摘要
[目的]针对传统的果蔬品质检测方法费时且破坏检测样本的问题,提出了一种新的苹果水心病检测方法。[方法]采用具有水心病的新疆阿克苏苹果作为样本,首先利用反向倍增法计算得出苹果每层组织的光学特性参数,然后利用苹果果皮、果肉、果核层的光学参数据构建3层迁移反演模型,使用蒙特卡罗法对苹果内部的光子轨迹进行仿真并获取苹果表面的仿真光亮度图。由于水心病通常发生在苹果果核层,通过仿真光亮度图学习苹果果核层的特征,然后迁移到实测苹果样本数据中进行学习和预测。基于光学特性参数反演的检测方法是利用蒙特卡罗法和卷积神经网络提取与水心病相关的苹果果核层光学特征,使用迁移学习将这些特征与实际苹果样本水心病病情建立联系,成功利用模拟数据提取实测数据中由于存在噪声而无法提取的有效特征,提高了水心病预测的速度和准确度。[结果]基于光学特性参数反演的水心病检测方法在二分类条件下的预测准确率达到94.3%,优于直接使用高光谱数据的卷积神经网络模型和SVM模型;在四分类条件下预测准确率能达到93.5%,相较于卷积神经网络模型提高约4%。[结论]试验结果说明所提出的基于光学特性参数反演的水心病检测方法能有效提取更多与水心病相关的苹果果核层高光谱特征,为苹果水心病的检测提供了一种新的思路。
[Objectives]Aiming at the problems that the traditional fruit and vegetable quality detection methods are time-consuming and will destroy the test samples,a new detection method for watercore disease in apples was proposed.[Methods]In the experiment,Xinjiang Aksu apples with watercore disease were used as samples.First,the reverse multiplication method was used to calculate the optical characteristic parameters of each layer of apple tissue,and then the optical parameter data of the apple peel,pulp,and core layers were used to construct a three-layer migration inversion model,and the Monte Carlo method was used to analyze the photons inside the apple.The trajectory was simulated and the simulated brightness map of the apple surface was obtained.Since watercore disease usually occurs in the apple core layer,the characteristics of the apple core layer were learned through the simulated luminance map,and then transferred to the measured apple sample data for learning and prediction.The detection method based on the inversion of optical characteristic parameters was that the Monte Carlo method and convolutional neural network were used to extract the optical features of the apple core layer related to watercore.The transfer learning was used to connect these features with the actual watercore disease condition of apple samples,and the simulated data was successfully used to extract effective features that could not be extracted due to noise in the measured data,which improved the speed and accuracy of watercore disease prediction.[Results]The watercore detection method based on the inversion of optical characteristic parameters had a prediction accuracy rate of 94.3%under the condition of two classifications,which was better than the convolutional neural network model and SVM model that directly used hyperspectral data;the prediction accuracy rate under four classifications could reach 93.5%,which was about 4%higher than the convolutional neural network model.[Conclusions]The experimental results showed that the prop
作者
张思旭
徐焕良
王江波
孙云晓
王浩云
ZHANG Sixu;XU Huanliang;WANG Jiangbo;SUN Yunxiao;WANG Haoyun(College of Artificial Intelligence,Nanjing Agricultural University,Nanjing 210095,China;Faculty of Plant Sciences,Tarim University,Alaer 843300,China)
出处
《南京农业大学学报》
CAS
CSCD
北大核心
2023年第5期986-994,共9页
Journal of Nanjing Agricultural University
基金
国家自然科学基金项目(31601545)
中央高校基本科研业务费专项资金(KYLH202006)
南京农业大学-塔里木大学科研合作联合基金项目(NNLH202006)
新疆建设兵团南疆重点产业支撑计划项目(2017DB006)。
关键词
苹果水心病预测
光学特性参数反演
高光谱
卷积神经网络
迁移学习
apple watercore disease prediction
optical property parameter inversion
hyperspectral data
convolutional neural network
transfer learning
