葡萄霜霉病是全球危害最严重的葡萄病害, 对该病进行早期检测和防治, 可提高葡萄品质和产量, 提出一种基于多光谱荧光成像技术(MFI)和支持向量机模型(SVM)的霜霉病早期检测方法。 对人工接种霜霉病的葡萄叶片(145个)和健康对照叶片(145个)从叶背面连续6天进行多光谱荧光成像, 获得试验叶片16个荧光参数(4个单独波段F440, F520, F690, F740及其相互比值)的图像。 在分析不同荧光波段图像随接种天数(DPI)变化规律基础上, 通过单因素方差分析和相关性分析, 优选出进行霜霉病早期检测的4个波段特征F520, F690, F440/F740, F690/740, 利用这4个特征构建基于SVM的霜霉病检测模型。 试验发现, 16个荧光参数都有早期检测霜霉病的潜力, 四个单独波段中F440和F520比F690和F740对霜霉病的侵染更敏感, 6DPI才显症的病斑能在F440和F520波段2DPI(接种后第二天)的荧光图像中凸显, 接种叶片F440和F520波段荧光强度均随着DPI增加快速升高, 在2DPI显著高于健康叶片(p<0.01), 并随着DPI增加更加显著(p<0.0001); 接种叶片F690和F740波段荧光强度均随着DPI增加逐渐减小, 1DPI—3DPI与健康叶片无显著差异, 从4DPI开始显著低于健康叶片(p<0.05), 并在5DPI—6DPI更加显著(p<0.01); 健康叶片荧光参数变化很小。 F440极易受干扰, 变异系数最大, F520最稳定。 随着DPI增加, 叶片被侵染程度加深, 4个特征融合的SVM模型对健康和接种叶片检测准确率逐渐提高, 1DPI的准确率为65.6%, 3DPI检测准确率为82.2%, 整个试验周期(1DPI—6DPI)的平均检测准确率达84.6%, 高于单一特征中最优波段F520的阈值检测结果(1DPI的准确率为61.1%, 3DPI检测准确率为78.9%, 整个试验周期为80.0%)。 结果表明利用MFI技术和SVM模型能实现霜霉病显症前的早期检测, 为便携式葡萄霜霉病早期诊断设备的开发提供了理论依据。

Grapevine downy mildew is the most serious grape disease worldwide. Early detection of this disease can achieve early control so that quality and yield are improved. A test method based on multicolor fluorescence images (MFI) on grape leaves and a Support Vector Machine (SVM) model was proposed in the current study. Multicolor fluorescence imaging was performed on 145 inoculated leaves and 145 healthy leaves from the backside at six consecutive DPI (Days Post Innoculation). 16 fluorescence parameters (F440, F520, F690, F740 and their respective ratios) were obtained. Based on the image variation of four independent fluorescence wavelengths as DPI proceeding, single-factor ANOVA and correlation analysis were conducted. Four wavelengths of F520, F690, F440/F740 and F690/F740 were best selected with stronger detection ability of early infection and low correlation. For better detection, an SVM model was constructed with all four features. The results showed that the four basic bands F440, F520, F690, F740 and their ratios had the ability to detect early infection of grapevine downy mildew. F440 and F520 were more sensitive to the infection than F690 and F740. Start from 2 DPI, the area of the lesion could be highlighted in the fluorescence images of F440 and F520, at which the fluorescence intensity of the inoculated leaves was significantly higher than that of healthy leaves (p<0.01), and the difference increased with the increase of DPI (p<0.000 1). At F690 and F740 bands, the fluorescence intensity of inoculated leaves decreased gradually with the increase of DPI, and there was no significant difference between inoculated and healthy leaves from 1DPI to 3DPI. At 4DPI, inoculated leaves’ fluorescence intensity was significantly lower than that of heathy leaves (p<0.05) and the difference increased at 5DPI and 6DPI(p<0.01). The fluorescence parameters of healthy leaves changed little. F440 was the most susceptible to interference with the maximum coefficient of variation among the four bands, while F520 was more stable with the least coefficient of variation. With the increase of DPI, the detection accuracy of SVM model for distinguishing healthy and inoculated leaves was gradually improved, at 1DPI, the accuracy of SVM with multi-features was 65.6%, the accuracy of 3DPI achieved 82.2%, and the average accuracy was 84.6% in the whole experimental period (6 d), which was better than the best threshold method (F520 with 61.1% at 1DPI, 78.9% at 3DPI and 80.0% in the whole experimental period). In conclusion, the MFI technology with SVM model can achieve the early detection of downy mildew before the onset of symptoms, which provides a theoretical basis and proof for the development of portable equipment for early diagnosis of grape downy mildew.