Abstract: In order to quickly and accurately estimate the nitrogen content in leaves of direct⁃seeded rice，taking Huanghuazhan，Chunliangyouchang 70 and Liangyou 185 as objects，the DJI Phantom 4 drone was used to obtain the canopy multispectral images of the four periods，tillering stage，jointing stage，booting stage and heading stage of direct⁃seeded rice. Five vegetation indices with high correlation with leaf nitrogen content were selected respectively，and four algorithms，support vector machine（SVM），random forest（RF），back propagation neural network（BPNN） and partial least squares regression（PLSR） were used to build an estimation model of rice leaf nitrogen content and verify the accuracy.The results showed that there were differences in the correlation between vegetation index and leaf nitrogen content（CNC） of rice plants at different growth stages.The correlation between enhance vegetation index（EVI） and leaf nitrogen content at tillering stage was higher. The correlation coefficient was 0.858.The correlation coefficient between ratio vegetation index（RVI） and leaf nitrogen content was higher at jointing stage（0.938），the correlation coefficient between RVI and leaf nitrogen content was higher at booting stage （0.793），the NDVI and SAVI indices at the heading stage showed a higher correlation with leaf nitrogen content，with a correlation coefficient of 0.782.The accuracy of the models at tillering stage and jointing stage was higher，while the accuracy of the models at booting stage and heading stage was lower，and the inversion effect was poor.The prediction model of rice leaf nitrogen content based on the RF algorithm performed relatively better，and the R_c² was 0.922，RMSE_c was 0.101%，R_cv² was 0.658，and RMSE_cv was 0.205% at tillering stage.At jointing stage，R_c² was 0.980，RMSE_c was 0.061%，R_cv² was 0.913，and RMSE_cv was 0.124%.The results showed that the leaf nitrogen content monitoring model based on RF had high prediction accuracy and could predict the leaf nitrogen content of direct⁃seeded rice.

Key words: Direct?seeded rice, Unmanned aerial vehicle, Vegetation index, Nitrogen nutrition monitoring, Multispectral image

摘要： 为实现快速、准确估算直播水稻叶片氮含量，以黄华占、春两优长70和两优185为对象，利用大疆精灵4无人机获取直播水稻分蘖期、拔节期、孕穗期和抽穗期4个时期的冠层多光谱影像。分别选取与叶片氮含量相关性较高的5个植被指数，利用支持向量机（SVM）、随机森林（RF）、BP神经网络（BPNN）和偏最小二乘回归（PLSR）4种算法构建水稻叶片氮含量的估算模型并检验精度。结果表明，植被指数与不同生育时期的水稻植株叶片氮含量相关性存在差异，其中，分蘖期的增强植被指数（EVI）与叶片氮含量的相关性较高，相关系数为0.858；拔节期的比值植被指数（RVI）与叶片氮含量的相关性较高，相关系数为0.938；孕穗期的RVI与叶片氮含量的相关性较高，相关系数为0.793；抽穗期的NDVI和SAVI与叶片氮含量的相关性较高，相关系数均为0.782。分蘖期和拔节期的模型精度较高，而孕穗期和抽穗期的模型精度较低，反演效果差。基于RF算法构建的水稻叶片氮含量预测模型效果相对较优，分蘖期的建模集决定系数（R_c²）为0.922，建模集均方根误差（RMSE_c）为0.101%，验证集决定系数（R_cv²）为0.658，验证集均方根误差（RMSE_cv）为0.205%；拔节期R_c² 为0.980，RMSE_c 为0.061%，R_cv² 为0.913，RMSE_cv 为0.124%。表明基于RF算法构建的叶片氮含量监测模型，预测准确性较高，可以实现直播水稻叶片氮含量的预测。

关键词: 直播水稻, 无人机, 植被指数, 氮营养监测, 多光谱影像