欢迎访问《麦类作物学报》编辑部网站

作者	单位
赵泽阳,李美玲,徐伟,刘冰雪,黄鹏宇,康迪，张改生，宋瑜龙	(西北农林科技大学农学院/国家杨凌农业生物技术育种中心/国家小麦改良中心杨凌分中心/小麦育种教育部工程研究中心/ 陕西省作物杂种优势研究与利用重点实验室,陕西杨凌 712100)

摘要点击次数: 288

全文下载次数: 164

中文摘要:

为实现冬小麦产量的准确估算，利用无人机采集266个小麦品种（系）的多时相冠层多光谱数据，提取多个植被指数，分别基于多元线性回归（PLSR）、支持向量机（SVR）、随机森林（RF）、反向传播神经网络（BP）等机器学习算法建立单一生育时期和多生育时期结合的小麦籽粒产量预测模型，并采用决定系数（R²）、均方根误差（RMSE）对所获模型精度进行评价。结果表明，基于植被指数进行单一生育时期的产量预测时，最佳预测时期是灌浆中期，最优模型是RF模型，其预测R²和RMSE 分别为0.655和0.550 kg·m^-2；多生育时期结合进行产量预测时，最优模型是基于5个生育时期（扬花期、灌浆中期、灌浆后期、蜡熟期和完熟期）多光谱数据的RF模型，其预测R²和 RMSE分别为0.834和0.381 kg·m^-2。在建模特征中加入纹理特征后，冬小麦产量预测模型的精度进一步提高，其中以蜡熟期和完熟期结合的SVR模型最优，其预测R²和 RMSE分别为0.924和0.253 kg·m^-2。因此，可基于植被指数加纹理特征对冬小麦产量进行无人机冠层光谱预测，其中以基于多生育时期（蜡熟期、完熟期）的SVR模型预测精度最佳。

英文摘要:

In order to realize the accurate estimation of winter wheat yield, this study used an unmanned aerial vehicle(UAV) to collect multi-temporal canopy multispectral data from 266 wheat varieties(lines), extracted multiple vegetation indices, and established a single-fertility and multi-fertility combined single-stage and multi-stage based on machine learning algorithms, such as multiple linear regression(PLSR), support vector regression(SVR), random forest(RF), and back-propagation neural network(BP), to set up. wheat grain yield prediction model. At the same time, the coefficient of determination(R²) and root mean square error(RMSE) were used to evaluate the models. It was found that the yield prediction with a single growth stage was best at mid-filling stage, and the optimal model was the RF model with the prediction accuracy of R²=0.655 and RMSE=0.550 kg·m^-2 and the optimal model for the yield prediction model with a combination of multi-growth stages was the RF model based on multi-spectral data from the five development stages(flowering, mid-filling, late-filling, dough and full-maturity), with the prediction accuracy of R²=0.834, RMSE=0.381 kg·m^-2. Meanwhile, the accuracy of the winter wheat yield prediction model was further improved after adding textural features to the modeling features. Based on a single stage, the optimal prediction stage for the vegetation index combined with textural features was the combination of dough and full-maturity stages, and the optimal model was the SVR model, with the prediction accuracy of R²=0.924 and RMSE=0.253 kg·m^-2. Therefore, unmanned canopy spectra can be used to predict winter wheat yield based on vegetation indices plus textural features, with the best prediction accuracy based on SVR models with multiple growth stages(dough and full-maturity).

查看全文查看/发表评论下载PDF阅读器

关闭