Rapid and accurate crop yield estimation is an imperative aspect of agricultural planning that is important for crop management, food security and commodity trading. There are many related factors affecting wheat yield and the relationship between them and the yield is complicated, with nonlinear spatial-temporal characteristics that are difficult to describe accurately with mathematical functions. Deep learning models can fit complex nonlinear functions efficiently and transform input data into high-dimensional features automatically. However, the feature learning process does not produce transparent information. There has been considerable evidence that the ability of attention mechanism for modeling interpretation has been demonstrated in many fields. Therefore, an attention mechanism-based multi-level crop network (AMCN) was proposed to estimate the county level wheat yield based on remote sensing data and meteorological data. To explore the difference in spatio-temporal feature extraction ability under parallel and series structures when combining CNN with LSTM, we designed the AMCN models with two forms of structure, one is a parallel module of LSTM and CNN (AMCN1) and the other is a serial connection module between LSTM and CNN (AMCN2). Our results showed that the AMCN1 model provided an improved estimation accuracy as compared to that of the AMCN2 model. We also found remote sensing data contributed significantly to crop yield estimation mainly at the late growth stages, meteorological data provided additional information mainly at the early growth stage. We assessed the estimated uncertainty using Monte Carlo dropout, and the results indicated that the uncertainty level decreased gradually as the growth stages proceeded. In addition, extreme events such as drought and uneven distribution characteristics of the samples were associated with much higher estimated uncertainties. The study highlighted that the proposed model provided more accurate yield estimations by taking advantage of multi-level crop networks while considering the uncertainty involved in model estimations.

中文翻译：

---

基于注意力机制的小麦产量估算和遥感变量不确定性分析的深度学习方法


快速准确的作物产量估算是农业规划的一个重要方面，对于作物管理、粮食安全和商品贸易非常重要。影响小麦产量的相关因素较多，其与产量的关系复杂，具有非线性时空特征，难以用数学函数准确描述。深度学习模型可以有效地拟合复杂的非线性函数，并将输入数据自动转换为高维特征。然而，特征学习过程不会产生透明信息。有相当多的证据表明，注意力机制建模解释的能力已在许多领域得到证明。因此，提出一种基于注意力机制的多级作物网络（AMCN），基于遥感数据和气象数据估算县级小麦产量。为了探究CNN与LSTM结合时并行和串行结构下时空特征提取能力的差异，我们设计了两种结构形式的AMCN模型，一种是LSTM和CNN的并行模块（AMCN1），另一种是LSTM和CNN的并行模块（AMCN1）。 LSTM 和 CNN 之间的串行连接模块（AMCN2）。我们的结果表明，与 AMCN2 模型相比，AMCN1 模型提供了更高的估计精度。我们还发现，遥感数据主要在生长后期对作物产量估算有显着贡献，气象数据主要在生长早期提供额外信息。我们使用蒙特卡罗辍学评估了估计的不确定性，结果表明随着生长阶段的进行，不确定性水平逐渐下降。 此外，干旱等极端事件和样本分布不均特征与较高的估计不确定性相关。该研究强调，所提出的模型通过利用多级作物网络，同时考虑模型估计中的不确定性，提供了更准确的产量估计。