A Framework for Predicting Winter Wheat Yield in Northern China with Triple Cross-Attention and Multi-Source Data...

Area of Science:

• Agricultural Science
• Data Science
• Machine Learning

Background:

• Existing yield prediction models struggle to fully integrate interactions between multiple influencing factors.
• Accurate winter wheat yield prediction is crucial for food security and agricultural management.

Purpose of the Study:

• To develop an advanced winter wheat yield prediction framework that effectively fuses multi-source data.
• To improve the accuracy and timeliness of winter wheat yield predictions.

Main Methods:

• A novel framework integrating satellite, climate, and soil data using a triple cross-attention fusion mechanism.
• Incorporation of a graph attention mechanism alongside multi-head self-attention in the prediction module's encoder.
• Utilized statistical data to construct multi-source feature sequence sets for comprehensive analysis.

Main Results:

• The proposed method achieved lower mean absolute error (385.99 kg/hm²) and root mean squared error (501.94 kg/hm²) compared to existing approaches.
• Identified the jointing-heading stage (March-April) as the most critical period influencing winter wheat production.
• Demonstrated the model's capability for early yield prediction, nearly a month in advance.

Conclusions:

• The triple cross-attention framework significantly enhances multi-source data fusion for accurate winter wheat yield prediction.
• The model's ability to capture both global dependencies and local feature information contributes to its superior performance.
• Early and accurate yield prediction is feasible, aiding in proactive agricultural planning and resource management.