Dynamic gating-enhanced deep learning model with multi-source remote sensing synergy for optimizing wheat yield...

Area of Science:

• Agricultural Science
• Remote Sensing
• Deep Learning

Background:

• Accurate wheat yield estimation is vital for global food security and efficient agricultural management.
• Traditional methods often struggle with the spatiotemporal complexities of crop growth and environmental factors.

Purpose of the Study:

• To introduce and evaluate the Spatio-Temporal Fusion Mixture of Experts (STF-MoE) model for precise wheat yield estimation.
• To leverage multi-source remote sensing and environmental data for enhanced prediction accuracy.

Main Methods:

• Developed an LSTM-Transformer based deep learning framework, STF-MoE.
• Integrated a heterogeneous Mixture of Experts (MoE) with an adaptive gating network.
• Fused remote sensing data (NIRv, Fpar) and environmental variables (relative humidity, DEM) for yield prediction.

Main Results:

• Achieved high accuracy in recent yield estimation (R² = 0.827, RMSE = 547.7 kg/ha).
• Demonstrated robust performance across historical data and extreme climate events, outperforming baseline models.
• Identified relative humidity and DEM as key yield-influencing factors and enabled 1-2 month pre-harvest estimation.

Conclusions:

• The STF-MoE model effectively addresses spatiotemporal data complexities through dynamic gating and expert specialization.
• The model offers a scalable solution for pre-harvest wheat yield estimation, despite challenges in extreme yield regions.
• Future research will focus on optimizing computational efficiency and incorporating higher-resolution data.