GOA-optimized deep learning for soybean yield estimation using multi-source remote sensing data.

Area of Science:

• Agricultural Science
• Remote Sensing
• Artificial Intelligence

Background:

• Accurate crop yield estimation is vital for global food security.
• Soybean yield prediction presents unique challenges due to large-area cultivation.
• Existing methods often struggle with complex temporal dynamics and diverse data sources.

Purpose of the Study:

• To develop a precise county-level soybean yield estimation framework for the United States.
• To leverage multi-variable remote sensing data for enhanced prediction accuracy.
• To establish a new benchmark in agricultural yield estimation.

Main Methods:

• Utilized a state-of-the-art Convolutional Neural Network-Bidirectional Gated Recurrent Unit (CNN-BiGRU) model.
• Enhanced the model with a novel attention mechanism, the Gated Channel-wise Attention (GCBA).
• Integrated multi-source remote sensing data, including photosynthesis-related parameters.

Main Results:

• The GCBA model outperformed five leading machine learning and deep learning models.
• Achieved superior performance in 2019 and 2020 evaluations with remarkable R², RMSE, MAE, and MAPE values.
• Demonstrated that integrating multi-source remote sensing data significantly enhances yield estimation precision.

Conclusions:

• The developed deep learning framework sets a new standard for soybean yield estimation accuracy.
• Synthesizing diverse remote sensing data and photosynthesis parameters is key to improving predictions.
• Provides critical insights for precision agriculture and informed decisions on food security.