 本研究采集了来自无锡、徐州和宁波3个不同产地的‘白凤’水蜜桃高光谱图像, 通过提取感兴趣区域获得各产地的光谱信息, 并利用提取的光谱信息进行建模。分别比较在3种预处理方法下的AdaBoost、随机森林、长短期记忆网络、径向基函数神经网络产地分类模型, 并选出最优模型。随后, 利用连续投影算法(successive projections algorithm, SPA)和无信息变量剔除进行特征波长的提取, 以进一步优化模型。结果 实验结果表明, 全光谱下基于标准正态变量(standard normal variate, SNV)预处理的AdaBoost模型分类效果最好, 训练集和测试集的准确率分别达到100.0%和98.7%, 召回率和F1值分别为95.8%和97.9%。在对AdaBoost模型采用SPA提取9个特征波长后, SPA-SNV-AdaBoost模型训练集和测试集准确率都达到了100.0%, 召回率和F1值也达到100.0%。结论 本研究基于高光谱成像技术和AdaBoost算法, 构建了一种全新的水蜜桃不同产地分类模型, 经过预处理和特征提取, 该模型获得了较高的精度, 能够有效地进行水蜜桃产地溯源, 为水蜜桃品牌鉴别真伪打下了坚实的基础。

英文摘要:

      Objective To establish an optimal classification model for Prunus persica from different origins using hyperspectral imaging technology and the adaptive boosting (AdaBoost) algorithm. Methods This study collected hyperspectral images of ‘Baifeng’ Prunus persica from 3 origins: Wuxi, Xuzhou, and Ningbo. Spectral information from the 3 origins was obtained by extracting the region of interest, the spectral data were then modeled to compare the optimal models of AdaBoost, random, long short-term memory, and radial basis function networks under 3 different preprocessing methods for origin classification. Subsequently, feature wavelengths were extracted using the successive projections algorithm (SPA) and uninformative variable elimination to refine the model. Results The experimental results indicated that the AdaBoost model based on standard normal variate (SNV) preprocessing achieved the best classification performance under the full spectrum, with accuracy rates of 100.0% for the training set and 98.7% for the test set. The recall and F1-score were respectively 95.8% and 97.9%. After applying the SPA to extract 9 characteristic wavelengths for the AdaBoost model, the SPA-SNV-AdaBoost model achieved accuracy rates of 100.0% for both the training and test sets, with recall and F1-score also reaching 100.0%. Conclusion This study establishes a novel classification model for peaches from different origins based on hyperspectral imaging technology and the AdaBoost algorithm. After preprocessing and feature extraction, the model achieves high accuracy, enabling effective traceability of peach origins and laying a solid foundation for the authentication of peach brands.