高光谱成像的多种类柑橘病虫药害叶片检测方法研究

影响柑橘生长的病虫药害种类繁多，目前的检测方法大多针对单一病症，开发基于高光谱成像和机器学习的多种类柑橘病虫药害叶片快速精准检测方法，对果园精准施药和柑橘产业健康发展具有重要意义。以果园自然发病的柑橘叶片为研究对象，包括柑橘正常叶(50片)、溃疡病叶(50片)、煤烟病叶(103片)、缺素病叶(60片)、红蜘蛛叶(56片)和除草剂危害叶(85片),采集350～1 050 nm波段内的高光谱数据。分别利用一阶求导(1^stDer)、多元散射校正(MSC)和中值滤波(MF)方法对原始(Origin)高光谱数据进行预处理，对预处理后的高光谱数据采用主成分分析(PCA)和竞争性自适应重加权(CARS)算法提取特征波长，CARS降维得到的特征波长分别为10个、 5个、 12个和10个，4组PCA提取的特征波长均为7个，两种方法所得特征波长范围都集中在700～760 nm波段内。对全波段(FS)使用极限梯度提升树(XGBoost)算法，特征波长使用支持向量机(SVM)建立柑橘病叶多分类模型。采用XGBoost建立的检测识别模型有Origin-FS-XGBoost, 1^s...

Study on Detection Method of Leaves With Various Citrus Pests and Diseases by Hyperspectral Imaging

Numerous kinds of diseases and insect pests affect citrus’ growth, but most of the current detection methods are for a single condition. It is important for the accurate application of pesticides in orchards and the healthy development of the citrus industry that the development of a detection method based on hyperspectral imaging and machine learning to achieve rapid and accurate detection of multiple pests and diseases on citrus leaves. Naturally onset citrus leaves in orchards were used as research objects, including normal citrus leaves (50 pieces), ulcer disease leaves (50 pieces), soot disease leaves (103 pieces), nutrient deficiency disease leaves (60 pieces), and red spider leaves (56 pieces) and herbicide damage leaves (85 pieces), hyperspectral data in the 350～1 050 nm band were collected. First-order derivation (1stDer), multivariate scattering correction (MSC) and median filtering (MF) were used to preprocess the original (Origin) hyperspectral data, principal component analysis (PCA) and competitive adaptive weighting (CARS) algorithms were used to extract characteristic wavelengths from the prepossessed hyperspectral data. Characteristic wavelengths obtained by CARS were 10, 5, 12 and 10 respectively, and the 4 sets of characteristic wavelengths obtained by PCA were all 7, ranging in the 700～760 nm band. The limit gradient boosting tree (XGBoost) was used for the full band (FS), and the support vector machine (SVM) was used for the characteristic wavelength to establish a multi-classification model of citrus diseased leaves. The classification models established by XGBoost are Origin-FS-XGBoost, 1stDer-FS-XGBoost, MSC-FS-XGBoost and MF-FS-XGBoost, and the overall classification accuracy (OA) obtained from the detection of 6 kinds of diseases and insect pests leaves was 94.32%, 93.60%, 95.98% and 96.56% respectively; the classification models established by SVM are Origin-CARS-SVM, 1stDer-CARS-SVM, MSC-CARS-SVM, MF-CARS-SVM, Origin-PCA-SVM, 1stDer-PCA-SVM, MSC-PCA-SVM and MF-PCA-SVM, model OA was 93.63%, 90.26%, 87.90%, 91.95%, 87.53%, 90.82%, 83.50% and 90.98% respectively. The experimental results demonstrate that the recognition rate of the XGBoost model with FS as input was better than the SVM model with characteristic wavelength as input. The OA of the MF-FS-XGBoost model was 96.56%, the Recall was 95.91%, and the model training time was 63 s. The overall performance was the best; the CARS-SVM modeling effect was better than PCA-SVM. After pre-processing by all three methods, the recognition rate of the CARS-SVM model was above 87%, and the recognition rate of the PCA-SVM model was above 83%. The results show that hyperspectral imaging technology combined with machine learning methods can classify and identify multiple species of citrus pests and diseases, providing a theoretical basis for rapid and non-destructive detection and control of citrus pests and diseases.