基于高光谱成像与深度学习的乌菜叶片亚硝酸盐含量无损检测

Non-destructive Detection of Nitrite Content in Tatsoi Leaves Based on Hyperspectral Imaging and Deep Learning

摘要:

为实现乌菜亚硝酸盐含量快速无损检测，本研究提出一种基于高光谱成像与深度学习的高精度定量检测方法。在培养箱中设置4组不同光强环境，每组放置50株乌菜，获取不同光强环境下乌菜叶片样本，每组200片作为试验数据集。利用高光谱成像设备采集乌菜叶片样本高光谱图像信息，进行亚硝酸盐含量预测。通过感兴趣区域（ROI）和阈值分割处理图像获取平均光谱信息。利用不同预处理方法对光谱图像进行预测模型构建并分析性能，确立小波变换去噪（WTD）算法作为最佳预处理方法，并对使用WTD预处理的光谱数据进一步分析。使用4种传统降维方法对预处理后最佳光谱数据进行降维，分别通过支持向量机回归（SVR）建立模型与引入注意力块（SE）的一维卷积神经网络（1D-CNN）深度特征提取后建立的SVR模型进行分析。结果表明，基于1D-CNN-SE提取深度特征的SVR模型对不同光强环境中乌菜叶片亚硝酸盐含量预测效果较好，测试集决定系数（R2p）、均方根误差（RMSEP）和相对分析误差（RPD）分别为0.8849、0.07381 μmol/g和2.7631，表明在光强环境下建立的WTD-1D-CNN-SE-SVR模型性能较佳，深度学习方法结合高光谱成像无损检测技术能够有效实现乌菜叶片亚硝酸盐毒害的无损检测，为叶菜无机物毒害无损检测提供技术支持。

Abstract:

Aiming to achieve rapid and non-destructive detection of nitrite content in Tatsoi (Brassica rapa subsp.narinosa), a high-precision quantitative detection method was proposed based on hyperspectral imaging and deep learning.Four groups of different light intensities were set up in a growth chamber, with 50 Tatsoi plants placed in each group.Leaf samples of Tatsoi under different light intensity environments were collected, and 200 leaf slices per group were used as the experimental dataset.Hyperspectral images of the leaf samples were acquired by using a hyperspectral imaging device to predict their nitrite content.The average spectral information was extracted by processing the images through region of interest (ROI) selection and threshold segmentation.Various preprocessing methods were employed to construct prediction models on the spectral data, and their performances were compared.Wavelet transform denoising (WTD) was identified as the optimal preprocessing method.The preprocessed spectral data using WTD were further analyzed.Four traditional dimensionality reduction methods were applied to reduce the dimensionality of the optimal preprocessed spectral data.Support vector regression (SVR) models were established on the reduced data, and further analysis was conducted by using SVR models incorporating deep features extracted by a one-dimensional convolutional neural network (1D-CNN) integrated with a squeeze-and-excitation (SE) attention block.The results demonstrated that the SVR model based on deep features extracted by 1D-CNN-SE achieved superior predictive performance for nitrite content in Tatsoi leaves under different light intensity environments.On the test set, the coefficient of determination (R2p), root mean square error of prediction (RMSEP), and residual predictive deviation (RPD) were 0.8849, 0.07381 μmol/g, and 2.7631, respectively.These findings indicated that the WTD-1D-CNN-SE-SVR model established under light intensity conditions exhibited favorable performance.The combination of deep learning methods with hyperspectral imaging non-destructive detection technology can effectively achieve non-destructive detection of nitrite toxicity in Tatsoi leaves, providing technical support for non-destructive detection of inorganic toxin stress in leafy vegetables.