高光谱数据处理算法的小麦赤霉病籽粒识别

赤霉病是小麦的一种主要病害，它会导致小麦减产甚至绝收，严重影响小麦种子质量，此外小麦受侵染分泌的真菌毒素危害人类身体健康。因此，小麦赤霉病籽粒的识别具有非常重要的意义。起初普遍采用色谱法和酶联免疫法进行赤霉病检测，这些方法设备昂贵、检测速度慢、准确性低。近年来，高光谱成像技术被广泛应用于农作物的识别与检测中，但是在小麦赤霉病检测的应用研究中，大多采用抽样检测的方法，图像采集完成后需要通过ENVI软件手动选取感兴趣区域。前期准备工作冗杂，而且容易发生漏检，漏检的小麦籽粒在存储运输过程中向周边籽粒快速侵染，难以保障小麦安全健康。鉴于此，利用高光谱成像系统结合机器学习提出了一种用于对大量小麦赤霉病籽粒样本快速可视化识别的算法，以降低漏检率并提升检测效率。实验分别采集健康小麦和染病小麦469～1 082 nm波段的高光谱图像，通过直方图线性拉伸结合图像分割的方法获取小麦样本的掩膜图像信息。利用Savitzky-Golay平滑去噪法与标准正态变量变换法（SNV）进行数据预处理，通过主成分分析法（PCA）和连续投影法（SPA）进行特征变量提取，筛选特征变量个数分别为4个和8个。在掩膜图像位置采集健康小麦样本与染病小麦样本各400份，其中75%用于建模集，25%用于测试集。采用十折交叉验证法结合线性判别分析法(LDA) 、K-近邻算法（KNN）、支持向量机（SVM）分别建立分类模型，测试集准确率都达到90%以上。随后比较了网格法（GRID）、粒子群算法（PSO）、遗传算法（GA）三种核参数寻优方法对SVM模型的影响，其中，SG-SPA-SVM（PSO）模型分类效果最优，建模集准确率为95.5%，均方根误差为0.212 1，测试集准确率为98%，均方根误差为0.141 4。基于样本点预测的基础之上，对掩膜获得所有小麦样本的光谱曲线进行预测并将识别结果反馈回掩膜中再进行伪彩色显示，实现染病籽粒可视化识别。结果表明，高光谱成像结合SG-SPA-SVM（PSO）算法建立的分类模型可以高效快速、准确无损、可视化的实现小麦赤霉病籽粒识别，为研制小麦赤霉病自动识别设备提供了算法基础。

Recognition of Fusarium Head Blight Wheat Grain Based on Hyperspectral Data Processing Algorithm

Fusarium head blight (FHB) is a major disease of wheat, which can lead to wheat yield reduction or even crop failure, seriously affecting the quality of wheat seeds. Mycotoxins secreted by the diseased wheat are deposited in the food chain, and ultimately endanger human health. Therefore, recognition of wheat scab is very important. First of all, chromatography and enzyme-linked immunosorbent assay (ELISA) are widely used to detect scab. These methods were expensive, slow and have low accuracy. In recent years, hyperspectral imaging technology has been widely used in crop identification and detection, but in the application of wheat scab detection, sampling detection method is mostly used. After image acquisition, the region of interest (ROI) is manually selected through ENVI software. Preliminary preparations are complicated and easy to be missed detection. The undetected wheat grains quickly infect the surrounding grains during storage and transportation, which is difficult to ensure the safety and health of wheat. In view of this, this paper presents a fast visual recognition algorithm for wheat scab samples based on hyperspectral imagery and machine learning to reduce the rate of missed detection and improve the detection efficiency. The hyperspectral images of healthy wheat and infected wheat in 469~1 082 nm band were collected, and the mask image information of wheat samples was accurately obtained by histogram linear stretching combined with image segmentation. Savitzky-Golay smoothing denoising method and standard normal variable transformation (SNV) method are used for data preprocessing. Principal component analysis (PCA) and successive projections algorithm (SPA) were used to extract features, and 4 and 8 feature variables were selected respectively. There were 400 healthy wheat samples and 400 infected wheat samples collected in the mask image position, 75% of which were used for modeling set and 25% for testing set. Ten fold cross validation method combined with linear discriminant analysis (LDA), K-nearest neighbor algorithm (KNN) and support vector machine (SVM) was used to establish the classification model. The accuracy of the test set is over 90%, and the SPA dimension reduction model is better than the PCA dimensionality reduction model. Then, the effects of GRID, particle swarm optimization and GA three kernel parameter optimization methods on SVM model are compared. Among them, SG-SPA-SVM (PSO) model has the best classification effect. The accuracy of modeling set is 95.5%, REMS is 0. 2121, the accuracy of test set is 98%, and REMS is 0.141 4. Based on the prediction of sample points, the spectral curves of all wheat samples obtained by the mask were predicted and the recognition results were fed back to the mask and displayed in pseudo-color to realize the visual identification of infected grains. The results show that the classification model based on hyperspectral imaging technology combined with SG-SPA-SVM (PSO) algorithm can effectively, quickly, accurately, nondestructively and visually identify wheat scab, providing an algorithm basis for the development of automatic identification equipment for wheat scab.