Zhang Xiaodong, Wang Yafei, Zhou Zhankun, Zhang Yixue, Wang Xinzhong
College of Agricultural Engineering, Jiangsu University, Zhenjiang 212013, China.
Basic Engineering Training Center, Jiangsu University, Zhenjiang 212013, China.
Foods. 2023 Jan 25;12(3):535. doi: 10.3390/foods12030535.
Leaf mildew is a common disease of tomato leaves. Its detection is an important means to reduce yield loss from the disease and improve tomato quality. In this study, a new method was developed for the multi-source detection of tomato leaf mildew by THz hyperspectral imaging through combining internal and external leaf features. First, multi-source information obtained from tomato leaves of different disease grades was extracted by near-infrared hyperspectral imaging and THz time-domain spectroscopy, while the influence of low-frequency noise was removed by the Savitzky Golay (SG) smoothing algorithm. A genetic algorithm (GA) was used to optimize the selection of the characteristic near-infrared hyperspectral band. Principal component analysis (PCA) was employed to optimize the THz characteristic absorption spectra and power spectrum dimensions. Recognition models were developed for different grades of tomato leaf mildew infestation by incorporating near-infrared hyperspectral imaging, THz absorbance, and power spectra using the backpropagation neural network (BPNN), and the models had recognition rates of 95%, 96.67%, and 95%, respectively. Based on the near-infrared hyperspectral features, THz time-domain spectrum features, and classification model, the probability density of the posterior distribution of tomato leaf health parameter variables was recalculated by a Bayesian network model. Finally, a fusion diagnosis and health evaluation model of tomato leaf mildew with hyperspectral fusion THz was established, and the recognition rate of tomato leaf mildew samples reached 97.12%, which improved the recognition accuracy by 0.45% when compared with the single detection method, thereby achieving the accurate detection of facility diseases.
叶霉病是番茄叶片的常见病害。其检测是减少该病害造成的产量损失和提高番茄品质的重要手段。在本研究中，通过结合叶片内部和外部特征，开发了一种利用太赫兹高光谱成像对番茄叶霉病进行多源检测的新方法。首先，利用近红外高光谱成像和太赫兹时域光谱从不同病害等级的番茄叶片中提取多源信息，同时采用Savitzky Golay（SG）平滑算法去除低频噪声的影响。使用遗传算法（GA）优化近红外高光谱特征波段的选择。采用主成分分析（PCA）优化太赫兹特征吸收光谱和功率谱维度。利用反向传播神经网络（BPNN）结合近红外高光谱成像、太赫兹吸光度和功率谱，建立了不同等级番茄叶霉病侵染的识别模型，这些模型的识别率分别为95%、96.67%和95%。基于近红外高光谱特征、太赫兹时域光谱特征和分类模型，通过贝叶斯网络模型重新计算番茄叶片健康参数变量后验分布的概率密度。最后，建立了高光谱融合太赫兹的番茄叶霉病融合诊断与健康评价模型，番茄叶霉病样本的识别率达到97.12%，与单一检测方法相比，识别准确率提高了0.45%，从而实现了设施病害的准确检测。
