【摘要】 【目的】氮素的精准监测和合理施用对小麦健康生长、产量及品质提升、减少农田环境污染与资源浪费尤为重要。为精准监测小麦生长关键生育期植株氮含量，探索机器学习方法构建的植株氮含量预测模型的迁移能力。【方法】小区试验于2020—2022年在河南省商水县开展，在冬小麦拔节期、孕穗期、开花期和灌浆期，采用M600大疆无人机搭载K6多光谱成像仪获取5波段（Red、Green、Blue、Rededge、Nir）多光谱影像。基于5个波段冠层反射率提取20种植被指数和40种纹理特征，采用相关分析从65个影像特征中筛选冬小麦植株氮含量敏感特征。基于筛选出的敏感特征，采用BP神经网络（BP）、随机森林（RF）、Adaboost、支持向量机（SVR)4种机器学习回归方法构建植株氮含量预测模型，并对模型预测效果和在不同水处理条件下模型的迁移预测能力进行分析。【结果】（1）植株氮含量与影像特征的相关系数通过0.01极显著水平检验的包括22个光谱特征和29个纹理特征。（2)4种机器学习回归方法构建的冬小麦植株氮含量预测模型存在差异，RF和Adaboost方法预测植株氮含量集中于95%的置信区间，多分布于1:1直线附近，而BP和SVR方法预测的植株氮含量分布相对较为分散；RF方法构建的预测模型R ~2最大，RMSE最小，MAE中等，分别为0.81、0.42%和0.29%;SVR方法构建的预测模型R ~2最小，RMSE和MAE较大，分别为0.66、0.54%和0.40%。（3）以W1处理（按需灌溉）实测植株氮含量为训练集，采用BP、RF、Adaboost和SVR方法构建的模型对W0处理冬小麦植株氮含量迁移预测R ~2分别为0.75、0.72、0.72和0.66；以W0处理（自然状态）实测植株氮含量为训练集，BP、RF、Adaboost和SVR方法构建的模型对W1处理冬小麦植株氮含量迁移预测R ~2分别为0.51、0.69、0.61和0.45。【结论】4种机器学习方法构建的冬小麦植株氮含量预测模型均表现出了较强的迁移预测能力，尤以RF和Adaboost方法构建的模型预测效果和迁移能力为好。更多还原

【Abstract】 【Objective】Accurate monitoring and rational application of nitrogen are particularly important for healthy growth,yield and quality improvement of wheat,and reduction of environmental pollution and resource waste.The purpose of this study was to develop UAV-based models for accurately and effectively assessment of the plant nitrogen content in the key growth stages of wheat growth,and to explore the transferability of the models constructed based on machine learning methods.【Method】Winter wheat experiment were conducted from 2020 to 2022 in Shangshui county,Henan province,China.Based on the K6multichannel imager mounted on DJM600 UAV,5-band (Red,Green,Blue,Rededge,and Nir) multispectral images were obtained from a UAV system in the stages of jointing,booting,flowering and filling in winter wheat,to calculate 20 vegetation indices and40 texture features from different band combinations.Correlation analysis was used to screen the sensitive characteristics of nitrogen content in winter wheat plants from the 65 image features.Combining the sensitive spectral features and texture features of the nitrogen content of winter wheat plants,BP neural network (BP),random forest (RF),Adaboost,and support vector machine (SVR) machine learning regression methods were used to build plant nitrogen content models,and compared for the model performance and transferability.【Result】(1)The correlation coefficients between plant nitrogen content and image features passed the test of 0.01 extremely significant level,including 22 spectral features and 29 texture features.(2) 51 spectral and texture features were adopted to build four machine learning models.The estimates of plant nitrogen by the RF and Adaboost methods were relatively concentrated,mostly close to the 1:1 line;while the estimations from the BP and SVR methods were relatively scattered.The RF method was the best,with R~2,RMSE,and MAE of 0.81,0.42%,and 0.29%,respectively;The SVR method was the worst,with R~2,RMSE,and MAE of 0.66,0.54%and 0.40%,respectively.(3) The prediction effects of the four methods on the nitrogen content of W0 and W1 treatments trained using W1 and W0 treatments were the same as those trained using both W0 and W1 datasets,both of which were closer to the 1:1 line for the RF and Adaboost methods.The R~2 of transfer prediction results for the models constructed by BP,RF,Adaboost,and SVR methods were 0.75,0.72,0.72,and 0.66 for the prediction of nitrogen content in W0 treatment and 0.51,0.69,0.61 (trained using data under W1 treatment) and 0.45 for the prediction under W1 treatment (trained using data under W0 treatment),respectively.【Conclusion】All models showed strong transferability,especially the RF and Adaboost methods,in predicting winter wheat nitrogen content under rainfed and irrigation water management.更多还原