基于近红外光谱与深度域自适的芒果品种识别  

Using DANN to Classify the Mango Varieties With NIR Spectroscopy

作　　者：李统乐 陈潇 陈孝敬[1] 陈熙[1] 袁雷明 石文 黄光造 LI Tong-le;CHEN Xiao;CHEN Xiao-jing;CHEN Xi;YUAN Lei-ming;SHI Wen;HUANG Guang-zao(College of Electrical and Electronic Engineering,Wenzhou University,Wenzhou 325000,China;Wenzhou Institute of Industry&Science,Wenzhou 325000,China)

机构地区：[1]温州大学电气与电子工程学院,浙江温州325000 [2]温州工业科学研究院,浙江温州325000

出　　处：《光谱学与光谱分析》2025年第5期1251-1256,共6页Spectroscopy and Spectral Analysis

基　　金：国家自然科学基金项目(62105245);温州市科技专家项目(X2023011);温州市基础公益项目(N20220015)资助。

摘　　要：不同品种的芒果不仅具有不同的品质,还具有不同的经济效益。传统芒果品种识别方法通常较为依赖从业者的经验费时费力。如何快速地对芒果品种进行识别是需要解决的问题。近红外光谱是一种快速、无损的检测方法。通常可以通过结合机器学习方法和近红外光谱数据识别不同品种的芒果。由于仪器、季节或年份等原因,同一品种芒果的近红外光谱可能会有差别。差别会导致先前批次测量的样本(源域)和新批次测量的样本(目标域)之间的数据分布有所不同。数据分布的不同会导致先前建立好的分类模型不能对新测量的芒果样本准确分类。本工作重点研究工作温度和季节等因素导致芒果近红外光谱数据的分布差异。域自适应方法可以解决由于数据分布不同而导致的模型不适用。使用深度域自适应神经网络(DANN)模型解决这类问题。DANN模型通过对抗性的学习方式将两个域之间的特征对齐,有效实现了跨领域的样本分类。将DANN与无监督动态正交投影(uDOP)和联合分布适配(JDA)两种基于统计学习的传统域自适应方法进行了对比。应用这三种方法的实验结果表明,DANN模型在芒果种类的二分类任务中,对测试集的分类准确率达到94%。在芒果种类的多分类任务中,DANN模型的分类准确率相比uDOP和JDA方法高超过10%。结果表明,DANN模型可以很好地解决由于两个领域之间近红外光谱数据分布不同所导致的芒果品种识别的问题。Different cultivars of mangoes can not only represent different qualities but also produce different economic benefits.Traditional mango variety identification methods often rely more on the experience of practitioners and are time-consuming and laborious.Therefore,how to quickly classify mango cultivars is an emerging problem that needs to be solved.Near-infrared(NIR)spectroscopy technology is a fast and non-destructive approach.Users can often identify different mangoes by combining machine learning methods with near-infrared spectroscopy data.However,the NIR spectral information of the same variety of mangoes can vary due to variations in different instruments,seasons,and years.These differences result in a different distribution between the previously measured sample data(source domain)and the newly measured sample data(target domain).Consequently,the present classification model cannot correctly classify new mango samples.Domain adaptation methods can solve this problem of model inapplicability caused by different data distributions.This article focuses on the distribution differences of mango near-infrared spectroscopy data caused by factors such as working temperature and season.The domain adaptation methods can solve the problem of model unsuitability caused by different data distributions.This article used a deep domain adaptive neural network(DANN)model to solve this problem.The DANN model effectively achieves cross-domain sample classification models by aligning features between two domains through adversarial learning.This article compared DANN with unsupervised dynamic orthogonal projection(uDOP)and joint distribution adaptation(JDA),two traditional domain adaptation methods based on statistical learning.The experimental results of applying these three methods in this article showed that the DANN model could achieve a classification accuracy of 94%for the test set in the binary classification task of mango varieties.In the multi-classification task of mango varieties,the classification accuracy of th

关 键 词：芒果 近红外光谱 域自适应 神经网络 

分 类 号：O433.4[机械工程—光学工程]