Crop yield prediction from UAV images has significant potential in accelerating and revolutionizing crop breeding pipelines. Although convolutional neural networks (CNN) provide easy, accurate and efficient solutions over traditional machine learning models in computer vision applications, a CNN training requires large number of ground truth data, which is often difficult to collect in the agricultural context. The major objective of this study was to develope an end-to-end 3D CNN model for plot-scale soybean yield prediction using multitemporal UAV-based RGB images with approximately 30,000 sample plots. A low-cost UAV-RGB system was utilized and multitemporal images from 13 different experimental fields were collected at Argentina in 2021. Three commonly used 2D CNN architectures (i.e., VGG, ResNet and DenseNet) were transformed into 3D variants to incorporate the temporal data as the third dimension. Additionally, multiple spatiotemporal resolutions were considered as data input and the CNN architectures were trained with different combinations of input shapes. The results reveal that: (a) DenseNet provided the most efficient result (R² 0.69) in terms of accuracy and model complexity, followed by VGG (R² 0.70) and ResNet (R² 0.65); (b) Finer spatiotemporal resolution did not necessarily improve the model performance but increased the model complexity, while the coarser resolution achieved comparable results; and (c) DenseNet showed lower clustering patterns in its prediction maps compared to the other models. This study clearly identifies that multitemporal observation with UAV-based RGB images provides enough information for the 3D CNN architectures to accurately estimate soybean yield non-destructively and efficiently.

利用无人机图像预测作物产量在加速和彻底改变作物育种流程方面具有巨大潜力。尽管卷积神经网络 (CNN) 在计算机视觉应用中提供了比传统机器学习模型简单、准确和高效的解决方案，但 CNN 训练需要大量的地面实况数据，而这些数据在农业环境中通常很难收集。本研究的主要目标是开发一种端到端 3D CNN 模型，使用基于多时相无人机的 RGB 图像（包含约 30,000 个样地）进行地块规模大豆产量预测。2021 年，阿根廷利用低成本 UAV-RGB 系统收集了来自 13 个不同实验田的多时相图像。将三种常用的 2D CNN 架构（即 VGG、ResNet 和 DenseNet）转换为 3D 变体以合并时态数据作为第三维度。此外，多个时空分辨率被视为数据输入，并且 CNN 架构使用不同的输入形状组合进行训练。结果表明： (a) DenseNet在准确性和模型复杂性方面提供了最有效的结果 (R ^{2 0.69)，其次是 VGG (R 2} 0.70) 和 ResNet (R ² 0.65)；(b) 较精细的时空分辨率不一定能提高模型性能，反而会增加模型复杂度，而较粗的分辨率则取得了可比的结果；(c) 与其他模型相比，DenseNet 在其预测图中显示出较低的聚类模式。这项研究清楚地表明，基于无人机的 RGB 图像的多时相观测为 3D CNN 架构提供了足够的信息，以非破坏性且高效地准确估计大豆产量。