基于随机森林算法的玉米品种高光谱图像鉴别  

Identification of Maize Seed Varieties Based on Random Forest and Hyperspectral Technique

作　　者：邵琦 陈云浩 杨淑婷 赵逸飞 李京 SHAO Qi;CHEN Yun-hao;YANG Shu-ting;ZHAO Yi-fei;LI Jing(State Key Laboratory of Remote Sensing Science,Beijing Normal University,Beijing 100875;Faculty of Geographical Science,Beijing Normal University,Beijing 100875;Beijing Key Laboratory of Environmental Remote Sensing and Digital Cities,Faculty of Geographical Science,Beijing Normal University,Beijing 100875;State Key Laboratory of Earth Surface Processes and Resource Ecology,Faculty of Geographical Science,Beijing Normal University,Beijing 100875;Institute of Agricultural Economics and Information Technology,Ningxia Academy of Agricultural and Forestry Sciences,Yinchuan 750002,China)

机构地区：[1]北京师范大学遥感科学国家重点实验室,北京100875 [2]北京师范大学地理科学学部,北京100875 [3]环境遥感与数字城市北京市重点实验室,北京师范大学地理科学学部,北京100875 [4]地表过程与资源生态国家重点实验室,北京师范大学地理科学学部,北京100875 [5]宁夏农林科学院农业经济与信息技术研究所,宁夏银川750002

出　　处：《地理与地理信息科学》2019年第5期34-39,共6页Geography and Geo-Information Science

基　　金：宁夏农林科学院科技创新引导项目(NKYG-18-01)

摘　　要：玉米品种直接影响到玉米的产量和品质,事关农业收入和食品安全,因此,如何准确、高效、无损地鉴别玉米品种具有重要意义。该文基于高光谱成像系统采集3个品种共600粒玉米在533~893.4nm波段(共146个波段)范围的高光谱图像,对其进行校正和预处理,利用Boruta算法筛选有效波段。在全波段、全波段和纹理信息、有效波段以及有效波段和纹理信息4种特征组合下,利用随机森林算法进行玉米品种识别研究。结果表明:4种特征组合下,随机森林的平均分类准确率达76.25%,Kappa系数均在0.6以上,分类效果均优于传统的偏最小二乘判别分析方法;从4种特征组合的分类结果看,融合纹理信息的随机森林判别模型识别精度显著提升,分类准确率达77.20%,Kappa系数在0.64以上;基于有效波段和纹理信息判别模型的分类准确率达78.30%,Kappa系数为0.675。由此可见,有效波段和纹理信息特征组合下的随机森林算法能充分利用高光谱图像的光谱和纹理信息,准确地鉴别玉米品种,为玉米品种的自动识别提供了一种新方法。Maize varieties directly affect the yield and quality of corn,which is related to agricultural income and food safety.Therefore,it is of great significance to accurately and efficiently identify the varieties of maize seeds.In this paper,a hyperspectral imaging system was used to acquire hyperspectral images of 600 maize seeds from 3 varieties within the wavelength range of 533~893.4 nm (146 bands).The image was then corrected and preprocessed,and the effective band was screened by the Boruta algorithm.The random forest algorithm was used to identify maize varieties under the combinations of full-band,full-band and texture information,effective band,and effective band and texture information.The results show that using the four combinations,the average classification accuracy by the random forest algorithm is 76.25%,the Kappa coefficients are above 0.6,and the classification effect is better than the traditional partial least squares discriminant analysis.According to the classification results,the recognition accuracy of the random forest discriminant model with fusion of texture information is significantly improved,the classification average accuracy is 77.20%,the Kappa coefficient is above 0.64,and the classification average accuracy based on the effective band and texture information discriminant model reaches 78.30%.The Kappa coefficient reaches 0.675.The research shows that the random forest algorithm under the combination of effective band and texture information features can make full use of hyperspectral spectral features and texture features to accurately identify maize varieties and provide a new method for automatic identification of maize varieties.

关 键 词：高光谱图像 玉米 随机森林 偏最小二乘判别分析 

分 类 号：TP79[自动化与计算机技术—检测技术与自动化装置] S513[自动化与计算机技术—控制科学与工程]