Crop yield is a primary trait to select superior genotypes and evaluate breeding efficiency in breeding programs. Crops with high yield potential are usually selected from numerous breeding lines in multiple years and locations. However, the efficiency of conventional breeding programs is limited by the capacity of field phenotyping, which can be improved by developing high-throughput field phenotyping systems using emerging technologies, including Unmanned Aerial Vehicle (UAV) imagery and deep learning technologies. The goal of this study was to evaluate the performance of a UAV imaging system and convolutional neural network (CNN) in estimating yield of soybean breeding lines. In this study, 972 soybean breeding lines in three maturity groups were planted under rainfed conditions for testing their drought tolerance. Aerial images were taken at the late vegetation (V6), early (R1), and late reproductive (R6-R8) growth stages. Seven image features associated with plant height, canopy colour, and canopy texture were selected to estimate the yield of each breeding line. A mixed CNN model was built to estimate soybean yield by taking the seven image features and two categorical factors, i.e. maturity group and drought tolerance, as predictors. Results show that image features collected at the early and late reproductive growth stages are comparably promising in estimating soybean yield. The prediction model could explain 78% of the measured yield with a root mean square error of 391.0 kg·ha⁻¹ (33.8% to the average yield), indicating that the UAV imagery and deep learning models are promising in estimating yield for soybean breeding purposes.

作物产量是选择优良基因型并在育种计划中评估育种效率的主要特征。具有高产潜力的农作物通常选自多年和不同地点的众多育种系。但是，常规育种计划的效率受到现场表型分析能力的限制，这可以通过使用新兴技术（包括无人飞行器（UAV）图像和深度学习技术）开发高通量现场表型分析系统来提高。这项研究的目的是评估无人机成像系统和卷积神经网络（CNN）在估计大豆育种系产量方面的性能。在这项研究中，在雨育条件下种植了三个成熟组的972个大豆育种系，以测试其抗旱性。在植被后期（V6），早期（R1）和晚期生殖（R6-R8）的生长阶段拍摄了航空影像。选择了与植物高度，冠层颜色和冠层纹理相关的七个图像特征，以估算每个育种系的产量。建立了一个混合的CNN模型，以七个图像特征和两个分类因素（即成熟度组和耐旱性）作为预测指标，以估计大豆产量。结果表明，在早期和晚期生殖生长阶段收集的图像特征在估计大豆单产方面具有可比性。该预测模型可以解释78％的测得产量，均方根误差为391.0 kg·ha 选择树冠质地和树冠质地来估算每个育种系的产量。建立了一个混合的CNN模型，以七个图像特征和两个分类因素（即成熟度组和耐旱性）作为预测指标，以估计大豆产量。结果表明，在早期和晚期生殖生长阶段收集的图像特征在估计大豆单产方面具有可比性。该预测模型可以解释78％的测得产量，均方根误差为391.0 kg·ha 选择树冠质地和树冠质地来估算每个育种系的产量。建立了一个混合的CNN模型，以七个图像特征和两个分类因素（即成熟度组和耐旱性）作为预测指标，以估计大豆产量。结果表明，在早期和晚期生殖生长阶段收集的图像特征在估计大豆产量方面具有相当的前景。该预测模型可以解释78％的测得产量，均方根误差为391.0 kg·ha 结果表明，在早期和晚期生殖生长阶段收集的图像特征在估计大豆单产方面具有可比性。该预测模型可以解释78％的测得产量，均方根误差为391.0 kg·ha 结果表明，在早期和晚期生殖生长阶段收集的图像特征在估计大豆单产方面具有可比性。该预测模型可以解释78％的测得产量，均方根误差为391.0 kg·ha^-1（平均产量的33.8％），这表明UAV图像和深度学习模型有望为大豆育种目的估算产量。