Li Jian, Kang Junrui, Lu Jian, Fu Hongkun, Li Zheng, Liu Baoqi, Lin Xinglei, Zhao Jiawei, Guan Hengxu, Liu He, Liu Zhihan
College of Information Technology, Jilin Agricultural University, Changchun, China.
Jilin Province Cross-Regional Collaborative Innovation Center for Agricultural Intelligent Equipment, Jilin Agricultural University, Changchun, China.
Front Plant Sci. 2025 Jul 21;16:1640806. doi: 10.3389/fpls.2025.1640806. eCollection 2025.
Accurate wheat yield estimation is crucial for efficient crop management. This study introduces the Spatio-Temporal Fusion Mixture of Experts (STF-MoE) model, an innovative deep learning framework built upon an LSTM-Transformer architecture.
The STF-MoE model incorporates a heterogeneous Mixture of Experts (MoE) mechanism with an adaptive gating network. This design dynamically processes fused multi-source remote sensing features (e.g., near-infrared vegetation reflectance, NIRv; fraction of photosynthetically active radiation absorption, Fpar) and environmental variables (e.g., relative humidity, digital elevation model) across multiple expert networks. The model was applied to estimate wheat yield in six major Chinese provinces.
The STF-MoE model demonstrated exceptional accuracy in the most recent estimation year (R² = 0.827, RMSE = 547.7 kg/ha) and exhibited robust performance across historical years and extreme climatic events, outperforming baseline models. Relative humidity and digital elevation model were identified as the most critical yield-influencing factors. Furthermore, the model accurately estimated yield 1-2 months before harvest by identifying key phenological stages (March to June).
STF-MoE effectively handles multi-source spatiotemporal complexity via its dynamic gating and expert specialization. While underestimation persists in extreme-yield regions, the model provides a scalable solution for pre-harvest yield estimation. Future work will optimize computational efficiency and integrate higher-resolution data.
准确估计小麦产量对于高效作物管理至关重要。本研究介绍了时空融合专家混合模型（STF-MoE），这是一种基于长短期记忆网络-Transformer架构构建的创新深度学习框架。
STF-MoE模型将异构专家混合（MoE）机制与自适应门控网络相结合。这种设计通过多个专家网络动态处理融合的多源遥感特征（例如近红外植被反射率，NIRv；光合有效辐射吸收比例，Fpar）和环境变量（例如相对湿度、数字高程模型）。该模型被应用于估计中国六个主要省份的小麦产量。
STF-MoE模型在最近的估计年份表现出卓越的准确性（R² = 0.827，RMSE = 547.7千克/公顷），并且在历年和极端气候事件中均表现出稳健的性能，优于基线模型。相对湿度和数字高程模型被确定为最关键的产量影响因素。此外，该模型通过识别关键物候阶段（3月至6月），在收获前1至2个月准确估计了产量。
STF-MoE通过其动态门控和专家专业化有效地处理了多源时空复杂性。虽然在极端产量地区仍存在低估情况，但该模型为收获前产量估计提供了可扩展的解决方案。未来的工作将优化计算效率并整合更高分辨率的数据。
