Dynamic Gating-Enhanced Deep Learning Model with Multi-Source Remote Sensing Synergy for Optimizing Wheat Yield Estimation

Jian Li^1*Junrui Kang¹Jian Lu²Hongkun Fu³Zheng Li¹Baoqi Liu⁴Xinglei Lin¹Jiawei Zhao¹Hengxu Guan¹He Liu^5*Zhihan Liu¹

• ¹College of Information Technology, Jilin Agriculture University, Changchun, China
• ²College of Resources and Environment, Jilin Agriculture University, Changchun, China
• ³College of Agriculture, Jilin Agriculture University, Changchun, China
• ⁴Northeast Institute of Geography and Agroecology Chinese Academy of Sciences, Changchun, China
• ⁵College of Engineering and Technology, Jilin Agriculture University, Changchun, China

Accurate estimation of wheat yield is the cornerstone of efficient crop management. In this study, the STF-MoE (Spatio-Temporal Fusion Mixture of Experts) model was developed by innovatively introducing a mixture of heterogeneous experts (MoE) mechanism into an LSTM-Transformer framework. This mechanism employs an adaptive gating network to dynamically allocate fused multisource remote sensing features (e.g., near-infrared vegetation reflectance, NIRv; fraction of photosynthetically active radiation absorption, Fpar) to multiple heterogeneous expert networks for refined processing. By integrating outputs from these expert networks, the model efficiently processes these complex features, enabling high-precision yield estimation in six major wheat-producing provinces in China. In the most recent estimation year, the model demonstrated exceptional accuracy (R² = 0.827, RMSE = 547.7 kg/ha) and exhibited robust performance during historical years and extreme climatic events. Relative importance analysis of input variables revealed that relative humidity (RHum) and digital elevation model (DEM) are the most critical factors influencing yield estimation, while higher values of Fpar , NIRv, and leaf area index (LAI) show significant positive correlations with wheat yield. Furthermore, the STF-MoE model reliably identified key phenological stages for yield formation (March to June) and achieved high-precision yield estimation approximately 1-2 months before harvest. Notably, the model operates independently of future weather conditions, as its wheat yield estimation primarily relies on capturing yield-related features embedded in early-to midgrowing season data.