CAJ | 학술논문

为了快速、准确地进行水稻穗瘟病害程度分级，以实现水稻品种抗性评价或精准的田间化学防治，该研究提出了一种光谱词袋（bag of spectrum words, BoSW）模型分析方法，分析稻穗的高光谱图像，自动评判穗瘟病害程度。首先，稠密规整地将高光谱图像分割成小立方格，计算每个立方格像素的平均全波段包络矢量，用K-Means算法聚类形成典型光谱包络词典。词典中光谱包络“词”（word）用作高光谱图像表达的“基”，直方图统计各光谱“词”在高光谱图像样本中的出现频度，形成光谱图像的词袋表达。采用HyperSIS-VNIR-QE光谱成像仪获取田间采集的170株稻穗样本高光谱图像，用BoSW方法生成其词袋表达；植保专家根据病害程度类别确定光谱图像样本标签。随机选择2/3“词袋表达-病害程度等级标签”数据对构成训练集，采用卡方-支持矢量机（chi-square support vector machine, Chi-SVM）分类算法建立穗瘟病害程度分级模型。余下的1/3样本构成测试集，测试穗瘟病害等级模型的预测性能，分类识别精度为94.72%，高于主成分分析（principle component analysis, PCA）、敏感波段选择等传统光谱分析方法，其识别精度分别为83.83%和79.83%。该研究提高了穗瘟病分级的自动化程度和准确率，也可为其他病害分级检测提供参考。
위료쾌속、준학지진행수도수온병해정도분급，이실현수도품충항성평개혹정준적전간화학방치，해연구제출료일충광보사대（bag of spectrum words, BoSW）모형분석방법，분석도수적고광보도상，자동평판수온병해정도。수선，주밀규정지장고광보도상분할성소립방격，계산매개립방격상소적평균전파단포락시량，용K-Means산법취류형성전형광보포락사전。사전중광보포락“사”（word）용작고광보도상표체적“기”，직방도통계각광보“사”재고광보도상양본중적출현빈도，형성광보도상적사대표체。채용HyperSIS-VNIR-QE광보성상의획취전간채집적170주도수양본고광보도상，용BoSW방법생성기사대표체；식보전가근거병해정도유별학정광보도상양본표첨。수궤선택2/3“사대표체-병해정도등급표첨”수거대구성훈련집，채용잡방-지지시량궤（chi-square support vector machine, Chi-SVM）분류산법건립수온병해정도분급모형。여하적1/3양본구성측시집，측시수온병해등급모형적예측성능，분류식별정도위94.72%，고우주성분분석（principle component analysis, PCA）、민감파단선택등전통광보분석방법，기식별정도분별위83.83%화79.83%。해연구제고료수온병분급적자동화정도화준학솔，야가위기타병해분급검측제공삼고。
Estimation of panicle blast level plays an important role in high-quality production of rice. It helps to quantitatively assess the level of blast resistance and severity in the field to make appropriate decisions in gauging cultivar resistance in rice breeding or precisely controlling blast epidemic. However, it is difficult to evaluate the blast disease degree automatically and accurately. In this study, a novel grading method for panicle blast severity based on hyperspectral imaging technology is proposed. The method defines a bag of spectrum words (BoSW) model for hyperspectral image data representation. The BoSW model based on hyperspectral image data representation is used as the input of a Chi-square kernel support vector machine (Chi-SVM) classifier for predicting the rice panicle blast level. More precisely, dense grids are firstly extracted over the spatial X-and Y-axes across the whole spectral Z-axis. The average spectrum curve of all the pixels within a grid cube is calculated. Then, K-Means clustering would be performed on the large collection of average spectrum curves from the training samples to form the dictionary of spectrum words. Next, each spectrum curve on the grid cube is quantized into one of spectrum words. Each hyperspectral image of rice panicle is transformed into a map of spectrum words. All the spectrum words are distributed evenly on the spatial XY-axis plane. BoSW model for each hyperspectral data cube is then formed by means of histogram statistics of spectrum word occurrences. Finally, a Chi-SVM classifier is trained using the BoSW representations of rice panicle hyperspectral images for predicting panicle blast infection levels. The proposed BoSW method uses both the image and full-spectrum information by means of regular grid cube extraction, which utilizes the full potential of the imaging sensing system. Meanwhile, the representation dimension for each hyperspectral image is significantly reduced, i.e. 100 here, and thus relieving modeling difficulty. The procedure of clustering helps to find the representative spectrum curves and quantizing helps to transform all the continuous-state spectrum curves into one of representative spectrum curve. Thus the proposed BoSW method is invariant to complicated noise and robust to rice cultivars.To verify the proposed BoSW method, a total of 170 fresh rice panicles covering more than 50 cultivars are collected from an experimental field for the performance evaluation. The experimental field is located in regional testing area for evaluating rice cultivars in Guangdong province. Therefore, all the rice plants in this area are naturally inoculated as the area is a typical source of rice blast fungus. The hyperspectral images of all the rice panicles are acquired using HyperSIS-VNIR-QE imaging spectrometer and then are transformed into 100-dimension BoSW representation for the construction of Chi-SVM classifier. Four-class label of hyperspectral image sample is determined by plant protection expert according to description of blast infection levels. Two thirds of the labeled BoSW representations are randomly selected for training and the rest for testing. Experimental results show that the proposed BoSW based method achieved high classification accuracy of 94.72%. This result is much better than traditional hyperspectral image analysis methods such as Principal component analysis (PCA), sensitive bands selection and etc. Moreover, the proposed BoSW demonstrates strong robustness to rice cultivars, which is vital for the wide and practical application. This research improves the classification accuracy of rice panicle blast grading and provides a reference to evaluate other disease level grading as well.