基于可见光谱特征波长提取和分类算法的柑橘黄龙病快检研究

Rapid Detection of Citrus Huanglongbing Based on Extraction of Characteristic Wavelength of Visible Spectrum and Classification Algorithm

柑橘黄龙病(HLB)是由亚洲韧皮杆菌引起的一种严重病害,目前无法根治。其防控具有重要意义和经济价值。当前利用健康和患病叶片的光谱差异对其进行诊断显示了良好的应用前景。因患病叶片在叶绿素反射区及O—H伸缩振动区的可见光谱与健康叶片存在显著差异,而可见光谱检测在采集和数据处理方面具有成本低、简便的优势,研究可见光谱的黄龙病快速检测方法具有可行性和重要意义。为了减少光谱数据冗余和计算量,实现精准的黄龙病的早期鉴别以及降低黄龙病相似病症的误诊率,采集了黄龙病患病地区共160个叶片样本。经qPCR测定分别将其分类标定为健康、轻度疾病、重度疾病和缺镁症四类。根据叶片样本在可见光波段450~800 nm的反射光谱特征,通过S-G平滑以及降采样等预处理光谱数据后,为了优选出尽可能囊括光谱特征信息的特征波长,分别使用遗传算法(GA)、连续投影算法(SPA)以及竞争自适应重加权采样法(CARS)对采集到的可见光谱数据进行特征波长提取和降维优选出特征波长,进一步降低模型复杂度,提高预测精度。综合泛化能力及检测速度的考量,在定性判别分析模型的选择中采用训练速度快,分析准确率高的最小二乘支持向量机(LS-SVM)以及随机森林(RF)对两种变量筛选算法降维后的数据进行分类判别。通过对不同的模型的验证优选,筛选出最佳的快检方案。对比发现,在建立的模型中,SPA-RF模型与其他模型比较,对于训练集和测试集的判别准确率分别达到了100%和97.5%。结果表明,连续投影算法以及随机森林的组合分类模型可以很好地实现黄龙病早期的病理鉴别,同时也能够很好地识别出黄龙病病叶与其他相似病症的差异,为柑橘黄龙病快速检测及防治提供了一种方法依据。

The Citrus Huanglongbing(HLB),caused by the Asian citrus psyllid,represents a severe disease with no current cure.Its control is of significant importance and economic value.Current diagnostic approaches utilizing the spectral differences between healthy and diseased leaves show promising applications.Diseased leaves exhibit notable differences from healthy ones in the chlorophyll reflection zone and the O—H stretching vibration region of the visible spectrum.With its low cost and simplicity in data collection and processing,the visible spectrum detection scheme presents a feasible and significant approach for the rapid detection of HLB.To reduce spectral data redundancy and computational load,achieving precise early identification of HLB and minimizing misdiagnosis of similar symptoms,this study collected 160 leaf samples from HLB-affected areas.These samples were classified into four categories—healthy,mild disease,severe disease,and magnesium deficiency-using qPCR determination.Reflecting on the characteristics of leaf samples in the visible light band(450~800 nm),the study involved preprocessing spectral data through S-G smoothing and down sampling.To select feature wavelengths that encapsulate maximum spectral information,Genetic Algorithm(GA),Successive Projections Algorithm(SPA),and Competitive Adaptive Reweighted Sampling(CARS)were employed for feature wavelength extraction and dimensionality reduction,further simplifying model complexity and enhancing prediction accuracy.Considering the generalization ability and detection speed,the study used the Least Squares Support Vector Machine(LS-SVM)and Random Forest(RF)to classify and discriminate the dimensionally-reduced data from the two variable selection algorithms.The best rapid detection scheme was selected by validating and optimising different models.-In comparison with others,the SPA-RF model achieved a discrimination accuracy of 100%and 97.5%for the training and test sets,respectively.The results demonstrate that the combination of SPA and RF in the classification model effectively accomplishes early pathological identification of HLB and distinguishes HLB-diseased leaves from similar symptoms,providing a basis for rapid detection and control of Citrus Huanglongbing.