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

Jian Li, J. S. Kang, Jian Lü, Hongkun Fu, Zheng Li, B. Liu, Xue Lin, Jiawei Zhao, Haibing Guan, Liu He, Zhihan LiuFrontiers in Plant Science 2025 · DOI

TL;DR

This paper introduces STF-MoE, a hybrid deep learning architecture that combines Transformers and LSTMs with a "Mixture of Experts" (MoE) mechanism to predict wheat yields across China. By dynamically routing data through specialized sub-networks, the model achieves high accuracy ($R^2 = 0.827$) and can provide reliable yield estimates up to two months before harvest.

What Problem Does This Paper Address?

Accurate crop yield estimation is the "holy grail" of agricultural management, yet it remains incredibly difficult due to spatiotemporal complexity. Factors that influence yield—weather, soil quality, and topography—don't just vary by location; their importance shifts throughout the growing season.

Existing deep learning models often struggle with this. LSTMs are great at capturing time-series sequences but can be inefficient with long-range data. Transformers excel at global dependencies but may miss local, short-term nuances. Furthermore, simply "dumping" multi-source data (satellite imagery, weather stations, soil maps) into a single model often leads to sub-optimal feature fusion. The authors tackle the challenge of how to effectively "synergize" these diverse data sources to handle the non-linear realities of farming.

Methods

The authors propose the Spatio-Temporal Fusion Mixture of Experts (STF-MoE) model. Its architecture is built on three pillars:

  1. Hybrid Backbone: The model uses a parallel structure. One branch uses a Transformer (4 encoder layers) to capture long-range temporal dependencies across the entire growing season. The other branch uses a Bidirectional LSTM to extract local contextual features and short-term dynamics.
  2. Heterogeneous Mixture of Experts (MoE): This is the core innovation. Instead of one large neural network processing everything, the model features five distinct "expert" sub-networks. These experts have different architectures (varying depths and activation functions like GELU, ReLU, and SiLU) designed to specialize in different data patterns.
  3. Dynamic Gating: A gating network acts as a "traffic controller." It evaluates the incoming data and uses a Top-2 sparse activation strategy, routing the information only to the two experts best suited for that specific input. This optimizes computational resources while allowing for specialized "expert" processing of complex environmental conditions.

Datasets

The study covers six major wheat-producing provinces in China from 2002 to 2021. The multi-source dataset includes:

  • Remote Sensing: MODIS data for vegetation indices (NDVI, EVI, NIRv, GNDVI) and biophysical parameters (LAI, Fpar).
  • Meteorology: ERA5-Land reanalysis data (Relative Humidity, Solar Radiation).
  • Topography: SRTM Digital Elevation Model (DEM) and Slope data.
  • Soil: OpenLandMap Soil Organic Carbon (SoC) content.
  • Ground Truth: County- and city-level statistical yearbooks for actual wheat yield.

Key Insights and Ideas

  • The Power of Specialization: The MoE approach proves that "one size does not fit all." By allowing different sub-networks to specialize in specific "regimes" of data (e.g., extreme weather events vs. normal growth cycles), the model handles outliers much better than monolithic architectures.
  • Humidity and Height Matter: While many models rely heavily on "greenness" (NDVI), this study found that Relative Humidity (RHum) and Elevation (DEM) were the most critical yield-influencing factors. High humidity often correlated with disease risk, while elevation impacted the accumulated temperature available for crop growth.
  • Weak Linear, Strong Non-Linear: The authors noted that linear correlations between single indices (like NDVI) and yield were surprisingly weak ($r < 0.12$). This highlights why deep learning is necessary: the relationship between a satellite pixel and a grain silo is deeply non-linear and context-dependent.
  • Early Warning Capability: The model hits its "stride" in terms of accuracy between March and June. This allows for accurate predictions 1–2 months before harvest, providing a vital window for policy adjustments and food security planning.

Key Results

  • Top-Tier Accuracy: In the 2021 test year, the STF-MoE achieved an $R^2$ of 0.827 and a Root Mean Square Error (RMSE) of 547.7 kg/ha.
  • Benchmark Superiority: It consistently outperformed baseline models. For example, it reduced the mean RMSE by approximately 9.7% compared to a standard LSTM.
  • Temporal Robustness: In "rolling" year-by-year tests from 2002–2021, the model maintained a high mean $R^2$ of 0.8712, proving it isn't just "getting lucky" on specific years but is learning the underlying physics of crop growth.
  • Spatial Granularity: The model successfully mapped yield disparities at both the prefecture and county levels, clearly distinguishing high-yield eastern plains from lower-yield western arid regions.

Limitations and Future Work

Despite its success, the authors acknowledge a few hurdles:

  • The "High-Yield" Ceiling: Like many regression models, STF-MoE tends to underestimate yields in extremely high-producing regions. This "regression to the mean" remains a challenge for capturing record-breaking harvests.
  • Computational Overhead: While the sparse gating helps, the hybrid Transformer-LSTM-MoE structure is computationally heavy. Future work will focus on model compression and "lightweight" designs for easier deployment.
  • Resolution Gaps: The current model uses 500m resolution data. The authors plan to integrate higher-resolution sources (like Sentinel-2 at 10m) to capture sub-field variability and the impact of specific local farming practices.

Summary updated

Abstract

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.
BibTeX
@inproceedings{li2025dynamic,
  author = {Jian Li and J. S. Kang and Jian Lü and Hongkun Fu and Zheng Li and B. Liu and Xue Lin and Jiawei Zhao and Haibing Guan and Liu He and Zhihan Liu},
  title = {{Dynamic gating-enhanced deep learning model with multi-source remote sensing synergy for optimizing wheat yield estimation}},
  booktitle = {Frontiers in Plant Science},
  year = {2025},
  doi = {10.3389/fpls.2025.1640806},
}
Figure