欢迎访问《中国粮油学报》编辑部网站！

李艳宏,戴丹,郑剑,金佳东,张帅杰,梁子乐,陈思思,董晨,郑辛煜.基于高光谱技术及SA-CNN模型的甘薯黑斑病早期预测研究[J].中国粮油学报,2025,40(4):186-196

基于高光谱技术及SA-CNN模型的甘薯黑斑病早期预测研究

Research on early prediction of sweetpotato black rot disease based on hyperspectral technology and SA-CNN model

投稿时间：2024-05-22 修订日期：2024-08-25

DOI：

英文关键词:hyperspectral technology sweetpotato black rot early prediction self-attention mechanism convolutional neural network

基金项目:国家青年科学基金项目（32301585），国家青年科学基金项目（42001354）

作者	单位	邮编
李艳宏	浙江农林大学数学与计算机科学学院	311300
戴丹^*	浙江农林大学数学与计算机科学学院	311300
郑剑	浙江农林大学食品与健康学院
金佳东	浙江农林大学数学与计算机科学学院
张帅杰	浙江农林大学数学与计算机科学学院
梁子乐	浙江农林大学数学与计算机科学学院
陈思思	临安区农林技术推广中心
董晨	浙江农林大学数学与计算机科学学院
郑辛煜	浙江农林大学数学与计算机科学学院

摘要点击次数: 0

全文下载次数: 0

中文摘要:

甘薯块根在储运过程中易受长喙壳菌侵染而引发黑斑病,因此在发病初期进行早期预测并有效预防尤为重要。鉴于此,本研究提出将高光谱技术和深度学习模型相结合对甘薯块根黑斑病进行早期预测,以接种长喙壳菌后的甘薯块根为实验对象,定期采集染病甘薯块根的高光谱图像和对应的总酚含量。首先基于全波段进行模型训练,使用Savitzky-Golay平滑(SG)、一阶求导(D1)和多元散射校正(MSC)三种光谱预处理方法分别与机器学习模型(BPNN、SVR和PLSR)和深度学习模型(CNN和LSTM)相结合进行对比分析。结果表明,经MSC预处理后的CNN模型的预测效果最好,预测集的决定系数(R2P)为0.865,均方根误差(RMSEP)为0.151 mg/g,剩余预测偏差(RPD)为2.724。因此在CNN模型的基础上引入自注意力机制进行优化,优化后的模型SA-CNN较原模型预测精度明显提升。为了进一步减少波段冗余信息,本研究基于特征波段进行模型训练,使用PCA、CARS和SPA三种特征提取方法建立SA-CNN模型进行预测,结果表明,MSC-CARS-SA-CNN模型具有最佳的预测性能,R2P 、RMSEP 和RPD分别为0.959、0.083 mg/g和4.961。以上说明,本研究结果有望为甘薯块根贮藏期间品质监测提供一种有效的方法。

英文摘要:

Sweetpotato tuberous roots often develop black rot disease during storage and transportation due to infection by Ceratocystis fimbriata. Therefore, early prediction and effective prevention at the initial stage of the disease became particularly important. In this regard, this study proposes combining hyperspectral technology with deep learning models for the early prediction of black rot disease in sweet potato tuberous roots. The experiment used sweet potato tuberous roots inoculated with Ceratocystis fimbriata and regularly collected hyperspectral images and corresponding total phenolic content from the diseased roots. Initially, model training was conducted based on the full wavelengths. The study used three spectral preprocessing methods—Savitzky-Golay smoothing (SG), first derivative (D1), and multiplicative scatter correction (MSC)—and combined them with machine learning models (BPNN, SVR, and PLSR) and deep learning models (CNN and LSTM) for comparative analysis. The results indicated that the CNN model with MSC preprocessing achieved the best prediction effect, the prediction set determination coefficient (R2P) was 0.865, the root mean square error (RMSEP) was 0.151 mg/g, and the residual predictive deviation (RPD) was 2.724. Consequently, the study introduced a self-attention mechanism to optimize the CNN model, and the optimized model (SA-CNN) significantly improved the prediction accuracy over the original model. To further reduce redundant spectral information, the study conducted model training based on characteristic wavelengths. Using PCA, CARS, and SPA for feature extraction, the study developed the SA-CNN model for prediction. The results showed that the MSC-CARS-SA-CNN model exhibited the best prediction performance, and the R2P, RMSEP and RPD were 0.959, 0.083 mg/g, and 4.961. These findings suggested that the study results could provide an effective method for quality monitoring of sweet potato tuberous roots during storage.

关闭