基于紫外-可见-近红外光谱的苹果叶片褐斑病早期识别与诊断方法

Early Identification and Diagnosis of Apple Leaf Brown Spot Disease Based on Ultraviolet-Visible-Near-infrared Spectroscopy

摘要:

褐斑病是危害苹果叶片的真菌病害之一,严重影响苹果树的健康生长与果实品质。 本文提出了一种结合紫外可见近红外光谱技术和机器学习算法的苹果叶片褐斑病早期识别与诊断方法,展现出较好的识别诊断效果。 首先,利用日立 UH4150 型分光光度计(波长 200 ~ 3 050 nm)采集了健康、1 级病害和 3 级病害的苹果叶片高光谱数据。 然后,采用标准正态变换(Standard normal variate,SNV)、多项式平滑滤波(Savitzky-Golay,SG)、均值标准化 (Z-score standardization,ZS)、多元散射校正(Multiplicative scatter correction,MSC)和一阶导数(First derivative,FD)方法对原始光谱数据进行了预处理。 接着,采用主成分分析(Principal component analysis,PCA)和递归特征消除 (Recursive feature elimination, RFE)算法提取了与褐斑病叶片相关的特征波长。 最后, 建立了基于随机森林 (Random forest,RF)算法与卷积神经网络(Convolutional neural network,CNN)的判别模型,评估了使用特征波长进行疾病早期诊断的能力。 研究结果表明,SG-RFE-RF模型表现出优异的诊断性能,其在建模集和预测集上的分类准确率分别为 96. 03% 和 94. 34% 。 该研究为开发苹果叶片褐斑病早期检测传感器提供了技术支持。

Abstract:

Brown spot disease is one of the fungal diseases affecting apple leaves, severely impacting the healthy growth of apple trees and the quality of their fruit. An early detection and diagnosis method for brown spot disease in apple leaves was presented by integrating ultraviolet visible near-infrared spectroscopy with machine learning algorithms, demonstrating promising diagnostic performance. Firstly, hyperspectral data of healthy, grade 1, and grade 3 diseased apple leaves were collected by using a Hitachi UH4150 spectrophotometer (wavelength 200 ~ 3 050 nm). Then, the raw spectral data were preprocessed using standard normal variate (SNV), Savitzky-Golay (SG), Z-score standardization (ZS), multiplicative scatter correction (MSC), and first derivative (FD). Subsequently, principal component analysis (PCA)and recursive feature elimination (RFE)algorithms were employed to extract feature wavelengths related to brown spot in the apple leaves. Finally, discriminant models based on the random forest (RF)algorithm and convolutional neural network (CNN)were established to evaluate the capability of using feature wavelengths for early disease diagnosis. The results showed that the SG RFE RF model achieved classification accuracies of 96. 03% on the modeling set and 94. 34% on the prediction set, demonstrating the best performance. The proposed method showed strong robustness and generalization ability, indicating its potential for practical applications in orchard disease monitoring. Compared with traditional inspection methods, this approach can provide a rapid, non-destructive, and objective solution for early disease detection. The research result can offer valuable technical support for the development of intelligent detection systems and precision agriculture practices for apple production.