基于无人机高光谱遥感与机器学习的小麦品系产量估测研究

Yield Estimation of Wheat Lines Based on UAV Hyperspectral Remote Sensing and Machine Learning

摘要:

为快速、准确地估测小麦产量，有效提高育种工作效率，本文以小麦品系为研究对象，收集小麦灌浆期无人机高光谱数据和产量数据。首先基于递归特征消除法筛选出特征波长作为模型输入变量，然后利用岭回归（Ridge regression，RR）、偏最小二乘回归（Partial least squares regression，PLS）、多元线性回归（Multiple linear regression，MLR）3种线性算法和随机森林（Random forest，RF）、梯度提升回归（Gradient boosting regression，GBR）、极限梯度提升（eXtreme gradient boosting，XGB）、高斯过程回归（Gaussian process regression，GPR）、支持向量回归（Support vector regression，SVR）、K最邻近算法（K-nearest neighbor，KNN）6种非线性算法构建单一算法产量估测模型并进行精度比较，最后基于Stacking算法构建多模型集成组合，筛选最佳集成模型。结果表明，基于不同算法的产量估测模型精度差异显著，非线性模型优于线性模型，基于GBR的产量估测模型在单一模型中表现最优，训练集R2为0.72，RMSE为534.49kg/hm2，NRMSE为11.10%，测试集R2为0.60，RMSE为628.73kg/hm2，NRMSE为13.88%。基于Stacking算法构建的集成模型性能与初级模型和次级模型的选择密切相关，以KNN、RR、SVR为初级模型组合，GBR为次级模型的集成模型有效提高了估测精度，相比单一模型GBR，训练集R2提高1.39%，测试集R2提高3.33%。本研究可为基于高光谱技术的小麦品系产量估测提供应用参考。

Abstract:

Rapid and accurate estimation of wheat yield can improve the efficiency of breeding. Yield data of wheat lines and hyperspectral data during grain filling period were collected. Firstly, the feature wavelengths were selected as model input variables by using recursive feature elimination method. Then three linear algorithms (ridge regression, partial least squares regression, multiple linear regression) and six nonlinear algorithms (random forest, gradient boosting regression, eXtreme gradient boosting, Gaussian process regression, support vector regression, K-nearest neighbor) were employed to establish single algorithm yield estimation models for precision comparison. Finally, the Stacking algorithm was adopted to develop multi-model ensemble combinations, aiming to identify the optimal ensemble model. The results showed that the accuracy of yield estimation models, based on different algorithms, varied significantly, and that the nonlinear models were better than the linear models. The yield estimation model based on GBR performed best in the single models, with R2 of 0.72, RMSE of 534.49kg/hm2 and NRMSE of 11.10% in the training set, R2 of 0.60, RMSE of 628.73kg/hm2, and NRMSE of 13.88% in the testing set. The performance of the ensemble models based on Stacking algorithm was closely related to the selection of primary and secondary models. The model with KNN, RR, SVR as primary models and GBR as the secondary model effectively improved the yield estimation accuracy. Compared with the single model GBR, the training set R2 was increased by 1.39% and the testing set R2 was increased by 3.33%. The research result can provide an application reference for yield estimation of wheat lines based on hyperspectral technology.