【摘要】 [目的]收获前获取水稻育种小区的产量信息是高通量表型监测的重要组成部分，也是水稻高产育种的迫切需求。当前，水稻产量预测方法大多基于少数品种，且由线性回归、机器学习等方法建模，因此估产模型一般都存在迁移性较差、精度不高等问题。本研究旨在利用无人机影像和深度学习网络构建适用于多水稻品种的产量预测模型。[方法]利用无人机获取水稻育种试验的多时相RGB和多光谱影像，系统比较了线性回归、机器学习、深度学习算法在产量预测上的表现；提出一种基于注意力机制和卷积神经网络的水稻产量预测方法，并对比分析了ResNet50、MobileNetV3、ShuffleNetV2三种网络的表现。[结果]线性回归和机器学习算法在水稻育种小区产量预测上表现较差（R²<0.3）。MobileNet模型的收敛速度和预测精度是最高的，测试结果的R²、RMSE、RRMSE分别为0.55、1.06 t·hm^-2、12.62%。引入注意力机制的MobileNet模型的收敛速度和预测精度得到了一定的提高，测试结果的R²、RMSE、RRMSE分别为0.58、1.03 t·hm^-2、12.26%。利用时域卷积网络（temporal convolutional network）构建的时间序列模型对水稻产量预测精度有一定提升，R²、RMSE、RRMSE分别达到0.64、0.96 t·hm^-2、11.4%。[结论]卷积神经网络为水稻育种材料产量预测提供了可靠方法，为基于无人机平台的水稻高通量表型研究提供了较好的技术支撑。更多还原

【Abstract】 [Objectives]Obtaining yield information of rice breeding plots before harvest is an important part of high-throughput phenotyping, and it is also an urgent need for high-yield rice breeding. At present, most of the rice yield prediction models are based on a few varieties, and are modeled by linear regression, machine learning and other methods. Therefore, the yield estimation models generally have poor mobility and low accuracy. This study aimed to use UAV images and deep learning networks to construct a yield prediction model, which was suitable for multiple rice varieties. [Methods]The multi-temporal UAV-based RGB and multispectral images of rice breeding experiments were obtained, and the performance of linear regression, machine learning and deep learning algorithms in yield prediction was systematically compared. Furthermore, a new methodology of rice yield prediction was proposed using attention mechanism and convolutional neural network, and the performance of ResNet50,MobileNetV3 and ShuffleNetV2 were compared. [Results]The linear regression and machine learning algorithms performed poorly in predicting the yield of rice breeding plots（R²<0.3）. The convergence speed and prediction accuracy of the MobileNet model were the highest with R²,RMSE,and RRMSE of 0.55,1.06 t·hm^-2,and 12.62%,respectively. The convergence speed and prediction accuracy of the MobileNet model with attention mechanism were improved to a certain extent with R²,RMSE and RRMSE of 0.58,1.03 t·hm^-2 and 12.26%,respectively. The time series model constructed by the temporal convolutional network（TCN）had a certain improvement in the prediction accuracy of rice yield, with R²,RMSE and RRMSE reaching 0.64,0.96 t·hm^-2 and 11.4%,respectively. [Conclusions]The convolutional neural network provided a reliable method for yield prediction in rice breeding experiments, and provided a good technical support for UAV-based high-throughput phenotyping of rice.更多还原