| 基于多源遥感数据和机器学习算法的冬小麦产量预测研究 |
| |
| 引用本文： | 甘 甜,李 雷,李红叶,宋成阳,谢永盾,陶志强,肖永贵,孟亚雄.基于多源遥感数据和机器学习算法的冬小麦产量预测研究[J].麦类作物学报,2022(11):1419-1428. |
| |
| 作者姓名： | 甘 甜 李 雷 李红叶 宋成阳 谢永盾 陶志强 肖永贵 孟亚雄 |
| |
| 作者单位： | （1.甘肃农业大学农学院，甘肃兰州 730070；2.中国农业科学院作物科学研究所，北京 100081） |
| |
| 基金项目： | 国家重点研发计划（2021YFD1200601），中国农业科学院基本科研业务费专项(S2018QY02) |
| |
| 摘 要： | 为探讨基于多源遥感数据和机器学习算法预测冬小麦产量的可行性，利用中麦175/轮选987重组自交系F7代群体中70个家系开展田间试验，通过无人机遥感平台和地面表型车平台及手持式冠层鉴定平台，获取冬小麦灌浆期光谱数据，分别用4种机器学习方法和集成方法建立产量预测模型。结果表明，在61个光谱指数中，除MCARI、DSI、PVI外，其余指数均与产量显著相关或极显著相关，700 nm和800 nm组合的高光谱指数能够比较准确地预测产量。相对于高光谱和多光谱，RGB传感器预测产量精度最高，平均决定系数（r2）为0.74，平均均方根误差（RMSE）为517.78 kg·hm-2。相对于决策树（DT）、随机森林（RF）、支持向量机（SVM）三种传统机器学习算法，岭回归（RR）算法预测产量的精度最高，平均r2为0.73，平均RMSE为516.1 kg·hm-2。与单一的传统机器学习算法相比，DT、RF、SVM、RR结合集成算法的预测精度高且稳定，r2高达0.77，RMSE也较低。SVM 、RF、DT、RR四种机器学习算法和RGB、ASD、UAV、UGV四个传感器构成的算法-传感器集成方法的预测精度提升，r2为0.79，RMSE降至469.98 kg·hm-2。因此，利用Stacking集成方法将不同算法、传感器进行结合，能够有效地提高冬小麦产量预测精度。
|
| 关 键 词： | 遥感 冬小麦 产量预测 机器学习 集成学习 |
Winter Wheat Yield Prediction Based on Multi-Source Remote Sensing Data and Machine Learning Algorithms |
| |
| GAN Tian,LI Lei,LI Hongye,SONG Chengyang,XIE Yongdun,TAO Zhiqiang,XIAO Yonggui,MENG Yaxiong.Winter Wheat Yield Prediction Based on Multi-Source Remote Sensing Data and Machine Learning Algorithms[J].Journal of Triticeae Crops,2022(11):1419-1428. |
| |
| Authors: | GAN Tian LI Lei LI Hongye SONG Chengyang XIE Yongdun TAO Zhiqiang XIAO Yonggui MENG Yaxiong |
| |
| Abstract: | In order to explore the feasibility of predicting winter wheat yield based on multi-source remote sensing data and machine learning algorithms,field experiments were carried out with 70 F7 families derived from Zhongmai 175/Lunxuan 987 were used as materials. The spectral data of winter wheat during grain filling period were obtained through UAV remote sensing platform,ground phenotypic vehicle platform and handheld canopy identification platform. Four machine learning methods and integrated methods were used to establish yield prediction models. The results showed that among the 61 spectral indices,except for MCARI,DSI,and PVI,the other indices were significantly correlated with yield,and the combination of 700 nm and 800 nm hyperspectral indices predicted yield more accurately. Compared with hyperspectral and multispectral,the RGB sensor had the highest yield prediction accuracy,with an average coefficient of determination (r2) of 0.74 and an average root mean square error (RMSE) of 517.78 kg·hm-2. Compared with the three traditional machine learning algorithms of decision tree (DT),random forest (RF) and support vector machine (SVM),the ridge regression (RR) algorithm had the highest accuracy in predicting yield,with an average r2 of 0.73 and an average RMSE of 516.1 kg·hm-2. Compared with the single traditional machine learning algorithm model,the prediction accuracy of DT,RF,SVM,and RR combined with the integrated algorithm was high and stable,with an r2 of up to 0.77 and a low RMSE. The prediction accuracy of sensor integration algorithm,which was composed of four machine learning algorithms of SVM,RF,DT and RR and four sensors of RGB,ASD,UAV and UGV,was improved,with r2 of 0.79,where the RMSE was reduced to 469.98 kg·hm-2. Therefore,using the stacking integration method to combine different algorithms and sensors can effectively improve the yield prediction accuracy. |
| |
| Keywords: | Remote sensing Winter wheat Yield prediction Machine learning Ensemble learning |
|
| 点击此处可从《麦类作物学报》浏览原始摘要信息 |
| 点击此处可从《麦类作物学报》下载免费的PDF全文 |
|
