基于SAGA-SVR预测模型的水稻种子水分含量高光谱检测  

Hyperspectral detection for moisture in rice seeds by SAGA-SVR prediction model

作　　者：芦兵[1,2] 孙俊[1] 杨宁[1] 武小红[1] LU Bing;SUN Jun;YANG Ning;WU Xiao-hong(School of Electrical and Information Engineering,Jiangsu University,Zhenjiang,Jiangsu 212013,China;Information Center,Jiangsu University,Zhenjiang,Jiangsu 212013,China)

机构地区：[1]江苏大学电气信息工程学院,江苏镇江212013 [2]江苏大学信息化中心,江苏镇江212013

出　　处：《南方农业学报》2018年第11期2342-2348,共7页Journal of Southern Agriculture

基　　金：国家自然科学基金项目(31471413);江苏省高校优势学科建设工程项目(苏政办发[2011]6号);江苏省六大人才高峰项目(ZBZZ-019)

摘　　要：【目的】利用高光谱技术测定水稻种子的水分含量,为其品质监测和筛选提供参考依据,从而提高水稻良种筛选率。【方法】通过电烘箱恒重法制备120份不同水分含量的水稻种子样本作为研究对象,利用多项式平滑(Savitzky-Golay, S-G)算法对原始光谱数据进行降噪平滑处理,采用连续投影算法(Successive projections algorithm, SPA)对预处理后的数据进行特征波长的优选。为提高建模效率,提高各水分含量区间光谱特征值的区分度,使用模糊C-均值聚类(Fuzzy C-means clustering,FCM)算法对各区间的样本数据进行聚类处理,最后利用支持向量回归机(Support vector regression, SVR)定量检测模型建立特征光谱数据与水稻种子水分含量的映射关系。【结果】由于FCM未达到预期的聚类效果,而引入遗传模拟退火算法(Simulated annealing genetic algorithm,SAGA)进行聚类,分别对基于原始特征值、FCM及SAGA聚类的SVR训练结果进行比较,发现基于SAGA聚类的光谱样本数据训练效果更好,预测集决定系数可达0.8956,均方根误差3.75%。由于决定系数不够理想,引入松弛变量降低间隔阈值,最终模型预测集决定系数为0.9286,均方根误差为3.42%,此时模型达最佳性能,能满足实际应用需求。【建议】基于聚类算法,提高光谱数据的准确性;通过合理调整模型参数,提高预测模型性能;推动高光谱农产品检测相关装备的研制。【Objective】The moisture content of rice seeds was determined by hyperspectral technique,which could provide reference for monitoring and screening the quality of rice seeds,so as to improve the yield of rice fine seeds screening.【Method】One hundred and twenty rice seed samples with different moisture contents were prepared by oven constant weight method.Savitzky-Golay(SG)algorithm was applied in smoothing and denoising the original hyperspectral data,successive projections algorithm(SPA)was used to select characteristic wavelengths for these preprocessed data.In order to improve the modeling efficiency and increase the discrimination of hyperspectral eigenvalues of each moisture interval,fuzzy C-mean clustering(FCM)algorithm was applied to cluster the sample data of each interval.Finally,the mapping relationship between feature hyperspectral eigenvalues and moisture content of the rice seeds was established by a quantitative detection model called support vector regression(SVR).【Result】Whereas,FCM did not achieve the desired clustering result,genetic simulated annealing algorithm(SAGA)was introduced for clustering.Then,comparing the SVR training results based on original eigenvalues,FCM and SAGA clustering respectively,it was found that the best training effects of hyperspectral sample data was based on SAGA clustering,the determinant coefficient of the prediction set was up to 0.8956,and the root mean square error was 3.75%.The interval threshold was reduced by introducing relaxation variable because the coefficient of decision was not ideal.The final prediction coefficient was 0.9286 and the root mean square error was 3.42%.The model achieved its best function and could meet the need of actual practice.【Suggestion】The following suggestions are proposed:improving the accuracy of spectral data based on clustering algorithm,the prediction model performance should be improved by adjusting the model parameters reasonably,promoting the development of equipment for detection of hyperspectral agricultur

关 键 词：水稻种子 高光谱 模糊C-均值聚类(FCM) 遗传模拟退火(SAGA) 支持向量回归机(SVR) 

分 类 号：S126[农业科学—农业基础科学]