农作物生长发育过程中经常会遭到病虫害等外界因素侵染, 如果不能实施有效的监测诊断和科学的防治, 极易引起农药喷洒不当或过量, 不仅会影响作物的产量和种植户的经济效益, 还会造成严重的环境污染。 近年在广西大棚厚皮甜瓜上发生了一种严重的由瓜类尾孢( Cercospora citrullina)引起的甜瓜叶斑病, 导致甜瓜减产和种植户的经济损失。 故此应用高光谱成像开展甜瓜叶片的尾孢叶斑病检测, 获取健康甜瓜叶片和受瓜类尾孢感染的具有不同病变程度的甜瓜叶片在380~1 000和900~1 700 nm的高光谱图像, 选取感兴趣区域并获取相应的平均光谱反射率, 比较发现健康叶片和不同病变程度叶片染病区域的平均反射率差异显著。 在540 nm处附近, 健康叶片和病变程度轻微的叶片的光谱具备波峰形态, 随着病变程度增加, 波峰逐渐消失; 在700~750 nm处附近, 叶片反射率曲线急剧上升, 出现绿色植物光谱曲线显著的“红边效应”特征; 750~900 nm范围, 健康叶片与轻微病变区域的光谱反射率变化趋于平稳, 而其他病变区域的反射率呈上升趋势, 且健康叶片的反射率高于病变区域, 反射率随病变程度增加而下降, 这一变化规律一直持续到近红外波段的900~1 350 nm范围。 运用主成分分析、 最小噪声分离法观察叶片早期病变的特征, 经主成分分析和最小噪声分离法处理后, 特别是对于早期病变, 样本受感染后发病的区域更为明显。 基于高光谱图像提取的前三个主成分得分绘制三维散点图, 虽然不同病变程度的部分样本有重叠, 但病变样本与健康样本的分布区分明显。 应用K-近邻法和支持向量机方法建立叶片病变判别模型, 结果显示: KNN模型对健康样本测试集判别率为98.7%, 病变样本的判别率随病变程度加重而逐渐升高; 对病变程度较轻样本, 支持向量机模型相比于KNN模型而言, 判别正确率更高、 分类效果更好; 总体上, 高光谱图像对健康样本的判别率较高(>97%), 可用于健康样本与病变样本的识别, 但对不同病变程度的区分效果欠佳。 研究结果表明, 高光谱成像可用于甜瓜尾孢叶斑病的检测, 对不同病变程度的区分效果仍有待提高。

During crop growth, it is often infected by external factors such as pests and diseases. If effective monitoring, diagnosis and scientific control can’t be carried out, it will easily leds to improper or excessive spraying of pesticides. It will not only affect the yield of crops and the economic benefits of farmers, but also cause serious environmental pollution. In recent years, a serious muskmelon leaf spot caused by Cercospora citrullina occurred in Guangxi, which leads to yield reduction and economic losses. In this study, hyperspectral imaging technology was used to detect muskmelon Cercospora leaf spot. Hyperspectral images of healthy leaves and diseased leaves with varying degrees of lesionwere collected at 380~1 000 and 900~1 700 nm. Regions of interest were selected and the corresponding average reflectances spectra was obtained. It was found that the mean reflectance of healthy leaves and the diseased leaves were significantly different and changed regularly according to the degree of lesion. Near 540 nm, the spectra of healthy leaf and leaf with slight lesion had a peak, which disappeared gradually with the increase of lesion degree. In 700~750 nm, the leaf reflectance curve increased sharply, and there was a significant “red edge effect” of green plant spectral curve. In the range of 750~900 nm, the reflectance spectra of healthy leaves and leaves with mild lesions changed steadily. The reflectance of healthy leaves was higher than that of the lesion area. The reflectance decreased with the increase of lesion degree. And this change regularity lasted until 900~1350 nm in the near infrared region. Principal component analysis (PCA) and minimal noise fraction (MNF) were used to observe the characteristics of early leaf lesions. After pretreated with PCA and MNF, the area of infection was more obvious, especially for early lesions. Three-dimensional scatter plot was drawn based on the scores of the first three principal components extracted from hyperspectral images. Although some samples with different degrees of lesion overlap, the distribution of lesion samples and healthy samples is distinct. K-nearest neighbor (KNN) method and support vector machine (SVM) were used to establish the discriminant models. The correctness of KNN model for healthy sample discrimination in the test set was 98.7%. And the discriminant rate of lesion samples increases with the severity of lesion. For the lighter lesion samples, SVM model has higher discriminant accuracy and better classification effect than KNN model. Generally, hyperspectral images had a high discriminant rate (>97%) for healthy samples and lesion samples, however, the discrimination of different lesion degrees is not good enough. It can be concluded that hyperspectral imaging technology can be used to detect muskmelon Cercospora leaf spot disease, but the discrimination of different lesion degrees still needs to be improved in the future.