Abstract: Powdery mildew is one of the major diseases affecting the yield and quality of melon, hyperspectral technology was used to realize the early disease identification of melon powdery mildew. By using greenhouse melon as the research object, hyperspectral images of melon leaves containing 128 bands were collected, in which leaves within 1-4 days of inoculation with powdery mildew fungus were classified as the early diseased leaves and leaves with no fungus inoculation were healthy ones. Two algorithms, such as Successive Projections Algorithm (SPA) and Competitive Adaptive Reweighted Sampling (CARS), were used to extract characteristic wavelengths, and Principal Component Analysis (PCA) was applied to reduce the dimensionality of the original data. The original wavelength (Original), SPA characteristic wavelength (8), CARS characteristic wavelength (9) and PCA principal components (4) were used as the input variables of the recognition models, respectively. Combined with two ensemble leaning methods, Random Forests (RF) and Adaptive Boosting (AdaBoost). Eight early identification models of melon powdery mildew were constructed, including Original-RF, SPA-RF, CARS-RF, PCA-RF, Original-AdaBoost, SPA-AdaBoost, CARS-AdaBoost, PCA-AdaBoost. The model was evaluated by the ten-fold cross-validation method. The results showed that the accuracy of the proposed models were all above 90%, among which the Original-AdaBoost and Original-RF models had the highest average accuracy of 94.3% and 93.8%, respectively. SPA-AdaBoost effectively reduced the model input and achieved 93.3% recognition accuracy on the 1st day of the disease, with an average accuracy of 93.5%. 

Key words: melon, powdery mildew, hyperspectral, characteristic wavelength, machine learning, early identification

摘要： 白粉病是危害甜瓜产量和品质的主要病害之一，利用高光谱技术进行甜瓜白粉病早期病害识别研究。以温室甜瓜为研究对象，使用高光谱成像仪采集甜瓜叶片包含128个波段的高光谱图像，其中接种白粉病菌1~4天内的早期无病斑叶片为染病叶片，未接种病菌的叶片为健康叶片。采用连续投影算法(SPA)和竞争性自适应重加权算法(CARS)两种算法提取特征波长，运用主成分分析算法(PCA)对原始数据进行特征降维。分别以原始波长(Original)、SPA特征波长(8个)、CARS特征波长(9个)和PCA主成分(4个)作为早期识别模型的输入变量，结合随机森林(RF)和自适应增强(AdaBoost)两种集成学习算法，构建出8种甜瓜白粉病早期识别模型：Original-RF、SPA-RF、CARS-RF、PCA-RF、Original-AdaBoost、SPA-AdaBoost、CARS-AdaBoost、PCA-AdaBoost，并使用十折交叉验证方法对模型进行评价。结果表明，所建模型准确率均在90%以上，其中使用全波段的Original-AdaBoost和Original-RF模型平均准确率最高，分别为94.3%和93.8%；SPA-AdaBoost有效降低模型输入，在染病第1天识别准确率就达到93.3%，平均准确率达到93.5%。

关键词: 甜瓜, 白粉病, 高光谱, 特征波长, 机器学习, 早期识别