基于高光谱遥感和集成学习方法的冬小麦产量估测研究

【目的】利用2种灌溉处理下不同发育阶段的冬小麦冠层高光谱信息,通过机器学习方法对小麦籽粒产量进行估测精度研究,明确产量最佳估测模型,对于育种工作有着重要应用价值。【方法】以黄淮麦区207个主栽小麦品种为材料,于2018—2019和2019—2020年度连续2个生长季在河南省新乡基地的正常灌溉和节水处理下种植,并调查开花期、灌浆前期和灌浆中期的冠层高光谱数据,分别以6种机器学习方法和集成方法建立光谱指数产量估测模型。【结果】2种灌溉处理下,3个生育期各光谱指数均与产量呈极显著相关（P<0.0001）,且表现出较高的遗传力（0.61-0.85）,主要受遗传因素控制。在正常灌溉处理下,与传统机器学习方法表现最佳的模型相比,集成学习方法在3个生育期的平均决定系数(R²) 分别由0.610、0.611和0.640提高至0.649、0.612和0.675,平均均方根误差 (RMSE) 分别降低至0.607、0.612和0.593 t·hm^-2;节水处理下,3个生育期的平均R²分别由0.461、0.408和0.452提高至0.467、0.433和0.498,平均RMSE分别降低至0.519、0.559和0.504 t·hm^-2。【结论】利用集成方法将不同模型估测结果进行结合,能够有效地提高产量估测精度,2种灌溉处理下均在灌浆中期估测精度最佳,可为冬小麦育种工作中产量估测提供参考。

Research on Winter Wheat Yield Estimation Based on Hyperspectral Remote Sensing and Ensemble Learning Method

【Objective】Using the hyperspectral data of winter wheat canopy at different development stages under two irrigation treatments, the estimation accuracy of wheat grain yield was studied by machine learning method, and the best yield estimation model was defined, which had the important application value for crop breeding. 【Method】 A total of 207 widely-grown wheat varieties in the Yellow and Huai Valleys Winter Wheat Zone (YHVWWZ) of China were planted under full irrigation and limited irrigation treatments in Xinxiang, Henan province during two consecutive growing seasons of 2018-2019 and 2019-2020, the canopy hyperspectral was investigated at three growth stages after flowering, and six machine learning methods and ensemble methods were adopted to establish yield prediction model by using spectral index as input features.【Result】 The spectral indices at each growth stage were significantly correlated with yield (P<0.0001) under both the two irrigation treatments, and also showed high heritability (0.61-0.85) across all the three growth stages under both the irrigation treatment, which were mainly controlled by genetic factors. Under the full irrigation treatment, compared with the model with the best performance of traditional machine learning methods, the average coefficient of determination (R²) of ensemble learning method in the three growth stages increased from 0.610, 0.611 and 0.640 to 0.649, 0.612 and 0.675, respectively, and the average root mean square error (RMSE) decreased to 0.607, 0.612 and 0.593 t·hm^-2, respectively; Under the limited irrigation treatment, the average R ² increased from 0.461, 0.408 and 0.452 to 0.467, 0.433 and 0.498, respectively, and the average RMSE decreased to 0.519, 0.559 and 0.504 t·hm ^-2, respectively.【Conclusion】Combining the prediction results of different models with the ensemble learning method could effectively improve the yield estimation accuracy, and the mid grain filling achieved the best prediction accuracy under both the two irrigation treatments. Overall, this study could provide the reference for yield estimation in winter wheat breeding.